Andrew Barto and Richard Sutton, who received computing’s highest honor this week for their foundational work on reinforcement learning, didn’t waste any time using their new platform to sound alarms about unsafe AI development practices in the industry.

The pair were announced as recipients of the 2024 ACM A.M. Turing Award on Wednesday, often dubbed the “Nobel Prize of Computing,” and is accompanied by a $1 million prize funded by Google.

Rather than simply celebrating their achievement, they immediately criticized what they see as dangerously rushed deployment of AI technologies.

“Releasing software to millions of people without safeguards is not good engineering practice,” Barto told The Financial Times. “Engineering practice has evolved to try to mitigate the negative consequences of technology, and I don’t see that being practiced by the companies that are developing.”

Their assessment likened current AI development practices like “building a bridge and testing it by having people use it” without proper safety checks in place, as AI companies seek to prioritize business incentives over responsible innovation.

The duo’s journey began in the late 1970s when Sutton was Barto’s student at the University of Massachusetts. Throughout the 1980s, they developed reinforcement learning—a technique where AI systems learn through trial and error by receiving rewards or penalties—when few believed in the approach.

Their work culminated in their seminal 1998 textbook “Reinforcement Learning: An Introduction,” which has been cited almost 80 thousand times and became the bible for a generation of AI researchers.

“Barto and Sutton’s work demonstrates the immense potential of applying a multidisciplinary approach to longstanding challenges in our field,” ACM President Yannis Ioannidis said in an announcement. “Reinforcement learning continues to grow and offers great potential for further advances in computing and many other disciplines.”

The $1 million Turing Award comes as reinforcement learning continues to drive innovation across robotics, chip design, and large language models, with reinforcement learning from human feedback (RLHF) becoming a critical training method for systems like ChatGPT.

Industry-wide safety concerns

Still, the pair’s warnings echo growing concerns from other big names in the field of computer science.

Yoshua Bengio, himself a Turing Award recipient, publicly supported their stance on Bluesky.

“Congratulations to Rich Sutton and Andrew Barto on receiving the Turing Award in recognition of their significant contributions to ML,” he said. “I also stand with them: Releasing models to the public without the right technical and societal safeguards is irresponsible.”

Their position aligns with criticisms from Geoffrey Hinton, another Turing Award winner—known as the godfather of AI—as well as a 2023 statement from top AI researchers and executives—including OpenAI CEO Sam Altman—that called for mitigating extinction risks from AI as a global priority.

Former OpenAI researchers have raised similar concerns.

Jan Leike, who recently resigned as head of OpenAI’s alignment initiatives and joined rival AI company Anthropic, pointed to an inadequate safety focus, writing that “building smarter-than-human machines is an inherently dangerous endeavor.”

“Over the past years, safety culture and processes have taken a backseat to shiny products,” Leike said.

Leopold Aschenbrenner, another former OpenAI safety researcher, called security practices at the company “egregiously insufficient.” At the same time, Paul Christiano, who also previously led OpenAI’s language model alignment team, suggested there might be a “10-20% chance of AI takeover, [with] many [or] most humans dead.”

Despite their warnings, Barto and Sutton maintain a cautiously optimistic outlook on AI’s potential.

In an interview with Axios, both suggested that current fears about AI might be overblown, though they acknowledge significant social upheaval is possible.

“I think there’s a lot of opportunity for these systems to improve many aspects of our life and society, assuming sufficient caution is taken,” Barto told Axios.

Sutton sees artificial general intelligence as a watershed moment, framing it as an opportunity to introduce new “minds” into the world without them developing through biological evolution—essentially opening the gates for humanity to interact with sentient machines in the future.

Edited by Sebastian Sinclair

Figure AI, a U.S.-based startup focused on building AI-powered humanoid robots, severed its ties with OpenAI last week, with CEO Brett Adcock claiming a “major breakthrough” in robot intelligence that made the partnership unnecessary.

The split came just months after the two companies announced their collaboration alongside a $675 million funding round that valued Figure at $2.6 billion to kick-start its Figure 02 robot.

