For decades, semiconductor design has been driven by humans coming up with bright ideas that unlock new innovations. But the benefits of better chip design have been reaped, including the rise of AI, which now means there could be another party involved in making chip designs smarter: AI itself.

‘Chip designer’ isn’t one of the roles on the chopping block as AI automation upends the job market. But in the narrow pockets of the design flow where problems are structured, and evaluators are robust, it is starting to be adopted — with benefits.

Google DeepMind's AlphaChip reinforcement-learning system has produced designs for three generations of the company's Tensor Processing Units (TPUs), with DeepMind claiming "superhuman" layouts compared with those produced by human designers. They’re not alone: Synopsys has passed 100 production tape-outs with its DSO.ai design-space-optimization tool, reporting productivity boosts of more than three times and power reductions of up to 25% for customers including STMicroelectronics and SK hynix.

"Like every new technology, AI may have multiple uses," said Borivoje Nikolić, professor of electrical engineering and computer sciences at the University of California, Berkeley, in an interview with Tom’s Hardware Premium. Nikolić drew a parallel with Moore's Law, which has historically been exploited in two ways: to reduce the cost of an existing product by porting it to cheaper processes, or to add features that were previously impossible. "I think AI will be used in both ways," he says. "At the moment, the industry seems to be focused on the first item — how to make things cheaper, how to automate things in a better way than they were in the past."

By contrast, academics are more interested in using AI to discover things humans haven't yet thought of, an approach that mirrors breakthroughs in areas such as drug discovery and protein folding with the likes of AlphaFold.

Nikolić and his colleague Sagar Karandikar have been exploring that territory in their own research on cache replacement policies, a subject deep in the weeds of processor microarchitecture. Their ArchAgent system, built on Google DeepMind's AlphaEvolve framework, generated a cache replacement policy in two days that beat the prior state-of-the-art by 5.3% in IPC speedup on Google's multi-core workload traces. On the heavily worked-over single-core SPEC06 benchmarks, it took 18 days to eke out another 0.9%. That’s a "first sign of life" for Karandikar that large language models can design genuinely new logic, rather than just tinkering with existing parameters.

"There is still a lot of human guidance, and it kind of up-levels the kind of thinking humans have to do," said Karandikar, a computer architecture researcher at Berkeley, in an interview with Tom’s Hardware Premium. "The humans involved in that project are doing more of the high-level thinking — coming up with new ideas and guiding the LLM — and the LLM does a lot of the finer policy development around that."

Where AI is making breakthroughs

For Igor Markov, a chip design researcher who has spent years at the frontline of electronic design automation, the places where AI is adding real value are specific and often mundane. Some of the biggest wins, he says, come at the low end of the flow, such as tasks that previously required engineers to interpret informal specifications written in natural language and convert them into formal descriptions that a tool can act on.

Take power and ground networks, the intricate webs of metal that feed electricity across a chip. "They’re sometimes designed just with descriptions in natural language," Markov said in an interview with Tom’s Hardware Premium. "People explain the geometry, and then it's implemented, and at some point, you need to formalize it. This is a step that was done manually, and it's pretty straightforward to automate using AI." The productivity dividend isn’t massive; “it took a couple of days, now it's a couple of hours,” he explained. But is still better than nothing, even if the output still needs to be checked.

Where Markov is most bullish is on what he calls the agentic space: the high-level orchestration of chip design flows, including deciding whether a run is doomed or whether a flow needs to be restarted entirely. “If you take a zero multiplied by something, you get a zero,” he said. “But if you already have something decent, then this high-level control can be very, very enabling.”

The most stubborn corners of the industry are starting to think about adopting AI. Analog design has long been seen as the last redoubt of human craft, but researchers have begun producing generative AI systems such as AnalogGenie, which uses a GPT-style model to discover new circuit topologies, and Princeton's AI-enabled design-space discovery for millimeter-wave and sub-terahertz power amplifiers operating between 30 and 120 GHz.

It’s in these areas that what’s often seen as AI’s failing, that it doesn’t have an inherent knowledge or muscle memory of its own, becomes a strength. Humans have a tendency when porting a design from one process node to another to assume the old topology must be close to optimal for the new one. “AI may not have those kinds of barriers,” said Nikolić.

Testing versus real life

However, some caution is needed. AI can be trained to ace demos, but can flunk the messier problems engineers face in practice. "Whether something that works in five cases works in general, and allows you to innovate, that's the key," says Markov.

There is also the problem of what it is you are asking AI to do in the first place. Ask a model to design the best chip for AI, and without a formal, unambiguous specification of what best means, the model will produce something — or anything. "You will play whack-a-mole," Markov said when it comes to making it work in practice.

He added that every previous jump in design automation has provoked similar debates about whether machines can really think. Shortest-path algorithms for wire routing, once seen as a distinctly human capability, became undergraduate coursework. Placement algorithms now routinely outperform human designers. Logic synthesis, once considered too abstract to automate, is handled by for loops and conditionals. “EDA has always been a type of AI, because it automated what people did,” Markov said. “We are just moving along the straight line, and there's no stopping.”

For now, AI is acting as a force multiplier, Markov said, squeezing more output from teams rather than shrinking them. Who’s in those teams and what they bring is also shifting: engineers who are fluent with AI coding assistants are now in demand where they weren't six months ago.

Jevons’ paradox also looms large over the potential of AI in the chip design process. As AI makes certain parts of the process dramatically cheaper and faster, Nikolić expects engineers to use that freed-up capacity to explore territory they wouldn't otherwise have dared tackle, including the design of the AI chips driving the whole cycle in the first place.

After all, if any class of silicon is ripe for the kind of optimization that hasn't yet been systematically studied, Markov argues, it is the highly structured, performance-critical accelerators powering the current boom. “There’s plenty of opportunity for humans to be improving other parts of the design flow to make it more amenable to these AI-based systems,” said Karandikar. As models become more advanced, so too might their capacities to assist in chip design and development.

Google will leverage data from one of the most complex and multi-layered sci-fi MMORPGs to train its AI. With a quarter million monthly active users, Eve Online's deep living simulation of economics and politics, with strategic aspects involving exploration and combat, presents the opportunity to expand the capabilities and horizons of AI. Google DeepMind is now primed to learn from interactions in this expansive game world after acquiring a minority stake in the newly independent Fenris Creations, reports Bloomberg.

Readers are likely more familiar with Eve Online being stewarded by CCP Games. Fenris Creations was recently formed when the developers sought to buy back the Eve Online game rights from Korean game-maker Pearl Abyss. The new Icelandic company paid $120M in cash and crypto to set up. While that may sound like a huge sum, Pearl Abyss spent more than double that amount to acquire the game maker back in 2018.

It is important to understand why Google DeepMind would part with "millions" to secure a minority stake in Fenris Creations. Thankfully, Bloomberg has spoken to executives from both firms to more clearly assess how the deal benefits both parties.

A DeepMind director quoted by the source indicates that success in Eve Online relies on skills that are far from mastered by current-generation AIs. This MMORPG is (in)famous for some of its players succeeding using tactics ranging from politics to deceit to outright scams. Eve Online is also lauded for the long-term planning and continual learning involved in success.

DeepMind has previously dabbled with gaming, and among all the big names of AI, it seems like it is the most interested in tapping into this rich vein for training.

From the perspective of Fenris, the deal looks pretty irresistible too, earning millions from a partner with deep pockets and what seems to be a hands-off stance. The Fenris CEO is quoted as previously joking that Eve Online would be “the final boss for AI in games.” He went on to ponder that Eve Online gameplay shines a light on society and the human condition. We are also getting hints that AIs learning from Eve Online could learn something about humans pushed to extremes…

It is stated by Fenris that DeepMind’s initial research will look at player behavior on isolated servers, having no impact on the live game. Moreover, there may be benefits from this research outside of AI training to improve the game, or even inspire new experiences, according to the Fenris CEO.

Fenris is beginning to start work on the successor to Eve Online, dubbed Eve Frontier, as well as an extraction adventure shooter dubbed Eve Vanguard (see video above). So we’re sure the extra cash injection will be very useful.

Eve Online is 23 years old in 2026, but it still attracts a devoted following with between 200,000 and 300,000 monthly active users. Q4 2025 was the game’s second most lucrative ever, with November being notable for breaking records.

As companies pour billions of dollars into AI infrastructure, demand for AI talent to program and run these AI data centers is also greatly increasing. Deedy Das, a Venture Capitalist at Menlo Ventures and a former Google Search staff member, posted on X that Meta has an over $2 million annual pay package for AI talent, but still keeps losing its people to OpenAI and Anthropic. He said that he’s personally heard three such cases this week alone, which is major news given the size of Meta’s compensation.

Statistics say that for every 10.6 from DeepMind, 8.2 from OpenAI, and 2 from Hugging Face that move to Anthropic, it only loses one employee for each company. This movement shows that the startup is quickly growing and that many people from competing AI labs want to work for it. We don’t know how much the company offers, though, but we can safely assume that it’s at least on par or, more likely, substantially larger than what the competition pays.

According to the SignalFire research, beyond salary, Anthropic's edge is a unique culture that embraces "unconventional thinkers" and gives employees true autonomy, as well as flexible work options, a lack of title politics and forced management tracks. Furthermore, employees report an embrace of intellectual discourse and researcher autonomy, compared to bureaucracy elsewhere.

Google is making use of aggressive noncompete clauses and extended notice periods, contends former GoogDeepMinder Nando de Freitas in a recent post on X. In some cases, Google DeepMind’s employment contracts may lock an AI developer into doing nothing for as long as a year, notes Business Insider, to prevent its AI talent from moving to competing firms. That’s a long time away from working on the cutting edge in the rapidly developing world of AI.

De Freitas says the best way to avoid such contractual chains is to simply “don’t sign these contracts.” However, folks with a pressing need for a steady income flow who get offered a well-paying position might easily be swayed into signing terms they would normally turn their nose up at. We know that the FTC banned noncompletes in states like California last year, but Google’s DeepMind is UK-based, so the company probably still has some latitude to push would-be employees to sign such agreements.

So, it is clear that Google is happier to pay AI talent for up to a year to do nothing than let them pass into the hands of rivals. De Freitas can perhaps speak more freely than some DeepMinders, as he has already comfortably migrated to Microsoft, where he is a VP of AI.

As one that got away, de Freitas indicates that current DeepMind employees frequently contact him. “Every week one of you reaches out to me in despair to ask me how to escape your notice periods and noncompetes,” he says in his social media post.

While de Freitas certainly sympathizes with those AI devs who find themselves in limbo, he advises them not to contact him, suggesting a couple of other names who are current leads at the Google-owned AI firm.

Ultimately, though, perhaps the best advice is not to sign a contract with undesirable extended notice periods and noncompetes. The competition for experienced AI staff is such that people should be able to avoid such contracts, which de Freitas describes as an “abuse of power.”

Google responds

BI received a statement from Google about the employee contract issues raised above. "Our employment contracts are in line with market standards," a Google spokesperson told the publication. "Given the sensitive nature of our work, we use noncompetes selectively to protect our legitimate interests."

Elon Musk’s lawsuit against OpenAI has opened a can of worms, as emails between OpenAI co-founders Greg Brockman, Ilya Sutskever, Sam Altman, and Musk himself have revealed turbulent behind-the-scenes drama. According to the Transformer, an email from Musk in 2016 said, “Deepmind [sic] is causing me extreme mental stress. If they win, it will be really bad news with their one mind to rule the world philosophy.” He further aired concerns that “Demis [Hassabis] could create an AGI dictatorship.”

Demis Hassabis is one of the founders of DeepMind, an artificial intelligence research firm founded in 2010, and acquired by Alphabet, Google’s parent company, in 2014. The success of DeepMind is probably one of the major reasons behind the founding of OpenAI.

Altman emailed Musk in May 2015, saying, “Been thinking a lot about whether it’s possible to stop humanity from developing AI. I think the answer is almost definitely not. If it’s going to happen anyway, it seems like it would be good for someone other than Google to do it first. Any thoughts on whether it would be good for YC [Y Combinator] to start a Manhattan Project for AI?” Musk responded with, “Probably worth a conversation.” And, after around six months, the company behind ChatGPT was founded.

Aside from this inter-company rivalry, there’s also some tension going on behind the scenes at OpenAI. Two co-founders — Greg Brockman and Ilya Sutskever — had concerns about Altman and Musk. In September 2017, they emailed Altman asking him why he wanted to be CEO of the company and mentioned that they didn’t fully trust his judgment. Simultaneously, they were skeptical of Musk’s intentions. They said that even though he (Musk) claims to not want control of the final AGI, the negotiations revealed that absolute control was crucial to him. So, they (Brockman and Sutskever) fear that the OpenAI’s structure could allow Elon Musk to be a dictator in the company, should he decide to become one.

The Transformer says that both of these emails did not sit well with their recipients. Shivon Zillis — another OpenAI board member and a close associate of Musk — updated him regarding her conversation with Altman. She said that Sam lost trust with his two other co-founders, Greg and Ilya, through the negotiation process, adding that he felt they were inconsistent and childish. As for Elon, he replied directly to the two and said that it was the final straw for him and that the group should look for funding on its own or let OpenAI remain as a nonprofit entity.

These emails reportedly happened between 2015 and 2018, when Musk was involved in OpenAI. Although it was founded as a nonprofit organization, OpenAI’s leadership team eventually realized that it would need billions to fund the massive computer power required to train its models. They realized that they needed to convert OpenAI to a for-profit entity to secure this funding, but negotiations between Musk, Altman, and its other board members faltered when they couldn’t agree on terms.

Eventually, Musk left OpenAI and founded his AI startup, xAI, which now has a 100,000-GPU-strong data center with a 150-MW power supply. As for OpenAI’s funds, Microsoft eventually stepped in 2019, when it invested a billion dollars in the company. Today, OpenAI is one of the leaders in the field, with ChatGPT, and Microsoft is using its technology as the basis for Copilot. It also plans to invest billions more, as the software giant sees a future in AI. On the other hand, Intel got left behind after it declined to purchase 15% of OpenAI for $1 billion in 2017, saying that the company didn’t see AI providing a return on investment anytime soon.

