Nvidia’s CEO stars at the “AI Superbowl”

Leading AI chipmaker Nvidia had its annual GPU Technology Conference this week, an event sometimes called the “Superbowl of AI”, and CEO Jensen Huang’s 2-hour keynote was a showcase of the company’s latest technology. Cruising the stage in his signature leather bomber jacket, Huang wowed the crowd of 25,000 attendees (and 300,000 online viewers) with a cornucopia of leading-edge tech goodies.

• New superchips: Huang outlined the company’s roadmap for new chips over the next 2 years: one codenamed Blackwell Ultra which will run at 20 trillion computer operations per second, or 20 petaflops, debuting later this year, then the Vera Rubin (50 petaflops) in 2026, and the Rubin Ultra (100 petaflops) in 2027.

• Robotics: Nvidia is betting big on humanoid robots, announcing Groot N1, an open-source AI model for controlling them.

• Personal Supercomputers: The company released a personal supercomputer known as Project Digits earlier this year at the Consumer Electronics Show, which runs at 1 petaflop for the startlingly low price of $3,000. Now Nvidia has rebranded the line, with the existing machine named DGX Spark, and a more powerful 20 petaflop version named DGX Station, which is likely to have a price tag in excess of $10,000. (High end PCs currently run at 0.04 petaflops, 25 times slower than the smaller Nvidia machine.)

• Self-driving cars: Nvidia has allied with General Motors. GM appears to have had the bejeesus scared out of it by China’s fast-growing electric vehicle manufacturers, who are now also making humanoid robots to assemble their cars. GM will use Nvidia chips and expertise to remake their design process, their factories, and their in-car intelligence, including self-driving.

• Disney partnership: Nvidia and Google’s DeepMind are teaming up with Disney to create Newton, a next-generation physics engine designed to make robot interactions more lifelike than Disney’s old-school “Animatronics.” The new robots may hit Disney theme parks as soon as next year.

  

  CEO Jensen Huang says Nvidia chips will power more lifelike robots at Disney’s parks.

Mistral leaps ahead in small models, but the race is still on

For much of the last 2 years, bigger AI models were generally smarter than smaller ones. Now, after DeepSeek’s hyper-efficient R1 model showed in January that less can be more, the race is on to miniaturize AI models while preserving performance. This week, plucky French AI startup Mistral released a 24 billion-parameter model (the big boys can have more than a trillion) which it claims is outperforming all other models in its size class (see chart below.) Their moment of glory was brief. Yesterday, South Korea’s electronics giant LG released a high-performing competitive model of 32 billion parameters, with a tiny 2.4 billion parameter version which still performs surprisingly well.

Mistral leaps ahead in small models, but the race to miniaturize is still on

LG releases Exaone 32B model, along with tiny 2.4 billion parameter version

OpenAI releases tools for building AI Agents

Everyone in AI-land agrees that the Next Big Thing in artificial intelligence will be Agents, AI systems that can actually do stuff in the real world, not just talk about it. Last week, OpenAI released a suite of tools to help developers and large companies build their own Agents. So far, Agents have been more hype than reality, other than OpenAI’s own Deep Research product, which continues to get highly positive reviews from the deep-pocketed subscribers willing to pony up the $200 per month fee. The potential utility of Agents is obvious, so developers will likely keep plugging away until at least one of them develops the Killer App that unlocks the Agentic future.

OpenAI releases tools for building AI Agents

Current Y Combinator startups are the fastest-growing ever, due to AI

Y Combinator CEO Gerry Tan says that this year’s crop of startups is growing faster and more profitably than any previous group, thanks to AI. Y Combinator is a legendary tech startup accelerator, providing seed funding, mentoring, and lots of valuable connections to startup founders who are accepted into the program. The organization has helped startups become household names, such as Airbnb and Dropbox. Tan notes that over the last nine months, this year’s group of companies has been growing 10% a week (that’s over 100x in a year) and is much more profitable than prior groups. He attributes this to artificial intelligence. With AI, founders can have a smaller team, because AI can handle a lot of the repetitive work. In addition, AI is becoming so good at coding that companies no longer need so many software engineers. Tan states that one-quarter of this year’s crop of startups has code that is 95% produced by AI.