“Today, I made the decision to leave our Collaboration Agreement with OpenAI,” Adcock tweeted. “Figure made a major breakthrough on fully end-to-end robot AI, built entirely in-house”. The move marked a stark reversal for Figure, which previously planned to use OpenAI’s models for its Figure 02 humanoid’s natural language capabilities.

In a separate post, Adcock explained that, over time, maintaining a partnership with OpenAI to use its LLMs started to make less sense for his company.

“LLMs are getting smarter yet more commoditized. For us, LLMs have quickly become the smallest piece of the puzzle,” Adcock wrote. “Figure’s AI models are built entirely in-house, making external AI partnerships not just cumbersome but ultimately irrelevant to our success.”

The decision came amid broader changes in the AI landscape. OpenAI itself had been rebuilding its robotics team, filing a trademark application mentioning “humanoid robots“—alongside a wide array of other technologies like virtual reality, augmented reality, agents, and wearables. It began hiring for its first robotics positions last month.

Some AI enthusiasts were quick to note that the move could just be another consequence of the DeepSeek effect—which has already forced the most powerful AI companies in the world to lower the prices of all their SOTA models to remain competitive against open-source alternatives.

OpenAI provides one of the most expensive LLMs in the market—with DeepSeek R1 providing better results than OpenAI o1 while being available free, open source, uncensored, and highly customizable. Figure could simply be betting on an in-house foundational model to power its lineup without depending on OpenAI’s offerings.

Figure has already secured a deal with BMW Manufacturing to integrate humanoid robots into automotive production, and recently struck a partnership with an unnamed major U.S. client that would be its second big commercial client.

“It gives us potential to ship at high volumes—which will drive cost reduction and AI data collection,” Adcock posted on LinkedIn a week ago. “Between both customers, we believe there is a path to 100,000 robots over the next four years.”

Figure developed a data engine that powered its “embodied artificial intelligence” systems, enabling its robots to learn and adapt in real time through cloud and edge computing solutions. The company’s technology allowed its robots to respond to language prompts and perform tasks that incorporated language, vision, and action.

“We’re working on training the robot on how to do use case work at high speeds and high performance” Adcock said. “Learning the use case with AI is the only path”

OpenAI still maintains investments in other robotics ventures, including Norwegian startup 1X.

Adcock promised to reveal the fruits of Figure’s “breakthrough” within 30 days, and he wasn’t subtle with his words. He promised the announcement would be “something no one has ever seen on a humanoid.”

Guess he learned from the best.

Edited by Andrew Hayward

Presearch—a decentralized and privacy-oriented search engine—has just launched PreGPT 2.0, marking the company’s latest attempt to challenge Big Tech’s dominance in the AI chatbot space.

The new release brought improved language models and a broader selection of open-source AI options, all running on a network of distributed computers rather than centralized data centers.

“Why am I so excited? Because PreGPT 2.0 is so powerful and unrestrained, that it has the potential to fundamentally disrupt the echo chamber effect that has long been manipulating conventional wisdom, amplifying the herd instinct into blind conformity,” Brenden Tacon, innovation and operations lead for Presearch, told Decrypt.

The upgraded chatbot comes with two subscription tiers: a $2 monthly basic plan running Mistral AI’s 7B model and a $5 pro version powered by Venice.ai’s more sophisticated LLMs. Both options promise to keep user data private and conversations unmonitored, with chats permanently erased upon deletion.

PreGPT 2.0’s model lineup features six of the most well-known names in the open-source AI space: Meta’s Llama-3.1-405b (a gigantic model), Llama-3.2-3b (a very small model built for efficiency) and Llama-3.3-70b (its latest LLM), Alibaba’s Qwen 32b.

It even leverages the old Dolphin 2.9 model, previously known in AI circles for being fully uncensored and powerful—and good at roleplay. The company also seems to have fine-tuned the Mistral 7B model to offer a custom version.

“This model gracefully handles a context of 8,000 Tokens, which equates to about 5,000 words, and you will be throttled to 1000 messages per month,” according to the company’s website.