The CEO of Google DeepMind has compared the IQ levels of contemporary artificial intelligence (AI) agents to domestic cats. “We’re still not even at cat intelligence yet, as a general system,” remarked Hassabis, answering a question about DeepMind’s progress in artificial general intelligence (AGI). However, research is progressing fast, with some huge cash and compute investments propelling it forward. Some expect it to eclipse human intelligence in the next half-decade.

Demis Hassabis, the co-founder and CEO of Google DeepMind, made the artificial intelligence vs. cat IQ comparison in a public discussion with Tony Blair, one of Britain’s ex-Prime Ministers. The talk was part of the Future of Britain Conference 2024, organized by the Institute for Global Change.

Hassabis highlights that his work is not focused on AI but on AGI. It gives us more perspective on how he is looking at the computer vs cat comparison. Yes, a contemporary AI can sometimes write, paint, or make music in a convincingly human-like fashion, but an ordinary house cat has a lot more general intelligence. “At the moment, we’re far from Human-level intelligence across the board,” admitted Hassabis. “But in certain areas like games playing [AI is] better than the best people in the world.”

On the potential of AI to shape our lives, Hassabis boldly reckons that it will be as big as the Industrial Revolution or the harnessing of fire or electricity. In the future, and more specifically, the DeepMind CEO thinks one of the most exciting ways AI will become a leading light will be accelerating scientific discovery in energy, materials science, health care, climate, and mathematics – and it is already doing this. Interestingly, he said we were all talking about ‘big data’ in the noughties, and AI systems are the answer.

Hassabis took the opportunity to plug a DeepMind project called Project Astra. This removes AI from the restrictions of being a mere Chatbot like ChatGPT or Google Gemini, with much more awareness of a user’s situation, environment, preferences, history, and so on. In this way, Project Astra aims to deliver a ‘universal AI agent’ that is helpful in everyday life.

Meanwhile, the biggest hurdles remaining for outfits like DeepMind and human-level AGI achievements include planning, memory, tool use, and smart questioning. The DeepMind CEO knows that big breakthroughs and compute scaling are still needed for AGIs to achieve human IQ levels.

Inflection AI, a new startup found by the former head of deep mind and backed by Microsoft and Nvidia, last week raised $1.3 billion from industry heavyweights in cash and cloud credit. It appears the company will use the money to build a supercomputer cluster powered by as many as 22,000 of Nvidia's H100 compute GPUs, which will have peak theoretical compute power performance that is comparable to that of the Frontier supercomputer.

"We will be building a cluster of around 22,000 H100s," said Mustafa Suleyman, the founder of DeepMind and a co-founder of Inflection AI, reports Reuters. "This is approximately three times more compute than what was used to train all of GPT-4. Speed and scale are what's going to really enable us to build a differentiated product."

A cluster powered by 22,000 Nvidia H100 compute GPUs is theoretically capable of 1.474 exaflops of FP64 performance — that's using the Tensor cores. With general FP64 code running on the CUDA cores, the peak throughput is only half as high: 0.737 FP64 exaflops. Meanwhile, the world's fastest supercomputer, Frontier, has peak compute performance of 1.813 FP64 exaflops (double that to 3.626 exaflops for matrix operations). That puts the planned new computer at second place for now, though it may drop to fourth after El Capitan and Aurora come fully online.

While FP64 performance is important for many scientific workloads, this system will likely be much faster for AI-oriented tasks. The peak FP16/BF16 throughput is 43.5 exaflops, and double that to 87.1 exaflops for FP8 throughput. The Frontier supercomputer powered by 37,888 of AMD's Instinct MI250X has peak BF16/FP16 throughput of 14.5 exaflops.

The cost of the cluster is unknown, but keeping in mind that Nvidia's H100 compute GPUs retail for over $30,000 per unit, we expect the GPUs for the cluster to cost hundreds of millions of dollars. Add in all the rack servers and other hardware and that would account for most of the $1.3 billion in funding.

Inflection AI is currently valuated at around $4 billion, about one year after its foundation. Its only current product is a generational AI chatbot called Pi, short for personal intelligence. Pi is designed to serve as an AI-powered personal assistant with generative AI technology akin to ChatGPT that will support planning, scheduling, and information gathering. This allows Pi to communicate with users via dialogue, making it possible for people to ask queries and offer feedback. Among other things, Inflection AI has outlined specific user experience objectives for Pi, such as offering emotional support.

At present, Inflection AI operates a cluster based on 3,584 Nvidia H100 compute GPUs in Microsoft Azure cloud. The proposed supercomputing cluster would offer roughly six times the performance of the current cloud-based solution.

In a recent interview with Wired, Google DeepMind CEO Demis Hassabis explained that he and his team expect their next development to take over ChatGPT in its breadth and capability. The new AI development will be known as Gemini and integrate with algorithms used by AlphaGo that successfully defeated a Go champion in 2016. 

Gemini is still in the developmental phase and is planned to be a large-scale text-based language learning model that works much like ChatGPT-4. However, this system will involve newer algorithms and learning models that are anticipated to take Gemini beyond ChatGPT's current capabilities. Just earlier in April, DeepMind merged with Brain—another AI lab operated by Google. This appears to be another step in the same direction.

Hassabis goes on to confirm the development process will require a fair bit of time before it's ready to release. He expects it will likely be a matter of months before anything is unveiled. This project is a considerable undertaking with a hefty price tag to match. Hassabis suggests it could cost hundreds of millions to complete their targets.

The DeepMind team is also experimenting with how Gemini could learn. They hope that by creating new models and ways to process information, they will be able to expand its capabilities. Hassabis mentioned fields like robotics and neuroscience, as well as the ability to learn from the physical world, as ways in which it could become an even more vital tool.

One area that Hassabis expressed concern about is the necessity to implement more research into tests. AI can be tricky to evaluate, and finding ways to investigate its scope and performance will aid in understanding its capabilities and how well engineers can control the AI. This is crucial to preventing the unintended development of dangerous threats like malicious software using their tools.

The road ahead is rocky and uncertain, but plans are underway to enhance AI while hopefully reigning it into a manageable system. Currently, all we can do is wait and take advantage of the systems we have in the meantime. Check out the original article shared by Wired to see what else Hassabis had to say about the upcoming Gemini project.

Elon Musk has confirmed that his companies Tesla and Twitter were buying tons of GPUs when asked to confirm whether he was building up Twitters compute prowess to develop a generative artificial intelligence project. Meanwhile, the Financial Times reports that Musk's AI venture will be a separate entity from his other companies, but it could use Twitter content for training.

Elon Musk's AI project, which he began exploring earlier this year, is reportedly separate from his other companies, but could potentially use Twitter content as data to train its language model and tap into Tesla's computing resources, according to Financial Times. This somewhat contradicts the earlier report which claimed that the AI project would be a part of Twitter. 

To build up the new project, Musk is recruiting engineers from top AI companies, including DeepMind, and has already brought on Igor Babuschkin from DeepMind and approximately half a dozen of other AI specialists.

Musk is also reportedly negotiating with various SpaceX and Tesla investors about the possibility of funding his latest AI endeavor, according to an individual with firsthand knowledged about the talks, which may confirm that the project is not set to be a part of Twitter.

In a recent Twitter Spaces interview, Musk was asked about a report claiming that Twitter had procured approximately 10,000 of Nvidia compute GPUs. Musk acknowledged this stating that everyone, including Tesla and Twitter, are buying GPUs for compute and AI these days. This is true as both Microsoft and Oracle have acquired tens of thousands of Nvidia's A100 and H100 GPUs in the recent quarters for their AI and cloud services. 

"It seems like everyone and their dog is buying GPUs at this point," Musk said. "Twitter and Tesla are certainly buying GPUs."

Nvidia's latest H100 GPUs for AI and high-performance computing (HPC) are quite expensive. CDW sells Nvidia's H100 PCIe card with 80GB of HBM2e memory for as much as $30,603 per unit. On Ebay, these things sell for over $40,000 per unit if one wants this product fast. 

Recently Nvidia launched its even more powerful H100 NVL product that bridges two H100 PCIe cards with 96GB of HBM3 memory on each one for an ultimate dual-GPU 188GB solution designed specifically for training of large language models. This product will certainly cost well above $30,000 per unit, though it is unclear at which price Nvidia sells such units to customers buying tens of thousands of boards for their LLM projects.

Meanwhile, the exact position of the AI team in Musk's corporate empire remains unclear. The renowned entrepreneur established a company called X.AI on March 9th, Financial Times reported citing business records from Nevada. Meanwhile, he recently changed the name of Twitter in the company's records to X Corp., which may be a part of his plot to build an 'everything app' under the 'X' brand. Musk is currently the sole director of X.AI, while Jared Birchall, who happens to manage Musk's wealth, is listed as its secretary. 

The rapid progress of OpenAI's ChatGPT, which Elon Musk co-founded in 2015 but no longer is involved with, reportedly inspired him to explore the idea of a rival company. Meanwhile, this new AI venture is expected to be a separate entity from his other companies possibly to ensure that this new project will not be limited by Tesla's or Twitter's frameworks.

Despite advocating for an industry-wide halt to AI training, Elon Musk has reportedly kicked off a major artificial intelligence project within Twitter. The company has already purchased approximately 10,000 GPUs and recruited AI talent from DeepMind for the project that involves a large language model (LLM), reports Business Insider.

One source familiar with the matter stated that Musk's AI project is still in its initial phase. However, acquiring a significant amount of additional computational power suggests his dedication towards advancing the project, as per another individual. Meanwhile, the exact purpose of the generative AI is unclear, but potential applications include improving search functionality or generating targeted advertising content. 

At this point, it is unknown what exact hardware was procured by Twitter. However, Twitter has reportedly spent tens of millions of dollars on these compute GPUs despite Twitter's ongoing financial problems, which Musk describes as an 'unstable financial situation.' These GPUs are expected to be deployed in one of Twitter's two remaining data centers, with Atlanta being the most likely destination. Interestingly, Musk closed Twitter's primary datacenter in Sacramento in late December, which obviously lowered the company's compute capabilities. 

In addition to buying GPU hardware for its generative AI project, Twitter is hiring additional engineers. Earlier this year, the company recruited Igor Babuschkin and Manuel Kroiss, engineers from AI research DeepMind, a subsidiary of Alphabet. Musk has been actively seeking talent in the AI industry to compete with OpenAI's ChatGPT since at least February. 

OpenAI used Nvidia's A100 GPUs to train its ChatGPT bot and continues to use these machines to run it. By now, Nvidia has launched the successor to the A100, its H100 compute GPUs that are several times faster at around the same power. Twitter will likely use Nvidia's Hopper H100 or similar hardware for its AI project, though we are speculating here. Considering that the company has yet to determine what its AI project will be used for, it is hard to estimate how many Hopper GPUs it may need. 

When big companies like Twitter buy hardware, they buy at special rates as they procure thousands of units. Meanwhile, when purchased separately from retailers like CDW, Nvidia's H100 boards can cost north of $10,000 per unit, which gives an idea of how much the company might have spent on hardware for its AI initiative.

Non-X Marks the Spot

10/20/2020 Update:  The AMD Ryzen 5 3600 is still an impressive CPU, but it will soon be supplanted by newer Ryzen 5000 processors. If you find the 3600 processor on sale at a steep discount during Black Friday or over the holidays, it's still worth considering. Just know that this CPU does not include AMD's latest Zen 3 architecture. So if you want the best single-core performance and other features that come with AMD's newest CPUs, you should probably spend more for a Ryzen 5 5600X when it arrives in late 2020. Those new chips have now taken over the top ranks on our CPU Benchmark Hierarchy.

AMD's value proposition has always been straightforward -- more for less. While we typically think of AMD offering more CPU cores than Intel for less money, the strategy also applies to the company's unrestrained feature sets for each processor, regardless of price. That includes in-box coolers, Hyper-Threading (AMD calls it SMT), and unlocked multipliers that enable easy overclocking, all of which are features that Intel either leaves out or disables on some of its chips in the name of segmentation.

Instead of squeezing out extra dollars from its customers, AMD gives you the same basic underlying features with the $199 six-core 12-thread Ryzen 5 3600 that it gives you with its full-fledged counterpart, the $249 Ryzen 5 3600X that we recently named the best mid-range processor on the market. That means the Ryzen 5 3600 has the same six-core 12-thread design, 32MB of L3 cache, and access to 24 lanes of PCIe 4.0, with the only tradeoff being a step back to the 65W Wraith Stealth cooler, while the 3600X comes with the more-capable 95W Wraith Spire cooler.

What does that mean to you? While the Ryzen 5 3600 is a great processor that packs a wonderful amount of performance into a 65W TDP envelope, a boon for small form factor enthusiasts, you can also overclock it and attain similar performance in many applications, like gaming, to the Ryzen 5 3600X (one of our best CPUs). But you save fifty bucks in the process while still getting class-leading features, like the PCIe 4.0 interface.

This follows the same AMD trend we’ve seen in the past, with overclockability making the non-X models a better value for enthusiasts than the pricier X-series models. But if you’re chasing the absolute highest frame rates you can get out of a six-core processor, be aware that the Ryzen 5 3600 chips might not reach the peak overclocking speeds of 3600X models. In either case, the solid blend of features and overclockability makes the Ryzen 5 3600 the clear choice for enthusiasts looking for a great value on a mid-range processor.

AMD isn't sitting still though: The company recently released its own new flagship, the 16-core 32-thread Ryzen 9 3950X, to fend off Intel's new challengers. That chip slots into a much higher tier than the 3950X, but it brings competitive gaming performance and much more threaded horsepower for those looking for the ultimate in performance. 

Ryzen 5 3600

Like the other Ryzen 3000 chips, the six-core 12-thread Ryzen 5 3600 comes with a 7nm compute die (with two disabled physical cores) paired with a 12nm I/O die. These two components come together into a single package that fits inside a 65W TDP envelope, making it physically identical to the 95W Ryzen 5 3600X.