Current Y Combinator startups are the fastest-growing ever, due to AI

Sakana’s AI Scientist produces a peer-reviewed paper

Japanese AI company Sakana announced last week that their AI Scientist project has produced the first AI-generated scientific paper to successfully clear peer review at a prestigious machine learning conference. The AI Scientist system generated the hypothesis, coded the experiments, evaluated the results, and produced the paper including all text, graphics, and formatting, entirely without human intervention. With the knowledge and consent of the conference organizers, the paper was submitted to a double-blind peer review process as part of the conference workshop. Reviewers were told that they may be reviewing AI-generated papers as well as human-generated ones, but not told which was which. The paper was scored by reviewers as worthy of acceptance, but was then withdrawn prior to publication, since the scientific community has not yet reached consensus on the ethics and appropriateness of publication of AI-generated research. Using AI to advance scientific research is a major goal of the field, and AI-assisted research into AI is a natural, because all the experiments can be run on a computer - no messy test tubes or Petri dishes.

Sakana’s AI Scientist produces a peer-reviewed paper

OpenAI exec says AI will overtake humans at coding this year

It’s no secret that AI models are getting extremely good at producing computer code. Now Kevin Weil, OpenAI’s Chief Product officer says “This the year that AI gets better than humans at programming, forever.” OpenAI’s CEO Sam Altman teased this view last month at a tech event, proclaiming that the company had an internal model that was the 50th best programmer in the world, based on competitive coding challenges in the international Codeforces competitions. Altman went on to project that by the end of 2025 they would have a model that could best all humans. Now the company’s Chief Product Officer is forcefully proclaiming it in no uncertain terms.

OpenAI exec says AI will overtake humans at coding this year

Sudowrite releases fiction-writing AI

Sudowrite, an AI startup that develops AI-enabled writing tools for authors, has now released Muse, an AI model specifically designed for assisting with long form fiction. It can help the author brainstorm story ideas, critique drafts, and if given samples of the author’s writing, even generate paragraphs or more in the author’s style.

Sudowrite releases fiction-writing AI

AI has now passed the “meme Turing test” - it’s funnier than humans

Since John Henry lost a steel-driving competition to a steam drill, machines have been outpacing humans in more and fields of endeavor. Now with AI, the areas of computer dominance are expanding rapidly. Computer programming may be next (see story above.) But we could all agree that however smart AI was, it wasn’t funny. Until now. A study of memes created either by humans, by human-AI collaboration, or by AI alone, found that humans rated the AI-only memes highest on average. However, all the top-rated memes were generated by humans alone. So for now, AI is funnier than the average human, but not as funny as the funniest humans.

AI passes the “Meme Turing Test”

arXiv paper

Figure AI builds “BotQ” humanoid robot factory

Humanoid robotics startup Figure AI, flush from a $675 million funding round that values the 3-year-old company at $2.6 billion, has moved to a larger location and built out a robot factory they call “BotQ” that can produce 12,000 robots a year to start, with the aim of producing a fleet of 100,000 humanoid robots within 4 years.

Figure AI builds “BotQ” robot factory

Case Western team researches AI robots for assisting the elderly

An interdisciplinary team of physicians, scientists, and management experts at Case Western Reserve University are researching how to use AI-enabled robots to assist the elderly, particularly those that are suffering from Alzheimer’s disease. Current work is focusing on experiments with a 3-foot tall, wheeled robot they call Ruyi, which has advanced sensors and an AI “brain” that enables it to both monitor and interact with the elderly subjects. Ruyi is being tested at a senior living facility in Cleveland. The goal is to provide round-the-clock assistance to elderly residents with cognitive declines. Much of the study focuses not on just the technical performance of the robot, but its acceptability to the elderly individual, to their family, and to the senior living facility’s staff.

Case Western team researches AI robots for assisting the elderly

Harvard and MIT develop TxAgent for personalized healthcare advice

Researchers from Harvard and MIT have developed an AI system for personalized, up to date treatment recommendations. Dubbed “TxAgent”, the system can utilize multi-step reasoning and real-time biomedical information retrieval with a toolbox of 211 tools from trusted sources to analyze drug interactions, contraindications, and patient-specific treatment strategies. The goal of the system is to improve therapeutic decision-making in real time.

UVA AI finds new uses for existing drugs

Getting a new medication approved by the FDA and available to patients can take years, and millions if not billions of dollars. But drugs that are already approved by the FDA for one purpose can legally be used by physicians for other diseases, in what is called “off-label” use. Researchers at UVA School of Medicine in Charlottesville have developed an AI system, known as LogiRx, that searches through databases of FDA-approved medications to find drugs that may be useful for conditions other than the ones they were approved for. These “found” drugs can be used immediately for the new condition, if determined to be effective. Most recently, LogiRx identified the antidepressant Lexapro as potentially helpful for the prevention of heart failure. Research into the effectiveness of Lexapro for this purpose is being planned.

UVA AI finds new uses for existing drugs