This means the model will have a memory of 5,000 words and won’t be able to properly handle conversations that exceed such limit—or won’t process prompts that are that long.

What is Presearch?

Presearch, which launched in beta back in 2017 and went live in 2018, is basically a project that wants to reimagine search engine architecture with decentralized technology.

The platform processed over 12 million monthly searches through a web of independent nodes. Each node operator staked PRE tokens and lent computing power to the network, creating a self-sustaining ecosystem that scaled naturally with demand.

The idea is that a decentralized network makes the profiling of users—Google’s business model—harder and could help generate a business model that is more transparent and organic.

The platform’s advertising model is also different from what you see in Google or Bing, for example.

Instead of bidding wars for keywords, advertisers staked PRE tokens to gain visibility. The more tokens they stake, the better their placement will be—a system that reduces token circulation while creating predictable revenue.

A portion of these tokens get burned periodically, gradually decreasing the total supply from its current 590 million PRE in circulation.

PreGPT 2.0 leveraged this distributed infrastructure by collaborating with Venice.ai, a privacy-conscious AI service provider, and Salad.com, a community that shares decentralized GPU power.

The professional tier operates on Venice.ai’s high-performance network, while the basic plan is supported by Salad.com’s distributed GPU network.

Both pathways encrypt user interactions and refrain from storing chat logs, upholding Presearch’s commitment to privacy.

PRE’s tokenomics keeps the system running smoothly. Users earn up to 8 tokens daily for search queries, while node operators receive rewards based on their stake size and search volume.

This, at least in theory, seems like a win-win situation in which both users and advertisers are properly rewarded while helping the ecosystem grow.

PreGPT 2.0 is a separate AI feature added to Presearch’s toolkit; the company remains focused on its core mission of decentralized, private search.

The chatbot integration is intended to complement the search experience without overshadowing it.

The goal is to make the entire platform ideal for privacy-conscious users who want a replacement for traditional web searches and are curious about using AI tools in their everyday lives.

Hands-On with PreGPT 2.0: Promise and Limitations

Testing PreGPT 2.0 revealed a capable chatbot that prioritizes function over flash. The interface felt cleaner than competitors like Venice.ai or Hugging Chat, though it lacked image generation capabilities that have become standard elsewhere.

The integration of a system prompt feature lets users fine-tune the AI’s behavior through custom instructions, which is helpful for getting more precise responses—a sound system prompt can dramatically increase a model’s performance.

The overall experience will feel familiar to those used to tinkering with different chatbots.

This wasn’t a revolutionary leap in AI capability but rather a privacy-focused implementation of existing open-source models that are often less powerful than mainstream alternatives like GPT-4o or Deepseek.

The platform only manages plain text. It can craft a bedtime story or summarize trends, but it lacks support for Excel documents and cannot properly format CSV files, PDFs, or third-party docs.

Instead, users must actually copy the contents of a sheet and paste it, which is far from ideal. Those who confuse decentralization with slow speeds have nothing to worry about.

The replies were fast, and the chatbot never hung. But the models offered the quality you’d expect from open-source LLMs that are not really topping the charts in the LLM Arena—LLama 3.1 405b is currently in the 27th position and is the most powerful model in Presearch’s roster.

It’s not bad, but it’s also not impressive by today’s standards.

There are currently some open-source implementations that are a lot more powerful at arguably similar sizes.

For example, Llama-3.1-Nemotron-70B-Instruct could easily substitute the newer (but not better) Llama-3.3-70b, and Deepseek R1 is leaps ahead of Meta’s Llama 3.1 405b, being the best open-source model to date.

Overall, the experience was pleasant; the models performed as expected, and the interface was easier to use than Venice AI, its main competitor.

If you are looking for a privacy solution or want to try every AI tool available today, this feature is definitely worth a look. Just take into consideration that the search engine won’t replace Google, and the AI chatbot won’t replace ChatGPT—at least not yet.

Edited by Josh Quittner and Sebastian Sinclair