The Ryzen 5 3600 has slightly lower clock speeds than the 3600X, with its 3.6 GHz base and 4.2 GHz Precision Boost 2 frequencies, a difference of 200 MHz in both measurements.

The 3600’s 4.2 GHz boost frequency is lower than the $192 Core i5-9500’s 4.4 GHz boost, but its 3.6 GHz base frequency equates to a 600 MHz advantage that, paired with AMD's drastic improvement to its instruction per cycle (IPC) throughput, will equate to higher performance in heavy workloads, not to mention the six additional threads of the AMD part. It’s notable that, unlike the previous-gen Ryzen models and Intel’s chips, AMD only guarantees the peak boost frequency on one core, while other cores could have lesser capabilities. Head to our Not All Ryzen 3000 Cores are Created Equal article for more information on that front.

Compared to the $182 Core i5-9400F, the 3600 has an 800 MHz base and 100 MHz boost frequency advantage. The Ryzen 5 3600 comes with a bundled 65W Wraith Stealth cooler, and while both the Core i5-9500 and -9400F come with stock coolers, they are of significantly lower quality. However, both of the Intel processors come with integrated graphics, while the Ryzen 5 3600 requires a discrete graphics card. If you’re not planning on incorporating a discrete GPU in your build, the Intel processors are the obvious choice.

The Ryzen 5 3600 comes with a healthy 32MB of total L3 cache, a neat doubling of capacity over its predecessor and more than three times the cache of the -9500 and -9400F. That does come with a few caveats, however, as cache performance and efficiency has a big impact on how much cache capacity benefits the processor in typical applications. As usual, our benchmarks will tell the tale.

The Ryzen 5 3600 drops into the AM4 CPU socket on the new X570 motherboards, which you'll need for official support for the PCIe 4.0 interface. But those new boards are more expensive than previous-gen models and aren't a good fit for value chips like the Ryzen 5 3600. Luckily, you can also use an older 400-series motherboard (B450 is a good fit) as a value alternative. But if you go that route you'll lose access to PCIe 4.0, which is one of the key selling points of the new processors.

Ryzen 3000 chips officially support dual-channel DDR4-3200, a step up from the previous-gen's support for DDR4-2966. AMD has greatly improved its memory compatibility and overclocking capabilities, but you still have to abide by rules that dictate the maximum supported frequency based on DIMM type and slot population. Ryzen 3000 also supports memory overclocking, either by hand-tuning or one-click A-XMP profiles with pricier kits, to skirt those rules.

AMD also has its Precision Boost Overdrive (PBO) feature on offer, which is an automated overclocking tool that will tune your processor to its maximum achievable performance based on its cooling, motherboard, and power delivery accommodations. The quality of your cooling solution, and the vagaries of the silicon lottery, have a big impact on how well PBO can auto-tune your processor.

MORE: Best CPUs
MORE: CPU Benchmarks Hierarchy
MORE: All CPUs Content

Overclocking and Test Setup

AMD's Ryzen 3000 processors have drastically improved single-threaded performance, but you'll lose that benefit if you manually overclock. That's largely because the chips can't be manually overclocked on all cores to reach the same frequency as the single-core boost frequency. In fact, we often find the all-core overclock ceiling to be 200 to 300 MHz lower than the rated boost speeds, which is likely due to the AMD’s new binning strategy that finds the Ryzen 3000 chips with a mix of both faster and slower cores.

We've tested several of the Ryzen 3000 processors in manually-overclocked configurations, and the results are predictable: You gain some extra threaded performance over automatic overclocking with PBO, but lose too much performance in lightly-threaded apps to make it worthwhile. In other words, outside of a few edge cases, like systems that will only do heavily-threaded work, manual overclocking simply isn't worth your time -- or the egregious power consumption it requires for relatively small performance gains.

As we've seen, AMD’s PBO algorithms provide a speedup that improves threaded performance while preserving the single-core boost frequency. The feature also keeps the Ryzen processor in its power-to-performance sweet spot, which means that it doesn't require much additional power consumption or cooling. Unfortunately, PBO gains are slight (don't expect miracles), but it is worthwhile if you have adequate cooling. As we've found in the past, AMD's stock coolers tend to extract most of the full benefit.

AMD's Precision Boost Overdrive (PBO) is an adaptive overclocking approach that allows the processor to communicate with the platform to modulate performance based on the motherboard's power delivery subsystem and thermal dissipation capabilities. The processor monitors Package Power Tracking (PPT), which is total socket power, and the Thermal Design Current (TDC) variable, which is the motherboard's maximum available sustained current. Electrical Design Current (EDC) also indicates the maximum current possible from the VRMs during peak/transient conditions.

AMD enables two options for PBO: IPM is AMD's default PBO setting, which is activated if you leave the PBO setting to 'Auto' in the Godlike's UEFI (our testing board). But you can select 'Enabled' to activate a profile that's dictated by the maximum limits of the motherboard's power delivery subsystem. These limits vary by motherboard and are defined by the vendor. We chose the latter to unlock the full potential of PBO. This setting kicks the socket's maximum power delivery up to 1000W to offer the best of increased multi-core boost clocks while retaining the high single-core boost clocks. You won't need anything near that much power delivery, but we don't want to leave any performance on the table.

AMD says you can also further tune the chip with an Auto OC (AOC) feature. AMD designed the new feature to give you some control over the maximum attainable boost clocks by allowing you to add up to an extra 200MHz to the maximum boost clock, but it isn't guaranteed that the processor will reach those speeds at all times, or under all conditions.

Unfortunately, we've found that the PBO+AOC feature often comes at the expense of performance in single-threaded workloads even though it is billed as retaining, and even heightening, single-core boost clocks. In most cases, the feature has no observable impact.

Security Mitigations

The new AMD-optimized Windows scheduler is only present in Windows 10 1903 and promises to expose gains in several types of applications. As such, we updated our test image to the latest version of Windows 10 available (18362.207). All of our test results come from the aforementioned operating system and include all publicly available security mitigations and the latest motherboard firmware revisions. Intel is currently impacted by Spectre, Spectre v4, Meltdown, Foreshadow, Spectre v3a, Lazy FPU, Spoiler, and MDS, while AMD is only impacted by Spectre and Spectre v4. AMD has added hardware-based mitigations for both variants of Spectre, which should reduce the performance impact, but the requisite patches for both companies have performance penalties, which are reflected here in our testing.

Phoronix Benchmark

We added in several new tests from Phornix's open-source benchmark suite. While this suite is heavily focused on Linux test environments, the benchmark utility does have several powerful testing options for Windows systems, along with Apple OS X, GNU Hurd, Solaris, and BSD operating systems. The test also outputs deviation metrics that help ensure accuracy in our test results.We've integrated key tests, like GIMP productivity, web browser benchmarks, SVT-AV1 encoding, NAMD, and the build-llvm compile test.

MSI MEG X570 Godlike

We're using MSI's MEG X570 Godlike as our test platform for the second- and third-gen AMD processors. The pricey Godlike board retails for around $800, but has the 14+4+1-phase power delivery subsystem to support aggressive overclocking.

The MEG X570 Godlike comes with a few nifty accessories like a 10Gb “Super LAN” Ethernet card and a PCIe Gen 4 Xpander-Z M.2 expansion card. That lets you add two more M.2 drives to complement the three M.2 PCIe Gen 4 M.2 ports on the board. You also get four PCIe 4.0 x16 slots, an RGB Mystic Light Infinity II mirror over the IO shroud, and a tiny OLED screen, alongside the two-digit LCD display for error codes.

Comparison Products

Test System and Configuration

MORE: Best CPUs
MORE: CPU Benchmark Hierarchy
MORE: All CPUs Content

Power Consumption

Power consumption measurements are always a bit tricky. But as long as your 12V supply (EPS) readings, motherboard power supply sensor values, and voltage transformer losses plausibly coincide, everything is fine. Therefore, we're using pure package power to avoid possible influences from our motherboard. Results from the PWM controller are very reliable if you take them as averages over a few minutes.

We began with the non-AVX stress test in AIDA64 and found that the Ryzen 7 3600 drew 67W, beat only by the Core i5-9400F at 53W. With PBO active and the Corsair H115i cooler, the 3600 only drew 70W.

The y-cruncher benchmark computes pi using a heavy multi-threaded AVX workload and also generates a performance measurement that we can use for efficiency metrics. We also measure power with HandBrake in x264 and x265 flavors. The latter uses a heavier distribution of AVX instructions than the former, but both transcoders are great for stressing the processor with a real-world workload.

The small increases in the 3600's power consumption from overclocking equate to relatively minor performance improvements. At stock settings, AMD has tuned the processors right at the knee of the voltage/frequency curve where the chip provides the maximum frequency possible and great efficiency. This PBO configuration also retains some of those same characteristics, but that doesn't leave much headroom for explosive performance gains.

Like the Ryzen 5 3600X, the six-core Ryzen 5 3600 is basically an eight-core Ryzen 7 3700X, but with two cores disabled. However, the 3600’s lower 65W TDP envelope equates to less power consumption at stock settings although, after overclocking, the 3600 does come close to matching the 3600X. That leads to surprisingly similar power consumption measurements during our x265 and y-cruncher tests. 

We tested with both the stock cooler and the Corsair H115i to see how much extra cooling impacts the maximum performance the auto-overclocking algorithms can extract from the processor, and how that impacts power consumption. According to our measurements, the bundled Wraith Stealth cooler dissipates enough waste heat to nearly achieve the maximum amount of available performance in some applications, but it also doesn't offer much of a boost in others. A better cooler is a good investment if you're overclocking the 3600, but you won't need something as robust as our H115i (even a low-cost Hyper 212 should suffice).

Plotting power consumption over our performance measurements shows that the Ryzen 5 3600 is an incredibly efficient processor, giving a solid level of performance at impressively low power consumption. That low power consumption isn't all about your electricity bill, either; it also equates to a lower bar for your cooling solution.

MORE: Best CPUs
MORE: Intel & AMD Processor Hierarchy
MORE: All CPUs Content

Test Notes

Test results annotated with "PBO" reflect performance with AMD's auto-overclocking Precision Boost Overdrive feature activated. As noted in the charts, we tested the overclocked Ryzen 5 3600 with two cooling solutions, the Corsair H115i watercooler and the bundled Wraith Stealth cooler. The overclocked previous-gen Ryzen 5 2600X offers roughly the same performance as an overclocked Ryzen 5 2600, so consider this model as a stand-in for its cheaper counterpart.

VRMark, 3DMark

We aren't big fans of using synthetic benchmarks to measure performance, but 3DMark's DX11 and DX12 CPU tests provide useful insight into the amount of horsepower available to game engines.

The DX11 and DX12 CPU test results expose the full-threaded heft of the Ryzen 3000 series processors, so there are no surprises. You'll notice that, even after overclocking, the Ryzen 5 3600 just barely matches the stock 3600X's performance, and it looses by a decent margin to the overclocked 3600X. It is clear that AMD's binning makes more of a difference with the Ryzen 3000 processors. 

That said, even at stock settings, the Ryzen 5 3600 offers enough threaded horsepower to rival even the agile overclocked Core i5-9600K. The more cost-comparable 6C/6T Core i5-9400F can't contend with the 3600's twelve threads.

We compare the Precision Boost Overdrive (PBO) auto-overclocking with the bundled Wraith Stealth and the beefier Corsair H115i cooler, which unlocks 2% and 1.5% more performance in the DX12 and DX11 tests, respectively.

The VRMark test benefits heavily from per-core performance, and the Ryzen 3000 processors have made great strides compared to the first- and second-gen models. The Ryzen 5 3600 beats the Core i5-9400F by ~15 FPS, while tuning with a capable cooler gives us ~5 more FPS. 

Ashes of the Singularity: Escalation

Ashes of the Singularity: Escalation is a computationally intense title that scales well with thread count, but clock speeds and per-core performance play a big role. The Core i5-9400F falls behind the 3600 by 7.2 FPS and the $262 i5-9600K by 5.5 FPS. We typically expect Intel processors to take the lead after overclocking; they are much more capable in that regard than AMD's processors. But the tuned twelve-threaded Ryzen 5 3600 scores within 0.1 FPS, which is essentially a tie.

MORE: Best CPUs
MORE: Intel & AMD Processor Hierarchy
MORE: All CPUs Content

Civilization VI AI, Stockfish Test

Stockfish, an open-source chess engine, is the perennial world leader in computer chess competitions, beating other engines like Google's Deepmind AlphaZero engine. The engine is designed to extract the utmost performance from many-core chips, so it scales well up to 512 cores. As we can see, that equates to a big win over both the Core i5-9400F and the i5-9600K as the engine unleashes the power of Ryzen 5's six extra threads.

Again, you'll notice that the overclocked Ryzen 5 3600 just manages to match the stock 3600X, but it lags behind considerably after tuning the latter.

Civilization VI's AI performance test is highly dependent on per-core performance, and AMD has made impressive steps forward compared to the stock Intel processors in the competing price ranges. However, Intel still holds the overclocking advantage, so it takes the uncontested lead after tuning.

Civilization VI Graphics Test

Intel's -9400F once again finds itself trailing the Ryzen 5 3600, but while the -9400F isn't overclockable, you can enable one-click overclocking with the 3600 and unlock higher frame rates.

Warhammer 40,000: Dawn of War III

The Warhammer 40,000 benchmark responds well to threading, so the Ryzen 5 3600 contends with the pricey -9600K, and it even closes in on the overclocked 3600X after tuning.

MORE: Best CPUs
MORE: Intel & AMD Processor Hierarchy
MORE: All CPUs Content

Far Cry 5

The Ryzen 5 3600 lags the 3600X by 2.7 FPS at stock settings, and while tuning the 3600 improves its performance, it still lags the overclocked 3600X by the same amount.

The Core i5-9400F is faster than the 3600 at stock settings, but AMD's unlocked multiplier evens the score. 

