Top News

  • Will the US government ban Chinese models? A new report from Axios suggests that, while an outright ban is probably unlikely, the Commerce Department, NSA, and White House have been exploring multiple options for stemming the availability and spread of powerful open source Chinese AI models for US users and companies. Options under consideration include adding Chinese AI labs to sanctions lists, issuing direct security warnings, or imposing additional restrictions and regulations on US firms that host Chinese models. The President and his top advisors have reportedly been back and forth on the issue for months, but the strength of Moonshot’s new Kimi K3 release has apparently given the “tighter regulations” side a boost. Meanwhile, in a widely shared Washington Post editorial, VC legend Bill Gurley argues that open source models were inevitable once users and enterprises started seeking intelligence options they could own and control, at a lower cost than the frontier labs could provide. To that end, Gurley goes on to argue that “nearly everyone in the AI economy has a reason to prefer an open foundation…” except, that is, frontier labs Anthropic and OpenAI.

  • Z.ai’s all-Chinese chip data center: Meanwhile, in China, open source leaders Z.ai (the lab behind the popular GLM-5.2 model) completed work on a giant new data center running exclusively Chinese-made chips. The 1-gigawatt hub produces enough energy to power around 750,000 homes, and is stacked with multiple clusters each housing 10,000 chips. It’s not just a new data center but a proof of concept for Beijing, demonstrating that China can continue developing powerful AI models at lightning speed even without relying on export controlled GPUs from US firm Nvidia.

  • AI may have solved a nearly 90-year-old math problem: The Jacobian conjecture is an extremely difficult open problem in algebraic geometry, first posited by German mathematician Ott-Heinrich Keller in 1939. (I could explain it to you here in this newsletter, but I don’t feel like it.) Over the weekend, mathematician Levent Alpöge of Harvard University posted to X that he has cracked the puzzle, and determined that the Jacobian conjecture is false with help from Anthropic’s frontier AI model, Fable 5. His counter-example: ((1+xy)^3 z + y^2 (1+xy) (4+3xy), y + 3 x (1+xy)^2 z + 3 x y^2 (4+3xy), 2 x - 3 x^2 y - x^3 z): \C^3\to \C^3, has jacobian determinant -2, and sends (0, 0, -1/4), (1, -3/2, 13/2), and (-1, 3/2, 13/2) to (-1/4, 0, 0). See? Simple! This apparently now ranks as the toughest math problem ever solved by AI.

TWiST 500

Two-year-old UK-based CuspAI functions as a kind of “search engine for materials.” Users enter the chemical and/or physical properties they’re looking for, and Cusp’s platform — MIRA — recommends concepts for materials that could theoretically deliver on these requests. It’s essentially the exact reverse of how material sciences typically works, with researchers discovering new elements and then playing around with them to see what they can do.

Cusp then powers through the full discovery cycle. Generating candidate materials, simulating them to see how they might theoretically function in the real world, figuring out how to actually generate them, and even coordinating the lab tests to determine if things worked out properly. It does this by training its models to be “synthesis-aware.” That is, to know if it’s designing materials that could actually be created, rather than just daydreaming about fictional substances that can’t ever be fashioned in our reality.

This week, CuspAI launched the “AI Materials Foundry,” a coalition of 48 tech companies, industrial firms, and research facilities that will work together to develop these new materials, along with fine-tuning processes to generate them at a faster pace and lower cost. Big-name participants include Nvidia, Meta, Hyundai Motor Group, Samsung, Applied Materials, and Tokyo Electron. Nvidia will help out with the compute, while Meta’s research lab has developed a model specifically built to simulate atoms.

Cusp has also confirmed a $450 million Series B round backed by Kleiner Perkins and Amazon magnate Jeff Bezos’ personal investment fund, boosting the startup’s value up to $2.6 billion.

Obviously, new substances that could replace the rare earth metals at the heart of modern-day AI chipmaking — such as iridium and ruthenium — could significantly cut the cost of churning out fresh chips (with far-reaching geopolitical and supply chain ramifications as well). But CuspAI is chasing a number of other sectors at once, including chemicals and, obviously, the automotive industry. – Lon

A message from Rippling

Don’t settle for AI that’s all talk. Head to Rippling.ai/twist to get the only AI built to give you full visibility across your startup and take complex actions across your entire business.

This Week in Startups

E2313: Science fiction is full of surgery-bots, and Andromeda Surgical is hoping to bring this dream into reality (preferably in a less troubling way than that one scene in “Prometheus.”) CEO Nick Damiano stops by the show to explain why the company is focusing on software, and relying on off-the-shelf robot arms from Germany to actually complete the procedures. Plus Mogul co-founder/CEO Alex Blackwood shows Jason how his platform allows anyone to buy fractional shares of rental homes for as low as $250, benefitting from monthly dividends and appreciation without having to become a landlord.

E2312: Hustle Fund’s Eric Bahn and Chapter One’s Jeff Morris Jr. join our venture capital roundtable directly following the news that Stripe wants to buy PayPal. The possible transaction highlights how little interest today’s leading startups have in going public any time soon. With host Alex Wilhelm, Bahn and Morris dug into when startups should burn the boats à la Fin, physical AI as the next frontier, the power of Silicon Valley groupthink, and how startups are approaching the first days of widespread, company-specific AI evals!

E2311: Jason’s been talking a lot about his personal vibecoding projects lately, and on this TWiST ep, we finally get a look at what he’s making. PodMeme is a research AI tool that seeks out segments from across different podcasts about any common theme or idea, and then strings these outtakes together into one seamless feed. On TWiST, we get advice on how to promote a product even before it exists from Tweet Hunter and Taplio creator Tibo “Maker” Louis-Lucas, plus the latest news and streaming recommendations.

TWiST Partner Offers

  • Vanta: Get $1000 off your SOC 2 at https://www.vanta.com/twist

  • Superhuman: Get the AI that works where you work. Find out more at superhuman.com.

  • YSecurity: The on-demand security team for startups. Need enterprise-grade security without hiring a $400k CISO? YSecurity gives you 40+ expert engineers, matched to exactly what you need, by the hour, with your first six hours completely free. Go to YSecurity.io/TWIST.

Subscribe to Us On Your Favorite Podcast Platforms

SF Live-Work Space Now Available

Need a flexible living and working environment in San Francisco? This thoughtfully designed loft-style residence at 787 Bryant St., the heart of the vibrant SOMA district and the city’s creative hub, is now available for rent or purchase. Check the listing for more details.

The TWiST500 newsletter is the new, updated, and improved TWiST Ticker.