Final Fantasy XV

We run this test with the standard quality preset to sidestep the impact of a bug that causes the game engine to render off-screen objects. This title scales well with additional cores and threads, and tuning brings the 3600 within 0.4 FPS of the overclocked 3600X, meaning they offer essentially the same performance after overclocking.

Meanwhile, the stock Ryzen 5 3600 handily beats the Core i5-9400F.

MORE: Best CPUs
MORE: Intel & AMD Processor Hierarchy
MORE: All CPUs Content

Grand Theft Auto V

Grand Theft Auto V favors Intel architectures and, more generally, multi-core designs with high clock rates. Here we see that the Ryzen 5 3600 benefits more from improved cooling than its X-series counterpart, which means that investing in a better cooler is a good idea if you're interested in overclocking the chip. After overclocking, the difference between the 3600 and 3600X is negligible.

Hitman 2

Hitman 2 unleashes Intel's overclocking advantage with the Core i5-9600K, but the Core i5-9400F remains mired at the bottom of the chart due to its locked multiplier. With the right cooling, the Ryzen 5 3600 has comparable performance to the 3600X after overclocking.

MORE: Best CPUs

MORE: Intel & AMD Processor Hierarchy

MORE: All CPUs Content

Project CARS 2

Although Project CARS 2 is purportedly optimized for threading, clock rates obviously affect this title's frame rates. Intel's per-core performance, which is a mixture of IPC and frequency, pays big dividends in this title. Here the -9400F grapples with the stock 3600 models, but tuning again proves beneficial for Ryzen.

The Division 2

Here we see the two overclocked 3600 configurations offer the same performance regardless of the cooler. We also noticed the same trend with the 3600X in this title.

World of Tanks enCore

We can add Worlds of Tanks to the list of titles that respond extremely well to overclocking Intel's chips. The Ryzen 5 3600X offers much higher performance at stock than the 3600, and after tuning, we still see a similar delta.

MORE: Best CPUs

MORE: Intel & AMD Processor Hierarchy

MORE: All CPUs Content

Web Browser

Browsers tend to be impacted more by the recent security mitigations than other types of applications, so Intel has taken a haircut in these benchmarks of fully-patched systems.

The ARES-6 web browser benchmark focuses on the latest and greatest JavaScript features, with a heavy focus on forward neural networks used for machine learning tasks, and browser responsiveness. The 3600 profits from AMD's diligent work on improving IPC with the Zen 2 microarchitecture. Intel's faster processors exploit the company's frequency advantage, which equates to higher per-core performance, but the Core i5-9400F's lower base and boost clocks trail the Ryzen 5 3600.

We such much of the same with Speedometer 2 and Jetstream 2, with the Ryzen 5 3600 opening up a wider gap between itself and the -9400F. WebXRPT 3 puts the cap on the 3600's full sweep of the -9400F.

You'll notice that overclocking the Ryzen processors doesn't yield any improvement. That's because the processor is still limited to its maximum single-core boost speed during these lightly-threaded tasks.

Microsoft Office

The Microsoft Office suite of benchmarks runs via PCMark 10's new application test. This benchmark tests with real Microsoft Office applications, and we can see that the Ryzen 3000 series processors are very competitive in Excel, the Edge browser, and Word.

The Ryzen 5 3600 proves to be an agile performer in these common applications, even beating the speedy Core i7-9700K and i5-9600K in the overall score. We see some gains via overclocking Ryzen chips, and although the gains aren't as pronounced in these applications, the 3600 lands in a virtual tie with the Ryzen 5 3600X.

Productivity

The LLVM compiler benefits from extra threads, handing the 3600 an easy lead over the competing Intel chips, even after overclocking. The Core i5-9400F even trails the first-gen Ryzen 5 1600X, highlighting how much Intel's feature-trimming (in this case, Hyper-Threading) hamstrings its lineup against AMD's more well-rounded chips.

The application start-up metric measures load time snappiness in word processors, GIMP, and Web browsers under warm- and cold-start conditions. Other platform-level considerations affect this test as well, including the storage subsystem. The Ryzen 5 3600 actually loses a little performance when we kick on PBO with the Wraith Stealth cooler, but the result lands within the expected 3% variance of this test, implying there is no uplift from overclocking in some applications with the stock cooler. We see this trend repeat a few times throughout our test suite. In any case, the Ryzen 5 3600 is incredibly competitive given its price point, easily beating the -9400F.

Our video conferencing suite measures performance in single- and multi-user applications that utilize the Windows Media Foundation for playback and encoding. It also performs facial detection to model real-world usage. Here we see the 3600/Stealth combo beat the H115i-equipped configuration, but the deltas are small enough to chalk up to expected run-to-run deviation.

The photo editing benchmark measures performance with Futuremark's binaries using the ImageMagick library. Common photo processing workloads also tend to be parallelized, which plays well to Ryzen's multi-threaded design. Once again, we see very little uplift from overclocking with the Wraith Stealth cooler.

MORE: Best CPUs

MORE: Intel & AMD Processor Hierarchy

MORE: All CPUs Content

Rendering

As we can see throughout this series of tests, AMD's Ryzen processors undoubtedly sit atop the multi-threaded benchmark throne. The six-core -9600K and -9400F, which both lack Hyper-Threading, are outclassed in these tasks. Meanwhile, the 12-threaded Ryzen 5 chips dominate in these types of workloads with convincing wins across the board. The Ryzen 5 3600X at stock settings is comparable, or faster, than the overclocked Ryzen 5 3600 in most of these heavily-threaded workloads, meaning it does offer some benefit for the extra $50 investment, particularly if you're not interested in overclocking.

Encoding

The SVT-AV1 encoder is an Intel- and Netflix-designed software video encoder that became available earlier this year. This new encoder is more scalable than other encoders, thus offering faster performance paired with efficient compression. While it may seem counter-intuitive to use an Intel-designed encoder for testing AMD processors, consider that most encoders are inherently reliant upon per-core performance, which is a strength of Intel, while SVT-AV1 exposes the power of threading, a strength of Ryzen. At stock settings, the Ryzen 5 3600X beats the overclocked Ryzen 5 3600 by a slim margin, but overclocking the X-series model opens a wider gap. The Ryzen 5 3600 fares better against the -9400F, which lags by a decent margin.

Our LAME and FLAC tests, like many encoders, rely heavily upon per-core performance. That means Intel's frequency advantage comes into play, allowing the -9600K to take the lead. Conversely, the lower-clocked -9400F suffers. The -9600K has the advantage at stock settings, but overclocking propels it to the top of the chart.

Intel processors traditionally leverage high frequencies to dominate the HandBrake x265 test, which relies heavily on AVX instructions, and the x264 test. But Intel's higher clock speed isn't too much of an advantage in these tests when the similarly-priced competition has twice the number of threads, so the Ryzen 5 chips carve out nice leads in both x265 and x264 encoding. We also noticed that the Wraith Stealth-cooled 3600's overclocked configuration often trails the stock settings with the same cooler, which we confirmed with repeated testing. It looks like AMD might have some tuning left to do with its PBO algorithms in thermally-limited situations.

Compression, Decompression, Encryption, AVX

Our threaded compression and decompression 7-Zip and ZLib tests work directly from system memory, removing storage throughput from the equation. The combination of Ryzen 5 3600's improved memory subsystem and generous helping of cores helps it take an easy lead over the -9600K and -9400F while the Ryzen 5 3600X does provide at least some advantage over the 3600 via overclocking, although it isn't enough to justify a $50 premium for value-seekers.

We can also see the vast improvement in Ryzen's AVX performance in the y-cruncher tests: That's a massive generational leap, particularly in single-threaded performance. You'll notice that the overclocked 3600 with the Stealth cooler again suffers during a heavily-threaded AVX workload, furthering our suspicion that the PBO algorithms aren't fully optimized. In either case, the improved gen-on-gen AVX performance is truly impressive and benefits a wide range of applications.

MORE: Best CPUs

MORE: Intel & AMD Processor Hierarchy

MORE: All CPUs Content

Conclusion

After a simple one-click activation of the PBO automatic overclocking feature, the Ryzen 5 3600 offers much of the same performance as the Ryzen 5 3600X that retails for $50 more. AMD continues to offer the full complement of features with its lesser processors, like Hyper-Threading (SMT), overclockability, and capable stock coolers, making the Ryzen 5 3600 the value CPU to beat.

In the chart below, we plot gaming performance with both average frame rates and a geometric mean of the 99^th percentile frame times (a good indicator of smoothness). It's worth noting that AMD's previous-gen lineup is heavily discounted, so we’re departing from our standard practice of using official price lists. Instead, we’re using average pricing found online (temporary sales excluded). Volatility applies.

The Ryzen 5 3600 beats the Core i5-9400F in our gaming suite at stock settings, even coming within a mere 0.6 FPS of the $265 Core i5-9600K in the average of our 99th percentile results. The 3600 expands its lead over the Core i5-9400F after tuning, but the Core i5-9600K leverages its unlocked multiplier to take a big lead - but for $63 more. While we don't have one on hand, the $192 Core i5-9500 is a natural competitor with the 3600 as well, but 300 MHz of additional clock speed likely doesn't change the value proposition much.

It's clear the Ryzen 3000-series chips are the best bang for your gaming buck in this price segment. For gamers, the real choice boils down to the 3600 versus the 3600X. The Ryzen 5 3600X does have a bit more performance in the tank than the Ryzen 5 3600 in the overclocking department, which largely boils down to binning. It's also faster when we compare the two processors at stock settings, but the Ryzen 5 3600 slightly exceeds the stock X-model after tuning. In either case, the difference between the two chips boils down to a few FPS, which isn't worth the extra $50 unless you're chasing every last drop of performance. Also, bear in mind these deltas essentially vanish if you're gaming at higher resolutions.

The picture is a bit different when we switch over to productivity workloads. In threaded apps there really is no contest again: The Ryzen 3000 processors offer far more value than Intel's competing chips, and the lack of Hyper-Threading makes this a no contest for threaded applications. In gaming, we recorded slim differences between the overclocked Ryzen 5 3600 with the stock cooler and the Corsair H115i, but that delta widens in heavily-threaded tests. In fact, we noticed a few regressions with the stock cooler during overclocking, suggesting the all-aluminum cooler may get a bit overwhelmed when more voltage is put to the chip in heavily threaded workloads. As a result, the Corsair H115i cooler extracts more performance, especially in the AVX workloads, but the deltas are slight in most areas. Given the 3600's relatively low power draw, you could top it with a much lesser cooler, like a Hyper 212 Black, and get the same benefit. It really just boils down to how much noise you're willing to tolerate, but a beefy dual-radiator cooler is overkill.

If you need integrated graphics, the Ryzen 5 3600 and 3600X aren't for you. However, if you plan on using a discrete graphics card, the Ryzen 5 3600 is hands-down the best value on the market. The 3600X might be worth the extra coin if you aren't interested in overclocking, as it does provide more performance out of the box and comes with a better cooler. However, it's hard to justify the $50 premium over the Ryzen 5 3600. For small form factor (SFF) enthusiasts, AMD has packed in quite a bit of punch into a 65W envelope, giving it the uncontested lead for small systems.

A Ryzen 5 3600 paired with a B450 motherboard will make a great setup for mainstream gamers, and you have the option to upgrade to a PCIE 4.0-capable motherboard in the future. We aren't sure when, or if, AMD's partners will push out new B-series motherboards, but that could be a compelling upgrade path in the future.

Image Credits: Tom's Hardware

MORE: Best CPUs

MORE: Intel & AMD Processor Hierarchy

MORE: All CPUs Content

Developers often include dimwitted bots in training modes, single-player recreations of multiplayer game modes, and other parts of a game. These automatons rarely exhibit good decision making; most of us have probably watched bots run into walls or spin in circles. DeepMind's artificial intelligence is a bit more sophisticated, and on May 30, the company revealed that its bots can beat human players in Quake III's Capture the Flag mode.

DeepMind is no stranger to developing machines capable of serving humans a heaping helping of humble pie. The most well-known example is AlphaGo, which beat the world's best Go player so handedly that he said competing against the AI was like playing in his backyard while AlphaGo explored the universe. The company has also worked to teach AI agents to play StarCraft II. (Another organization, OpenAI, taught its agents Dota 2.)

Quake III is wildly different from Go. It's a three-dimensional game played in a first-person perspective, to start, and its Capture the Flag game mode is team-based rather than a one-on-one competition. DeepMind also chose a variant of the mode that relies on procedurally generated maps, which meant the agents couldn't simply learn the best solution for a pre-defined space, and instead had to learn and master the principles of Capture the Flag itself.

To give you an idea of how well the agents learned how to play Quake III: DeepMind dubbed these agents For The Win (FTW) and they definitely earned that moniker. Despite knowing nothing about Capture the Flag to start, they learned the rules of the game as well as the core principles for success pretty quickly. Eventually they overtook 40 human players in a tournament where agents and humans played with and against each other. For science.

DeepMind conceded that some of FTW's success could be attributed to the agent's literally inhuman reaction times. (Again: most of us have probably wondered how the heck a bot managed to kick our butts no matter how hard we tried to win.) So it slowed them down and, sure, enough, the agents still beat their human counterparts. To rub salt in the wound, the company said most players also rated FTW as more collaborative than humans.

That means it's probably time to stop calling bad players bots. FTW showed that AI could learn the rules of a game, grok the fundamentals, and then get so good at the game that it beat humans even when it was handicapped with reaction times comparable to our own. Since these bots are actually better teammates than most humans, too, it seems like calling someone a bot could actually be a compliment once these agents go mainstream.

What It Means for AI

None of which is to say that FTW's success at Capture the Flag--and later other Quake III game modes--was just fun and games. DeepMind said that the work on this project "highlights the potential of multi-agent training to advance the development of artificial intelligence: exploiting the natural curriculum provided by multi-agent training, and forcing the development of robust agents that can even team up with humans."

The company also noted that FTW learned how to play Capture the Flag in (perhaps surprisingly) human-like ways:

"Through unsupervised learning we established the prototypical behaviours of agents and humans to discover that agents in fact learn human-like behaviours, such as following teammates and camping in the opponent’s base. [...] These behaviours emerge in the course of training, through reinforcement learning and population-level evolution, with behaviours–such as teammate following–falling out of favour as agents learn to cooperate in a more complementary manner."

More information about this project is available via DeepMind's website, a paper that was published in Science, and a YouTube video on the subject from 2018.

A report by a House of Lords (upper Parliament chamber) committee recognized that AI advancements will not be without risks and, as such, ethics will need to play a vital role in the development of AI.

AI For The Public’s Benefit

The House of Lords report said that the number one priority in development AI should be that the AI is developed for the common good and benefit of humanity. This seems to mirror UK-based DeepMind’s own first AI principle. However, its now sister-company (under the Alphabet group) Google may be of a different opinion, as it seeks to help the U.S. government in the creation of autonomous drones.

Another principle for AI code as established by the UK report is that AI should never have the autonomous power to “hurt, destroy or deceive human beings.” This also seems to be a contrary principle to that of the U.S. government, which is looking to build drones that will decide on their own when and who to kill.

Another important principle is that AI should not diminish the data and privacy rights of individuals, families, or communities. This principle also seems to be in antithesis with how big tech companies have been using AI so far, trying to collect as much user data as possible. The committee report also warned against monopolization of data by big tech companies.

Other two principles resulted from the report say that AI should be intelligible and fair, and that citizens should have the right to be educated and flourish mentally alongside AI.

UK Committee’s Recommendations

In order to avoid data monopolization by big companies, governments will have to encourage more competition for AI solutions. Additionally, big companies will have to be investigated over how they use data. The UK Parliament also believes that liability in case of AI harm (such as self-driving car crashes caused by bad software) is not yet properly defined and that new laws may be needed to fix this.

The UK Parliament also drew attention to the issue of not enough transparency when AI solutions are used. The UK committee believes that people should know when AI was used to make significant or sensitive decisions.

The House of Lords committee warned against biases in AI, too, and recommended that AI specialists are recruited from a diverse background

The committee also said that individuals should have greater control of their data and how it’s used. The way in which data is collected by companies will also need to be changed through legislation, new frameworks, and concepts such as data portability and data trusts.

Recently, OpenAI talked about its new discovery, in which using “self-play” mechanisms to train AI agents resulted in significantly more advanced but also less complex neural networks. OpenAI noted that self-play will have a major role in the training of AI in the future, not only because the AI doesn’t need large data sets from which to learn anymore, but also because the AI can improve and achieve mastery for a given skill much more quickly.

No Human Intervention

DeepMind, an Alphabet company, has been experimenting with self-play, too. It used this mechanism to design a new version of AlphaGo, called AlphaGo Zero, which could learn the Go game from scratch simply by playing against itself.

The previous version of AlphaGo had to learn the Go game from “watching” thousands of recorded amateur and professional human games. Eventually, it figured out which were the best “winning moves” for most situations, and it would play those against any human player. However, for the most part, the older AlphaGo was still limited by human knowledge and experience. The new AlphaGo has no such limits.

Timeline For Mastery

AlphaGo Zero started out making completely random moves at first, while sparring with an equal-strength version of itself. It took three hours for the AI to achieve human amateur level of playing, where it would simply try to gain as many stones as possible.

Within 20 hours, it learned more advanced tactics that are typically used only by professional players. By the third day, AlphaGo Zero had already gotten to the level that the first iteration of the original AlphaGo was when it defeated Lee Sedol in four out of five games last year.

It then took 21 days to get to the level where it was when it defeated Ke Jie 3-0, earlier this year. At the time, Ke Jie called AlphaGo a “god of Go.”

By the 40th day, AlphaGo Zero could already beat the best iteration of the original AlphaGo 100-0. Suffice it to say that it’s unlikely any human could ever hope to beat AlphaGo Zero at this point.

Differences From The Original AlphaGo

The first version of AlphaGo included a small number of hand-engineered features so that it would work correctly. The new AlphaGo Zero knows only about the white and black stones and that it can move them anywhere on the board.

Additionally, the first AlphaGo used two neural networks: a “policy network” to select the next move to play, and a “value network,” which predicted the winner of the game from each position. The two are combined for AlphaGo Zero, allowing it to be trained more efficiently.

It also seems that AlphaGo Zero doesn’t need to use “rollouts.” These were fast, random games started from an existing position, so that the agent could calculate what was the best move to make at that point. AlphaGo Zero simply relies on the intelligence of its own neural network to tell it what is the best move to make in any scenario.

The first ever version of AlphaGo ran on 176 GPUs, with a TDP of 40,000W (40 kW). The newest versions needed only four Tensor Processing Units (TPUs) for training. This large increase in efficiency is due to both algorithmical changes in how AlphaGo works and the latest TPUs being much more efficient than older Nvidia GPUs.

Thousands Of Years Of Knowledge Achieved In Days

AlphaGo Zero showed that AI has gotten to the point where it could learn skills in a matter of days or weeks that took humans hundreds or thousands of years to master. DeepMind believes that this sort of technology could be a multiplier for human ingenuity.

Similar techniques could be applied to protein folding, reducing energy consumption, or searching for revolutionary new drugs and materials. All of these could ultimately have a large positive impact on society (especially if DeepMind can keep its AI under control).

Last year, DeepMind, a sister company of Google in the new Alphabet conglomerate, announced its new machine learning-based text-to-speech (TTS) system called “WaveNet.” The system represented a significant improvement in how natural the synthetic speech sounded, but it had one critical flaw: it was too computationally intensive to be used for anything other than research projects.

A year later, the DeepMind team changed this by making the system 1,000x faster, which means it’s now efficient enough to be used in consumer products, such as Google Assistant.

Old Models For Synthetic Speech Systems

The best way to do synthetic speech generation so far has been to use the concatenative TTS system, which uses a database of high-quality recordings from a single voice author. The recordings are split into tiny chunks that can then be combined or concatenated to generate the synthetic speech.

This is also why TTS systems have sounded so “robotic” for so many years. This system can’t easily be altered or improved without creating a whole new database of recordings, which is why progress in synthetic speech has been so slow over the years, too.

Another method—which has been even less common, because it sounds more robotic—is the parametric TTS system. With this system, voices are completely machine-generated based on grammar and mouth movement rules that are supposed to make the synthetic voice sound human.

The parametric system never worked quite as well as the concatenative TTS system because, as with other similar attempts, it’s been too difficult to program the complexity of human movements and actions through fixed algorithms and parameters. The reason deep learning has been so successful is because it does away with such human-programmed algorithms, and can instead generate its own parameters by “learning” how things are done by humans.

The Neural Network-Based WaveNet Model

The WaveNet system takes roots in machine-generated synthetic speech, but instead of using fixed parameters, it trains neural networks on large datasets of human speech samples so it can learn on its own how to generate “human-like” speech.

During the training phase, the neural network determined the underlying structure of speech, such as which tones followed each other and which were more realistic. It then synthesized one voice sample at a time, while taking into account the properties of the previous sample. The resulting voice contained natural intonation and even features such as lip smacks.

This approach doesn’t only generate more natural-sounding synthetic speech, but it should also be much easier to improve it in the future, as it will just be a matter of fine-tuning the neural network training or throwing more data and computational resources at it.

The new model also has the advantage of being easily modified to create any number of unique voices from blended datasets.

A 1,000x Increase In Performance

However good it was, WaveNet couldn’t have been deployed to real-world applications, even by a company such as Google. Therefore, the DeepMind team had to drastically improve the system’s performance before using it in any consumer products.

The original WaveNet could generate only 0.02 seconds of synthetic speech in 1 second. The new WaveNet is 1,000x faster, and it can now create 20 seconds of even higher-quality sound from scratch in 1 second.

The new WaveNet can generate 24kHz audio samples with 16-bit resolution for each sample (same resolution used for CD-quality music), while the older system could only generate 16kHz audio samples with 8-bit resolution for each sample.

According to human testers, the new WaveNet does generate higher-quality sounds. For the new U.S. English I voice, the mean-opinion-score (MOS) has increased from about 4.2 for the old WaveNet to about 4.35 for the new one. The human voice was rated at 4.67, so it’s getting rather close to the ideal.

DeepMind noted that the WaveNet system also gives it the flexibility to build synthetic voices by training them using data from multiple human voices. This can be used to generate high-quality and nuanced synthetic voices even when the voice datasets are small.

The new WaveNet has been launched into production in Google Assistant, and it’s also the first application to run on Google’s new cloud TPU chips. However, Google still seems to need more time to train voices for multiple languages, because for now only the English and Japanese languages take advantage of the new technology.

DeepMind announced the creation of a new “Ethics & Society” division inside the company, which will focus on ensuring that AI benefits society and doesn’t get out of control.

Avoiding Dangers Of AI

Some prominent personalities, including people such as Elon Musk, Stephen Hawking, and Bill Gates, have warned about the dangers of artificial intelligence if we let it run loose. We may still be decades away from AI gaining consciousness and then deciding to kill us all to save the planet, but it’s probably not a bad idea to start researching just how an AI should think and act, especially in relation to humans.

Besides the sci-fi dystopian future we can easily imagine, there are indeed some real dangers that AI can already create, even if not through its own fault but through the fault of humans who develop it. One such danger is that AI can develop, or rather replicate, human biases and then accelerate them to the extreme.

We’ve already seen something like this in action when Microsoft launched its own Twitter-based AI bot. In mere hours, what was otherwise a neutral technology and “intelligence” became a neo-Nazi racist AI, all due to much prodding and testing from humans in the real world. Fortunately for us, that AI was merely in charge of a Twitter bot, and it wasn’t in charge of a nuclear power’s defense systems.

However, obviously we can’t simply assume that AI will do good when left to its own devices, because it may incorporate ideas that its developers never thought it would (unless we believe Microsoft actually intended to build a neo-Nazi AI from the start).

There’s also the age-old idea of the “paperclip maximizer,” which says that an AI could simply follow its mission in a very rigid way (making as many paperclips as possible) to the point where that mission starts harming humans, even if the AI itself never had any harmful “thoughts.” It’s just that the AI would use all of our planet’s resources to build those paperclips, leaving us with nothing...except for a lot of paperclips.

Controlling AI

DeepMind’s AI technology is perhaps the most advanced in the world right now, having already proven that it can beat the best players in the world at a game that people thought AI could never conquer. It also proved to have more real-world uses such as cutting Google’s data center cooling costs by 40%, and the technology is being integrated into some UK hospital’s systems to improve healthcare.

The DeepMind team believes that no matter how advanced AI becomes, it should remain under human control. However, it’s not clear how true that will be in the future, because we won’t be able to monitor every single little action the AI takes. For instance, will a human always have to approve when an AI technology decides to switch to the green light or red light on a city’s streets? Probably not, as that would defeat the purpose of using an AI in the first place.

That one is an easy example, but what about having a human always approving what medicine patients should take? Perhaps this will be the default procedure in the beginning, but can we guarantee this will always be the case in the future? Perhaps hospitals will decide AI will have gotten smart enough 20 years from now that they will allow it to distribute 95% of the medicine to patients, without human supervision.

The bottom line is that it’s it’s not clear where to draw the line in the first place, and even if it was, it will likely be a moving goalpost, as AI keeps getting smarter. Somewhere along the way things could go wrong, and at the moment in time it may be too late to fix it, because we’ll have very little control over the AI.

In the hospital example above, the AI could, for instance, suffer from a software bug or a hack, and then distribute the wrong medicine to the whole hospital, while the human supervisors would trust the AI to do its routine job safely, just as it would have done thousands of times before.

They may not necessarily notice the wrong medicine in time, just like nobody would notice that a streetlight-managing AI turned the green light on too fast on some roads, because nobody would supervise these individual actions.

DeepMind’s Ethics & Society Group

To understand the real-world impacts of AI, the DeepMind team has started researching ethics for AI, so that the AI they build can be shaped by society’s concerns and priorities.

Its new Ethics & Society unit will abide by five principles, which include:

1. Social benefit. DeepMind’s ethics research will focus on how AI can improve people’s lives and how to build more fair and equal societies. Here, DeepMind also mentions previous studies showing that the justice system currently uses AI with built-in racial biases. The group wants to study this phenomena more so that the same biases won’t be built into its own AI or other AI systems in the future.
2. Evidence-based research. The DeepMind team is committed to having its papers peer-reviewed to ensure there are no errors in its research.
3. Transparency. The group promises not to influence other researchers through the grants it may offer and to always be transparent about its research funding.
4. Diversity. DeepMind wants to include the viewpoints of experts from other disciplines, too, outside of the technical domain.
5. Inclusiveness. The DeepMind team said that it will also try to maintain a public dialog, because ultimately AI will have an impact on everyone.

The DeepMind Ethics & Society division will focus on key challenges involving privacy and fairness, economic impact, governance and accountability, unintended consequences of AI use, AI values and morals, and other complex challenges.

DeepMind hopes that with the creation of the Ethics & Society unit, it will be able to challenge assumptions, including its own, about AI, and to ultimately develop AI that’s responsible and that’s beneficial to society.

DeepMind and Blizzard announced that StarCraft II is now an open artificial intelligence research environment where people can develop and train their own AI agents to beat others at the game. Unlike other game bots, the new AI agents will learn and see the game the same way human players do, with no programming shortcuts to give them unfair advantages.

DeepMind AI

DeepMind has tried from the beginning to build AI agents that can solve complex problems. Its AI agents started with playing Atari games and then moved on to beat the world’s top Go grandmasters, a feat that AI experts didn’t think was possible for at least another decade. In the meantime, DeepMind’s AI has also been used for real-world applications such as cutting Google’s data center cooling costs by 40%.

DeepMind’s latest project is to conquer the much more complex game of StarCraft II. If its AI agents can learn to regularly beat top human players in a strategy game such as StartCraft II with its own 3D worlds, the AI could then be used for even more advanced real-world applications.

Why Conquering "StarCraft II" Will Be Difficult

There are multiple reasons why a game such as StarCraft II is actually much more difficult for an AI agent to learn than the game of Go. One is that StarCraft II is significantly more open-ended than Go. It also has more gameplay rules, making it a much more complex game.

Another reason is that both Go players know exactly what’s happening within the game at any point in time, whereas in StarCraft II, the fog of war clouds what each player knows about their rival throughout the game. This should make StarCraft II quite unpredictable for the AI, although it’s possible the AI will eventually figure out the most common strategies employed by human players at any given point in a game.

The more game replays there are, the better the StarCraft II AI agents should become. StarCraft II is quite a popular game that's played competitively online, so there should be plenty of game replays that can be used as a data source.

StarCraft II also has 300 basic actions that you can take, as opposed to Atari games that only have ten (eg, up, down, left, right, etc.).

As seen in the video below, on the left side, an early-stage training agent is failing to even keep its workers mining, a task that’s trivial for human players. On the right side of the video, a trained agent can perform more meaningful actions, but according to DeepMind, it still fails to beat even the easiest built-in StarCraft II AI.

The DeepMind team seems to be well aware that that training the AI agents to play StarCraft II well is not a simple task. Therefore, they’re encouraging everyone to take advantage of both DeepMind and Blizzard’s tools to create their own agents, so that everyone can learn from the process.

The tools include:

• Machine learning API developed by Blizzard that gives researchers hooks into the game
• 65,000 game replays (more than half a million in a few weeks)
• Open source version of DeepMind’s PySC2 toolset to make it easier for developers to use Blizzard’s feature-level API
• A set of mini-games in which the researchers can test their AI agents
• A research paper written by DeepMind and Blizzard that reports initial results for for the test AI agents against the built-in StarCraft II AI

The Information Commissioner’s Office (ICO), which is the Data Protection Authority in the UK, ruled that the Royal Free NHS Foundation Trust failed to comply with UK’s Data Protection Act. According to the ICO, Royal Free gave DeepMind access to the data of 1.6 million patients without fully disclosing to them how it would be used.

Royal Free And DeepMind’s Deal

Royal Free entered a five-year deal with DeepMind, a British machine learning technology company that was acquired by Google in 2014, to collaborate on developing an application, called “Streams,” that would help doctors more accurately diagnose some diseases and improve treatment for patients. The app would also alert doctors and nurses when certain patients are at risk of getting ill based on the information it already has on them.

Inappropriate Legal Basis

UK’s National Data Guardian, which watches over how medical data is used, was the first agency to come to the conclusion that the deal between Royal Free and DeepMind was done on an “inappropriate legal basis.” The deal was supposed to cover access to patient data only for “testing” purposes, and not to be used for “direct care.”

Since then, ICO has done its own investigation, and it seems to have arrived at similar conclusions.

Elizabeth Denham, Information Commissioner, said:

There’s no doubt the huge potential that creative use of data could have on patient care and clinical improvements, but the price of innovation does not need to be the erosion of fundamental privacy rights.Our investigation found a number of shortcomings in the way patient records were shared for this trial. Patients would not have reasonably expected their information to have been used in this way, and the Trust could and should have been far more transparent with patients as to what was happening.We’ve asked the Trust to commit to making changes that will address those shortcomings, and their co-operation is welcome. The Data Protection Act is not a barrier to innovation, but it does need to be considered wherever people’s data is being used.

Required Changes

It doesn’t look like the ICO will fine Google and it seems to mainly hold Royal Free accountable for this privacy violation, as the hospital was the one controlling the data. Going forward, the ICO will want to see some changes in the deal. These will include establishing a proper legal basis for the sharing of patient data with DeepMind, requiring Royal Free to complete a privacy assessment of the deal, as well as commission a third-party audit of the trial.

Although DeepMind was not recognized as the main party at fault here, the company seems to have already taken some steps to improve the transparency of this collaboration between it and the NHS trusts. These include making the deal more transparent, as well as offering more details, and being more mindful about how the patients would be affected by the company’s processing of the data.

The company has also started working on Verifiable Data Audit technology, which uses a digital ledger to ensure the integrity of medical records and how or when others make use of that data. For instance, this could be used to verify if DeepMind itself or some other organization has been processing patient data without approval. DeepMind hopes to build an initial version of this digital ledger by the end of this year.

After 60 matches against top Go players online earlier this year under the name “Master,” and after recently defeating Ke Jie, the world’s best (human) Go player, twice, the AlphaGo AI has become the ultimate Go master. This latest development is proof that artificial intelligence may soon begin to surpass humans at other highly complex tasks.

The "God Of Go"

On Saturday, Ke Jie had his final match against AlphaGo by playing white, which gave Ke Jie a small advantage as the second mover. Ke Jie’s strategy this time around was to make the game as complex as possible, hoping he could force AlphaGo into making a mistake by poorly connecting its own stones.

In theory, this was quite a good strategy on Ke Jie’s part, because he was hoping AlphaGo is still just a machine that can’t “connect the dots” as well as humans can by looking at the bigger picture on the board. However, AlphaGo had both a strong local and global game, thus denying Ke Jie the opportunity to lure it into making a mistake.

AlphaGo may have “known” what Ke Jie was trying to do, so it may have made its moves taking into account those calculations, or it may have simply played the game as perfectly as possible, not allowing Jie any opportunity on the board.

Jie said in a post-game interview that AlphaGo sees the whole universe of Go (“Weiqi” in pinyin), while he could only see a small area around him. He also said it’s like playing Go in his backyard while AlphaGo explores the universe.

Before the matches, Jie presumed that AlphaGo--and AI in general--would be better with smooth transitions, as opposed to big clashes between itself and the opponent. However, in his second match especially, he noticed that the AlphaGo is not only dealing with big clashes easily now, but it even had much better solutions for getting out of them than humans do. In the same interview, Jie likened AlphaGo to a “God of Go.”

AlphaGo Gives Hope For The Future

After the game, Jie said that AlphaGo’s skill in beating humans so easily now gives hope for the future, because that skill may be harnessed for other tasks. Jie mainly hopes that AlphaGo’s technology can be used in medical science so it cure diseases or at least find better treatments for them.

DeepMind, a U.K. company Google acquired in 2014 for more than $500 million and the creator of the AlphaGo AI, has already started collaborating with the National Health Service (NHS) agency in UK to improve access to quality healthcare and the speed of care. The DeepMind technology may also invent new methods of diagnosis that could lead to earlier diagnostics of health issues.

Medical science is not the only area where DeepMind technology has been used. The technology is already turning up a profit for Google by cutting its data center cooling costs by up to 40%.

DeepMind technology has also been used to generate synthetic speech with near-human level voice quality. However, the company is still working on reducing the computing resources for such a task, so it has not yet been deployed at scale in Google’s products.

From Go To 'StarCraft 2'

DeepMind CEO Demis Hassabis said in a recent post that the final match between Ke Jie and AlphaGo was also the last time the company would organize a match event and that it would step back from competitive play. This has probably disappointed many players, as they may have wanted to take a shot at beating the “God of Go” in a tournament, however fruitless that effort may have been.

However, Hassabis did say that DeepMind would release 50 “special games” in which AlphaGo plays against itself. The company will also release a Go teaching tool that will be developed in collaboration with Ke Jie.

Even though DeepMind is looking towards more “serious” issues to tackle with its artificial intelligence, Go is not the last game its AI will try to master. StarCraft 2 seems to be the next in line.

In Go, the AI had “perfect information,” as it could “see” exactly what moves the opponent made. In StarCraft 2, players operate on much more limited information. They don’t know where each other are or what they’re building and training because of the fog of war, so the DeepMind AI has to make the best of that situation.

Unlike typical AI in games, which simply receives the information it needs from the game’s code, the DeepMind AI will have to learn StarCraft in the same way human players do: by watching others play and through trial and error. It will also be limited by how many actions it can do per minute, because humans are also limited by how fast their hands move on the keyboard.

DeepMind and Blizzard plan on opening up the project this year so other researchers can help improve the AI to the point where it can beat a top human player. DeepMind said that the “messiness” of the StarCraft 2 environment is quite close to the real world, so it should help its AI better understand how our world works. That could lead to new super-human problem-solving capabilities for the DeepMind AI, especially in the area of robotics, but potentially others, too.

A popular concept in science fiction is the singularity, a moment of explosive accelerating growth in technology and artificial intelligence that rewrites the world. One of the better explanations for how this could happen is described by the Scottish sci-fi author Charles Stross as “a hard take-off singularity in which a human-equivalent AI rapidly bootstraps itself to de-facto god-hood.”

To translate: If an AI is capable of improving (“boostrapping”) itself, or of building another, smarter AI, then that next version can do the same, and soon you have exponential growth. In theory this could lead to a system rapidly surpassing human intelligence, and, if you’re in a Stross novel, probably a computer that’s going to start eating people’s brains.

The singularity still seems to be a long ways off (until we crack Moore’s Law), but at Google I/O, we got a glimpse of our future robot overlords from Google CEO Sundar Pichai.

Lifelong Learning

The new technology is called AutoML, and it uses a machine learning system (ML) to make other machine learning systems faster or more efficient. Essentially, it’s a program that teaches other programs how to learn, without actually teaching them any specific skills (it’s the liberal arts college of algorithms).

AutoML comes from the Google Brain division (not to be confused with DeepMind, the other Google AI project). Whereas DeepMind is more focused on general-purpose AI that can adapt to new tasks and situations, Google Brain is focused on deep learning, which is all about specializing and excelling in narrowly defined tasks.

According to Google, AutoML has already been used to design neural networks for speech and image recognition. (Fun fact: The networks to accomplish these two tasks are usually nearly identical. Images are typically analyzed by looking at repeating patterns in pixels, and speech is analyzed by turning sound into a graph of frequency over time that’s analyzed the same way). Designed by AutoML, the image recognition algorithms were as good as those designed by humans, and the speech recognition algorithms were, as of February 2017, “0.09 percent better and 1.05x faster than the previous state-of-the-art model.”

Using AI to build machine learning systems has been a hot area of research since 2016, as researchers at Google, UC Berkeley, OpenAI, and MIT have worked to reduce the time needed to set up and test new neural architectures.

Google’s plan is not to bring about the AI apocalypse, but to lower the barrier to entry for companies interested in machine learning research or products. Instead of “automated” or “self-reinforcing,” Google’s AutoML software might be better said to be “self-assembling,” or “self-optimizing” (the term Berkeley researchers used for their similar algorithm).

No, That's A Civet

Traditional machine learning takes two main approaches. A computer is either fed thousands of labeled pieces of data (say, photos of a cat, and photos not of a cat), and eventually it will build a system to differentiate “cat” from “not a cat.” This system may be unique, and we won’t necessarily be able understand exactly how it’s working, but at the end it’ll reliably identify a tabby.

The other method is the way computers can be trained to solve more flexible problems, like an efficient way to walk or how to escape a virtual maze. The computer is given a set of parameters to work in as well as a failure condition. The computer will then be set to experiment. At first it will work at random, but as it rules out more and more failed attempts, it’ll zero in on a solution. A combination of these methods was how AlphaGo learned to play board games.

Automated machine learning is just one more layer of abstraction. Instead of learning how to identify a cat by examining thousands of cat photos, the algorithm is trying to build the most efficient system for learning to identify cats.

This has a few major benefits. One is that deep learning systems need to be tuned to the inputs they will be analyzing. Although any machine learning system improves itself over time, a poorly designed algorithm may never be as fast or as accurate as a well designed one, and designing an optimized algorithm is hard. Some programmers swear they operate by intuition, and no matter what the method, the task of tuning and refining an algorithm takes time and expertise. Also, a well designed algorithm may be quicker to train, and require less time and input until it’s proficient at a task.

Clouds on the Horizon

As you can probably imagine, the process of using a neural net to create and test a set of other neural nets is incredibly expensive in terms of time and computation. To create the image and speech recognition algorithms designed by AutoML, Google reportedly let a cluster of 800 GPUs iterate and crunch numbers for weeks.

This is likely not going to be a tool that you can run on your laptop, but it may become a selling point for Google Cloud. Access to AutoML and the ability to create and refine a machine learning system without a strong background in AI, could be a tool to give Google a leg up over Amazon, whose AWS cloud service Google has long trailed behind.

AutoML and similar tools may be the key to making machine learning accessible to a range of scientists and could help bring AI to new fields of study.

If it doesn’t eat our brains first, that is.

Baidu launched Deep Voice 2, the next generation of its neural text-to-speech technology. The new version is based on the same Deep Voice 1 pipeline, but it alleges a much higher performance and delivers significantly improved speech quality.

Improving On Deep Voice 1

When using Deep Voice 1, the company still needed about 20 hours of training for each voice. However, with the new improved Deep Voice 2 technology, only half an hour was needed to “train” a new voice.

This also allows the new system to scale to hundreds of different voices and accents. This could, for instance, make the “read-aloud” feature in many ebook reading applications more appealing due to all the unique and personalized voices you could choose when listening to an ebook.

The Deep Voice 2 technology can learn on its own, from scratch, all the shared qualities of different voices, and then it can imitate them.

“Deep Voice 2 can learn from hundreds of voices and imitate them perfectly,” said the company in a blog post.

Baidu uploaded some samples online, demonstrating the rather high quality of the voices, as well as the different accents that they use.

Outperforming DeepMind’s “Wavenet”

Wavenet is DeepMind’s “breakthrough” technology that has significantly reduced the gap towards human-like speech. The Mean Opinion Score (MOS) for DeepMind’s Wavenet was 4.21 for U.S. English, whereas for humans it was 4.55. For Mandarin, Wavenet’s MOS was 4.08, but humans scored 4.21.

However, despite Wavenet’s ability to produce near-human levels of speech quality, it also has one significant flaw at the moment, and that’s the high computing resources it requires to work. According to Baidu, it can take minutes or more to generate a few seconds of speech with a Wavenet-like technology. The company also said that its Deep Voice technology is capable of synthesizing speech up to 400 times faster than other models like Wavenet.

In a recent paper, Baidu also showed the MOS for its Deep Voice and Deep Voice 2 technologies, and it looks like Baidu’s technologies compromise on speech quality to achieve the higher performance. Deep Voice achieved an MOS of only 2.05 for single-speakers, while the new Deep Voice 2 received a score of 2.96, which is a 44% improvement in speech quality over the previous generation.

However, as you can see, Deep Voice 2 also has a significantly lower score than DeepMind’s Wavenet technology in speech quality. Baidu was also looking at developing a hybrid Deep Voice technology with tens of Wavenet (convolutional) layers.

The speech quality could increase up to 3.53 MOS with 80 Wavenet layers, but the company offered no mention of how much the performance would degrade with all the Wavenet layers slowing it down. Baidu said that it may further investigate the hybrid approach in the future.

A Future Powered By Synthetic Speech

The rise in popularity of digital assistants has happened in part because they have gotten much smarter due to all the latest improvements in machine learning, but also because they’ve begun to sound increasingly more like humans. The result has been easier and less awkward conversations with various AI assistants.

Baidu also believes that speech will be one of the main ways we’ll interact with computers in the future. The company said that it’s working hard to enable that future, through speech synthesis technologies such as Deep Voice, speaker identification technologies such as the recently announced Deep Speaker, and the end-to-end speech recognition system Deep Speech 2.

After a close second match. for the first 100 moves anyway, between Ke Jie and AlphaGo, the AI got the upper hand towards the end of the match and eventually defeated Ke Jie.

“Perfect Moves” From Ke Jie

In the first match between AlphaGo and Ke Jie, the AI scored a narrow win of only half a point against the world’s current best (human) Go player. However, AlphaGo seemed to only care about winning, and not necessarily winning by a large margin, which made it hard to deduce whether or not it was close to its game-playing limits.

Ke Jie said that in the second match, the AI played “opposite” of how he expected it to play to maximize its chance of winning, which means that AlphaGo changed its playing style from the first match.

Ke Jie was excited to play the game in the beginning and even thought that he may be able to win. The AI’s evaluation of Jie’s moves was also showing that the human Go player’s first 50 moves were “perfect.”

“For the first 100 moves it was the closest we’ve ever seen anyone play against the Master version of AlphaGo,” said DeepMind CEO Demis Hassabis in the post-game press conference.

However, around the middle of the game, AlphaGo started to gain an upper hand and Jie became nervous about his chances of winning. Eventually the AI forced the human player to resign.

AlphaGo Expert At Playing Against Itself

The updated version of AlphaGo was nicknamed “Master” by the DeepMind engineers late last year when they unleashed it online and allowed it to play 60 games against human players. AlphaGo proved to be unbeatable and went 60-0. Among those players was Ke Jie, who lost three times.

AlphaGo “Master” learns mainly by playing against itself, which means that it has become an expert at finding its own weaknesses and then either eliminating those weaknesses or learning how to properly counteract when its match rivals try to exploit them.

Hassabis mentioned that even though the AI may already know all the moves it needs to make against itself to win, that’s not necessarily true when playing against human players. However, as we mentioned in a previous post, even though it should theoretically be possible for humans to “surprise” AlphaGo with certain moves, chances are slim that AlphaGo can still be surprised in a major way. The AI was able to watch millions of matches between professional players and play against itself thousands of times during the training of its neural networks.

Final Match To Watch

Even though AlphaGo has already won two out of three scheduled matches against Ke Jie, Ke Jie asked DeepMind’s CEO if he could play “white” (second to move) in the final match, to experiment some more aggressive tactics against AlphaGo. If Jie’s tactics are successful, this could make the final match the most interesting one to watch yet. The final match will be streamed on YouTube on Saturday (10:30-17:30, UTC+8).

This Friday, there will be two other matches: One is called "Pair Go" (8:30-12:30, UTC+8), in which two iterations of AlphaGo will assist human players who will play against each other, and the other is "Team Tournament" (12:30-17:30, UTC+8), in which AlphaGo will face a group of five players, all playing together to defeat the AI.

Updated, 5/25/2017, 9:05am PT: The article was corrected to say that "white" is the second to move, rather than first, in the Go game.

A year after AlphaGo defeated Lee Sedol, an 18-time world champion at Go, the AI faced Ke Jie, who is currently considered the world’s best Go player. AlphaGo beat Ke Jie with only half a point difference--the smallest possible--but that may be due to the AI’s “safer” winning strategy.

AlphaGo AI

The AlphaGo AI is built on the core DeepMind technology that that has already been used to cut Google’s data center cooling costs and to discover better treatments for some diseases.

AlphaGo is a special version of DeepMind AI that only knows how to play Go. Google has “shown” it (as it learns through computer vision) millions of matches between professional (human) Go players, and the company also made AlphaGo play against itself, so it could learn from its own mistakes. Basically, AlphaGo learns similarly to how a human would learn, by watching others do a task well and then repeating that task over and over until its skill at the task improves.

This machine learning technique was not just an innovative way to teach AlphaGo how to play Go well, but it was quite necessary for AlphaGo’s successful learning. That’s because it wasn’t possible to program an artificial intelligence to play and win at Go in the same way chess-focused artificial intelligence was programmed before it.

There are more possible Go moves than atoms in the universe, according to Google, which would have made it impossible for the AI to “play ahead” all the moves until it found the winning ones. This is also why experts thought having an AI win at Go was going to take at least another decade, before AlphaGo was created.

By learning how humans play best, AlphaGo was able to not only replicate those winning moves in various Go-playing scenarios, but also create its own patterns for what a winning move would look like.

Ultimately, this allowed AlphaGo to beat even the best Go players in the world, such as Lee Sedol, and now Ke Jie, too. However, Ke Jie still has two more tries left before a final winner is declared.

AlphaGo Wins One Out Of Three

Although AlphaGo and Lee Sedol played five Go matches against each other, there will be only three matches between the AI and Ke Jie this time around.

In the first natch, Ke Jie lost to AlphaGo by the smallest margin possible: half a point. However, as is often the case in sports, the final score doesn't always clearly indicate how close a competition really was. According to Demis Hassabis, DeepMind’s founder, AlphaGo is not as interested in winning by large margins, as it’s interested in winning period. At the end of the day, winning is what matters in Go, and this seems to be what AlphaGo cares about, too.

Strategy Against AlphaGo

Unlike Lee Sedol, who didn’t really know what to expect from AlphaGo, Ke Jie was more prepared for this match. He’d seen the matches against Sedol, as well as other matches played by AlphaGo online under the nickname of “Master,” so he could understand a little better how the Google AI likes to play.

Ke Jie tried to use a strategy he’s seen AlphaGo use online before, but that didn’t work out for him in the end. Jie should’ve probably known that AlphaGo must have already played such moves against itself when training, which should also mean that it should know how to “defeat itself” in such scenarios.

A more successful strategy against AlphaGo may be one that AlphaGo hasn’t seen before. However, considering Google has shown it millions of matches from top players, coming up with such “unseen moves” may be difficult, especially for a human player who can’t watch millions of hours of video to train.

However, according to Hassabis, the AlphaGo AI also seems to have “liberated” Go players when thinking about Go strategies, by making them think that no move is impossible. This could lead to Go players trying out more innovative moves in the future, but it remains to be seen if Ke Jie will try that strategy in future matches against AlphaGo.

Although Google hasn’t mentioned anything about this yet, it’s likely that both AlphaGo’s neural networks as well as the hardware doing all the computations have received significant upgrades from last year. Google recently introduced the Cloud TPU, its second-generation “Tensor Processing Unit,” which should have not only have much faster inference performance, but now it comes with high training performance, too. As Google previously used the TPUs to power AlphaGo, it may have also used the next-gen versions to power AlphaGo in the match against Ke Jie.

Next AlphaGo Matches

The next match between Ke Jie and AlphaGo will happen on Thursday, and then the final one will be streamed on Saturday. At the “Future of Go Summit” in Wuzhen, China, where these matches take place, there will also be a match between five human players and one AlphaGo AI, as well as a match between two humans who are both assisted by AlphaGo AI instances.

Last year, Google’s AlphaGo AI, which was based on the core DeepMind AI technology, played a historic series of matches against Lee Sedol, an 18-time world champion at the Go board game. This year, after many requests from the Go community, Google plans to pit the AlphaGo AI against Ke Jie, the world’s current no. 1 player.

AI Wins At Professional Go For The First Time

Last year, AlphaGo beat Sedol 4-1 in a five-match game. Sedol held his own throughout most of the five matches, but ultimately he couldn’t defeat the AI more than once. This was a historic moment, similar to the one when IBM’s Deep Blue computer beat Garry Kasparov at chess about two decades ago.

Since then, computers and machine learning has evolved significantly, which is what made it possible for an artificial intelligence to defeat a human pro player at Go. However, chess had a manageable number of possible moves on the board. This eventually allowed computers to quickly test all the possible moves in a given scenario on the board and pick the “best” one.

Go, on the other hand, has many orders of magnitude more possible moves, which can’t be calculated by a traditional computer or program by trying them all out ahead of time. Therefore, for an AI to beat a human pro player at Go, it needs the same type of “intuition” as humans. AlphaGo will “consider” moving in multiple places on the board, and then pick a move that it believes has a higher chance of “winning,” but it can’t simulate the entire remaining match.

AlphaGo’s Upcoming Matches

Google’s DeepMind team announced that from May 23-27, it will collaborate with the China Go Association and the Chinese government to create another event in which AlphaGo will play against multiple top Go players, including Jie, the world’s #1 Go player at the moment, who has also beaten Sedol several times in the past.

The new matches will be set up as follows:

Pair Go - a game of Go between two human players who will be assisted by two different AlphaGo instances. The humans will alternate their moves with their AlphaGo assistants. The idea is to make the game more interesting, as well as to show that Go doesn’t have to disappear as a result of AIs beating humans, but it could evolve to include AI assistants, too.

Team Go - this should be another interesting match, which will test AlphaGo’s creativity against the creativity of a team of five top Go players. The idea here is to test AlphaGo’s weak points (if it has any) by coming at it with different styles of play.

Ke Jie vs AlphaGo - the final games will consist of three 1:1 matches between Ke Jie and AlphaGo. That’s when we’ll see if any single human has any chance of beating AlphaGo anymore.  

The chances should be quite low considering AlphaGo has had a whole year to improve by "watching" millions of recorded game plays from top Go players. The AI can also evolve by playing against a version of itself, as it did before the game with Sedol. This time, AlphaGo may also take full advantage of Google’s TPU chips, so it probably won’t be limited by hardware resources too much.

A Rapid Improvement In AI Technology

Google’s core DeepMind technology has continued to improve, not just for playing Go and other games, but also for solving real-world problems. The company revealed last year that it was using DeepMind machine learning to cut its cooling bill for a data center by 40%, for example, and the DeepMind team has also collaborated with hospitals and medical research teams to study various diseases and types of cancer.

Other machine learning teams inside Google have also made breakthroughs in object recognition in photographs, allowing Google Photos users to search for rather specific items within a photo, as well as in significantly improving the quality of Google Translate.

It’s still early days for machine learning, but we’re already seeing a strong focus on improving machine learning hardware as well as software from large industry players. This should lead to a quick evolution of machine learning over the next few years, making it increasingly more useful in solving more types of real-world problems.

OpenAI was created to democratize and decentralize artificial intelligence. The nonprofit came closer to that goal with the release of Universe, a platform that will allow anyone to create advanced AI agents as part of the group's ongoing research to make AI more human-like. Alphabet's DeepMind also released its own open source 3D virtual lab to help researchers train AI agents in real world-like environments.

OpenAI Universe

Universe allows AI agents to do all sorts of tasks a human would do with a keyboard and a computer while watching a screen. The platform comes with over 1,000 environments--including Flash games, browser tasks, and games like Grand Theft Auto V--to fill that purpose. OpenAI's ultimate goal is to create an AI agent that draws from the experience of the many other agents using the platform to learn new skills within all the available environments.

Games will be an area of significant focus for the Universe platform. AI researchers, or anyone willing to help, can let the AI agents train on games they play. Soon, the AI agents will be able to play those games as well as human-like bots that learn by visualizing what’s happening on the screen, rather than taking actions based on pre-programmed algorithms.

So far, Universe includes over 1,000 Flash games, Atari 2600 games, and environments where the AI agent interacts with websites’ user interfaces. OpenAI hopes to expand this list to include more games from partners such as EA, Valve, and others, as well as HTML5 games, Unity games, online educational games, and possibly other types of environments as well.

DeepMind Lab

Although the DeepMind AI has also been trained on a multitude of games, the company has also created its own virtual lab for building 3D environments in which AI agents could train. The environment is seen from the first person view of the AI agent, which has a “floating orb” for a body. It can look and move around in these 3D environments, and it can also do things like collect fruit, navigate mazes, traverse dangerous passages, play laser tag, and learn and remember procedurally generated environments.

DeepMind said its virtual lab emphasizes navigation, memory, 3D vision, motor control, planning, strategy, and time. Fully autonomous agents also have to learn on their own what tasks to perform in a given environment.

According to the company, the DeepMind Lab is highly customizable, and researchers can build their own levels for AI training. The levels can be modified with gameplay logic, item pickups, custom observations, level restarts, reward schemes, in-game messages, and more. All code, maps, and level scripts will be available on DeepMind’s Github page.

The company said that so far it has only scratched the surface on many of these tasks, and there is still much to be discovered through research in the areas of navigation, memory, and exploration. However, it hopes that other researchers using the DeepMind Lab can speed up the process of achieving a general-purpose artificial intelligence.

The DeepMind team believes that it’s fundamentally easier to train artificial general intelligence in 3D environments seen from a first person point of view. The team thinks that we, humans, wouldn’t have developed too much general-purpose intelligence if we were born in a Pac-Man environment. Following the same logic, it should be easier to create a general-purpose AI agent by training it in a 3D world.

Some of the top technology companies that have already invested heavily in machine learning technologies and services have joined together in a consortium to share research and create best practices on AI technologies. The Partnership on AI’s initial founders include Amazon, Facebook, IBM, Google/DeepMind, and Microsoft.

The consortium will accept other members in the future, including non-profits and activists, and they will get equal membership roles in the group along with the corporations. Other important AI players in the industry who seem to be missing from the group right now are Apple, Baidu, and OpenAI. However, they may eventually become members as well, once membership is open, unless they have some other objections to the goals of the Partnership on AI consortium.

Stated Goals

The group wants to support research and recommend best practices in areas such as ethics, transparency, interoperability, privacy, reliability, and robustness of technology. It also aims to increase awareness and understanding about AI for the public, presumably so the public keeps seeing the rise of AI as a positive evolution, rather than a negative or dangerous one.

Lastly, the group wants to create an open platform where researchers can communicate with each other about new issues or advances in technology.

Mission

The consortium will engage experts not just in machine learning, but also in psychology, philosophy, economics, finance, sociology, public policy, and law, to discuss emerging issues related to AI’s impact on society.

The group also wants to finance objective third-party studies on best practices for ethics, safety, fairness, inclusiveness, trust, and robustness for AI research and applications.

Another mission of the consortium will be to engage AI users and developers, as well as representatives from sectors such as healthcare, financial services, transportation, commerce, manufacturing, telecommunications, and media.

Oligopoly on AI?

The fact that some of the most invested companies in artificial intelligence research are partnering is probably an overall good thing. They will get to share research with each other, and advance AI that much faster, rather than operate in their own silos. They can also, potentially, act as checks and balances against each other before doing something dangerous with AI.

The biggest criticism this group is likely to face is the perception that what they’re actually trying to do is create an oligopoly (the market is shared by a limited number of competitors) on AI, not just a “partnership,” and ultimately try to kill or cripple emerging competition.

However, the group promises to open up the consortium to others and give them equal leadership roles. Unless they fail to live up to that promise, this is probably not something anyone needs to worry about too much.

What Privacy?

A second criticism would be related to privacy. If you were already worried about one of these companies having too much of your data, then them sharing all of their data with each other, ought to give you a panic attack. Google, Facebook, Amazon, and Microsoft’s services and products cover just about everything we do on our devices and online. The partnership already seems to have started with the wrong foot in terms of privacy, as the PartnershipOnAI.org site doesn't even use HTTPS encryption.

The companies haven’t mentioned any data sharing between each other yet, although that could be likely to happen in the future. AI is defined as much by smart algorithms as it is by data. The smarter the algorithms and the more data you have, the more advanced the AI becomes. Therefore, these companies sharing their users data with each other seems almost inevitable.

They may not do it automatically, but instead do what Facebook did with Whatsapp, and prompt users with a privacy policy update that says they can now share your data with the other companies in the consortium. The users’ choice is likely to be opt-out as well, as it was for WhatsApp users, because they know that opt-out choices means most users will agree by default without paying too much attention to the policy updates.

The companies promise to keep an eye on how they deal with privacy going forward as well. However, it's not clear what exactly that means, and whether they will even adopt privacy-friendly technologies such as "differential privacy" when mining users' data to improve their artificial intelligence technologies.

A Positive Perspective

This partnership could lead to many positive things, as well. Faster advancements in AI also means new cures for diseases, new materials and technologies that can bring interplanetary travel closer to reality, or simply significantly improved services.

If nothing else, OpenAI, the non-profit organization created by Elon Musk and a few others for the purpose of not leaving AI in the hands of a few powerful corporations or governments, seems to be quite excited about the Partnership on AI consortium, and hopes to be part of it soon, as well.

Early this year, after Google’s DeepMind artificial intelligence technology proved it can beat an 18-time world champion at a game as complex as Go, many wondered if that intelligence could be put to good use in the real world, and how soon. Since then, Google showed that DeepMind can be used to cut the cooling bill for one of its data centers by 40%. The company just announced another real-world breakthrough pertaining to text-to-speech (TTS) technology.

Synthetic Speech Breakthrough

According to Google, DeepMind’s technology, called WaveNet, succeeded in reducing the gap between Google’s best TTS technology and the human voice by 50%, in terms of how natural synthetic speech can sound.

Until now, Google would use concatenative TTS, where the company would first record speech fragments spoken by real humans and then combine them to form sentences. This approach is what gives text-to-speech its “robotic” tone, because the words are spoken with no context or emotion.

Google also uses the parametric approach, where all the information required to generate the data is stored in the parameters of the model, and the contents and characteristics of the speech can be controlled via the inputs to the model. However, so far the parametric technology has been successful only with non-syllabic languages such as Mandarin Chinese, but it makes syllabic languages such as English sound less natural than does the concatenative method.

How WaveNet Works

WaveNet is a fully convolutional neural network that can modify the raw waveform of the audio signal one sample at a time. That means that for one second of audio, WaveNet can modify 16,000 samples (16KHz audio), making synthesized speech sound much more natural. Sometimes, WaveNet even generates sounds such as mouth movements or breathing, which shows the flexibility of using raw waveforms.

At training time, the inputs are real waveforms recorded from human speakers. After the training, Google can sample the network to generate synthetic speech. The process for picking the samples one step at a time makes for a computationally expensive process, but Google said that it is essential to generate realistic-sounding audio.

To test how good the new text-to-speech engine was, Google did a blind test with human subjects who would give 500 ratings on 100 sentences. The results are shown below, and as you can see, WaveNet reduced the gap between Google’s best previous technology (either concatenative for English or parametric for Mandarin) and human speech by around 50%.

Google’s DeepMind team seemed surprised that directly generating each audio sample even worked at all with deep neural networks, and they were even more surprised that it could outperform the company’s previous cutting edge TTS technology.

The team will continue to improve WaveNet so that it can create synthesized speech that’s even more human-like. Presumably, the team will also want to reduce the computational costs so that Google can start using WaveNet commercially as soon as possible.

Google’s Go-playing version of DeepMind AI, AlphaGo, managed to score yet another win against Lee Sedol, the 18-time world champion, in the second match out of a total of five.

AlphaGo’s Near Perfect Game

Yesterday, Lee Sedol believed that AlphaGo didn’t do so well in the beginning, but the AI managed to squeeze a win against him towards the end of the game. In the second game, it was also Lee Sedol’s opportunity to take advantage of AlphaGo’s potential weaknesses after learning a little bit about its style. However, to his surprise, AlphaGo played a “near perfect game” this time.

“Yesterday, I was surprised, but today I am quite speechless. I would have to say, if you look at the way it was played, I admit that it was a very clear loss on my part. From the very beginning of the game, there was not a moment in time where I thought that I was leading the game,” Lee Sedol said in the post-game conference.“Yesterday, as I was playing the game, I felt that AlphaGo played some problematic positions, but today I really feel that AlphaGo had played a near perfect game. There was not a moment that I thought AlphaGo’s moves were unreasonable,” he added.

At the conference he was also asked if he found any weaknesses in AlphaGo’s game, but he said that he lost the game because he couldn’t find any.

Sedol also expects that the next games are only going to become more difficult, possibly because AlphaGo becomes ever so slightly better after each game it plays. Therefore, he added that he would need to focus even more for the next games. He would also need to try and get an edge in the early game, when AlphaGo could be at its weakest, even though AlphaGo’s weakest moves could still be too good to be easily exploited.

A Stronger Challenger?

Chinese Go Grandmaster Ke Jie, who some believe that in the past few years has been a better player than Lee Sedol, said that at this point in time, he would have a 60 percent chance of beating AlphaGo.

Google would likely not give everyone who says they can beat AlphaGo a chance to play five matches against the AI, but it could be interesting to see additional matches against other world-class players. However, this would have to happen soon, as even Ke Jie said that it may be a matter of months, or at most years, until even he couldn’t beat the AI anymore.

Ke Jie only became a 9-dan (the highest level in Go) player last year, and since then he defeated Lee Sedol with a score of 3-2 earlier this year.

The next match between AlphaGo and Lee Sedol will be played on March 11, 11pm ET (March 12, 1pm KST).

Lucian Armasu is a Contributing Writer for Tom's Hardware. You can follow him at @lucian_armasu. 

Follow us on Facebook, Google+, RSS, Twitter and YouTube.

Google’s DeepMind-based AlphaGo AI managed to beat Lee Sedol, the world’s premiere Go player and 18-time world championship, in what represents a historical moment in the progress of artificial intelligence. Granted, the victory came in just the first match of five; the remainder will be played out over the next few days, through until March 15.

Lee Sedol seemed to have the upper hand for most of the game, until the last 20 minutes when AlphaGo gained a bigger lead on him. This may have happened because it’s harder for the AI to figure out an overwhelmingly winning strategy in such a complex game as Go. It can’t calculate too many movements ahead of time due to the enormous processing power that would require.

However, towards the end of the game, fewer moves are possible, which could mean it became much easier for AlphaGo to determine the best move at any given moment. Therefore, any mistake from Lee Sedol, no matter how small, could then be turned in a bigger and bigger advantage for AlphaGo as the game approached its conclusion.

Still, AlphaGo managed to hold its own with the world’s best Go Player for more than three hours. Perhaps if Sedol had taken a larger lead early on, he may been able to fend off the AI bot.

Even if Sedol wins all of the other four matches, Google still achieved a historical breakthrough in artificial intelligence with last night's win. It would then likely be only a matter of months or a year at most before AlphaGo and its DeepMind core, as well as the hardware processing power behind it, would improve enough to become unbeatable by any human player.

Since Google announced that its DeepMind AI can play and finish games from the 1970’s only about a year ago, the AI has already gone to play games such as Go and the 1990’s Doom video game. The improvement rate has been quite astonishing. The DeepMind AI is likely not too far behind being able to play any modern game in the same way a human would - by learning everything about the game from scratch.

The next match between AlphaGo and Lee Sedol will happen today at the same time, 11pm ET. It will be interesting to see whether Lee Sedol will use a different, perhaps more aggressive, strategy early on against AlphaGo, now that he learned how the AI plays, in order to get a larger advantage early on.

Lucian Armasu is a Contributing Writer for Tom's Hardware. You can follow him at @lucian_armasu. 

Follow us on Facebook, Google+, RSS, Twitter and YouTube.

Last month, Google announced that its AlphaGo program, powered by the DeepMind AI technology that it acquired in 2014, managed to beat the European Go champion, Fan Hui, in five out of five games. The company announced that next month, it will pit AlphaGo against Lee Se-dol, who has been the highest-ranked Go player in the world for the past decade.

Se-Dol’s Match

Lee Se-dol has a 71.8 winning percentage and is supposedly a much better player than Fan Hui. However, it’s hard to tell whether he has a rather good chance of beating Google’s AlphaGo AI agent or whether it will be impossible for him to beat it, too. AlphaGo beat Fan Hui 5-0, but it could’ve been either slightly better than him or millions of times better, which would mean Lee Se-dol doesn’t stand a chance either.

For now, Se-dol seems confident that it can beat Google’s AI, but he knows it’s only a matter of time until that AI becomes more advanced or until Google runs it on more powerful machines.

“I have heard that Google DeepMind’s AI is surprisingly strong and getting stronger, but I am confident that I can win at least this time,” said Se-dol.

The match between AlphaGo and Lee Se-dol will consist of five games, each played on a different day, presumably because humans need rest after such intense matches, which can alter their performance on the next games. The games will be streamed live on YouTube on March 9, 10, 12, 13 and 15. If Se-dol wins, he would get a $1 million prize.

Deep Blue vs DeepMind

The previous announcement from Google that its AI managed to beat a Go grandmaster is highly reminiscent of when IBM’s Deep Blue beat Gary Kasparov, the world’s best chess player at the time (but only after Kasparov beat it first, a year earlier).

However, this time things are different for two reasons. One is that chess is orders and orders of magnitude easier to beat by an AI. There are on average 35 possible moves you can make on a chessboard, while there are 250 possible positions on a Go board, and after each one you get 250 more, and so on. That means there are more possible positions on a Go board than atoms in the universe, according to Demis Hassabis, DeepMind’s creator.

In other words, until we develop large quantum computers, at least, such computations are not possible in a short amount of time. This is why beating human Go players has always been believed as sort of a holy grail in AI research. It would be some kind of breaking point from which we could make AI that actually starts resembling the human mind, and it wouldn’t be just a bunch of pre-programmed instructions on how to do a very specific thing (like win at chess).

The second reason why this time is so different from the Deep Blue “AI” is because unlike Deep Blue, which beat Kasparov through “bruteforce” (by actually doing all the possible calculations ahead, and choosing the best one), Google’s DeepMind-powered AlphaGo agent didn’t use bruteforce calculations.

AlphaGo’s General Intelligence

The real breakthrough was that AlphaGo learned on its own how to play Go by “watching” 30 million Go moves played by human experts. AlphaGo is merely a version of DeepMind, which Google used to train the AI how to play Go, but in reality DeepMind is actually an “AGI,” or artificial general intelligence.

Google has talked before about how it has been training it to learn arcade games from the 70s and 80s, as well as the Doom game from the 90s. The DeepMind agent is not pre-programmed to play any of those games, including Go.

It actually learns everything from scratch, even the game rules or how to use “button actions” during a video game. It initially fails all the time until it eventually figures out what are the moves that “reward” it with moving forward into the game.

In AlphaGo’s case, it learned how to predict a human’s movement 57 percent of the time, which means that on average it could beat human experts, but it could still be beaten in some of those games. To make it an even better player, Google pitted it against itself, so it could constantly improve its gameplay. This made AlphaGo so good at playing Go that without any tree search (bruteforce computations) at all, it could beat the best Go AIs out there that rely on enormous search trees.

Google believes that one day it should be possible to use this Artificial General Intelligence to address real-world problems, from climate modelling to complex disease analysis.

Lucian Armasu is a Contributing Writer for Tom's Hardware. You can follow him at @lucian_armasu. 

Follow us on Facebook, Google+, RSS, Twitter and YouTube.