I believe the value of "play" cannot be overstated. When we play, we overcome obstacles, unlock new creativity, develop strategies, form relationships and teams, and change the way our brain works.
In 1980, @Biogen, then a young biotech company in Cambridge, made an unusual request. It asked the city to reinstate regulations on recombinant DNA research that had been repealed two years earlier.
At a time when everyone worried oversight would stifle a promising new field, why would a company ask for more of it?
The answer was simple. Biogen wanted to build a laboratory, which meant convincing lenders and investors to commit capital for years. Without clear rules, it risked sinking millions into a project that could lose permission to operate.
Clear regulation wasn’t an obstacle. It was what gave the company the certainty it needed to build.
In the end, Cambridge revised its ordinance and Biogen built its laboratory, becoming one of the early companies to establish a presence in Kendall Square, now one of the world’s leading biotechnology hubs.
The usual assumption in technological advancement is a simple tradeoff: more rules, less innovation. But, as the history of Biogen and Kendall Square shows us, the reality isn’t so straightforward.
I wrote about how regulation can be an engine for innovation in the age of AI in “The Overlooked Lesson of Kendall Square,” my latest essay for the @Principal_VC newsletter. Read it here: https://t.co/NKNtxY1veP
Through my wonderful time at NVIDIA and Stanford, one question has kept me up at night:
Should the future of AI be monolithic or plural?
To answer this question, we’ve raised $110M to get started, backed by @MenloVentures, @MadronaVentures, @NVIDIA, @NEA, @KindredVentures,@Wing_VC, @aixventureshq, @Principal_VC, @Cisco, @Snowflake, among others.
Today, Carlos (@guestrin) and I are stoked to launch Noeri, a neolab to push the frontier of plural intelligence.
Noeri means “the plural of intelligence” or many intelligences.
The future of AI we root for is plural: a healthy ecosystem of many intelligences, in all sizes, strengths, and quirks.
Discovery, in particular, has always been plural: many curious minds, each devoted to their own question, often on roads less traveled.
Our path there is self-improving AI that:
-> learns continually on the job, closing the gap between what AI can do off the shelf and what the dynamic world demands
-> learns to use other AIs as tools and make new tools to overcome its own limitations
-> picks the right-sized model for your problem, at a fraction of the frontier cost
Then the power of intelligence, from the craftsmanship of making AI to ownership, can be in the hands of all.
@noeri_ai mission reflects who we are as the founding team. My career has been anything but linear, literally continual learning on the job. Carlos has devoted his career to building tools that empower people through AI.
We are all a bit contrarian, a bit adventurous, and above all, big believers in pluralism.
Join us to build this future together: https://t.co/oAAoe0lzEH
Should someone who spends more money have an advantage over someone who puts in more time? Does skill matter more than popularity? What’s the line between cunning and cheating?
These are questions the executive producer of a video game has to answer as they build a world with its own values, rewards, and incentives. In doing so, they’re writing a constitution: rules that teach players what leads to success and what counts as fair.
As AI spreads, we’re grappling with a similar question with far higher stakes: when a system shapes the rules people live by, who decides what’s fair?
Give an AI system a world to develop and goals to reach, and it may shape that world in ways no one anticipated, with unexpected consequences. The system may find the tactics, but someone still chose the goals.
In gaming, this is generally harmless: a player who thinks a game is unfair can usually leave. But if a system helps decide whether you get a job or a loan, walking away is a different proposition.
I explored this and more in my essay “When the World You Build Starts Thinking for Itself,” published in the fall issue of Daedalus from the American Academy of Arts & Sciences (@americanacad).
In it, I take a stab at the question I keep coming back to: who is the executive producer in the world of AI, and what do they owe the people who have to live by the rules they set?
Read the full essay here: https://t.co/hPEJKl2Zfv
Congratulations to our portfolio company Resin, named to the first-ever Emergent list from @NewcomerMedia and @felicis, which highlights the startups founders themselves are most excited about.
It's a well-earned recognition for what Pedro Tsividis (@ptsividis) and the team are building: generative tools designed around the artist.
At @Principal_VC, the creative AI we’re most interested in isn’t about replacing musicians and makers. It’s the kind that makes artists more capable without taking the work out of their hands. That’s the future we’re backing, and it’s exactly what Resin is building toward.
Excited for everything still to come, and very proud to back the Resin team.
Thrilled that @c_valenzuelab of Runway nominated Resin for the first-ever Emergent list from @NewcomerMedia and @felicis. At Resin, we’re building generative tools for music producers, designed around artist authorship and control. Much more to share soon. https://t.co/NB3EN5IVTa
Most AI today lives in the cloud. The next wave is moving somewhere more useful: onto the device in your hand, the factory floor, or the car you’re driving.
@Principal_VC portfolio company @liquidai is making that shift real, in a new partnership with @MercedesBenz covered today by @WSJ.
Mercedes is building Liquid’s models directly into its cars, so drivers can talk to the vehicle with the AI running on the device itself, not the cloud. That means faster responses, more privacy, and intelligence that works even without a connection.
This is the future I’m watching: efficient, purpose-built intelligence running close to where it’s needed, in the things we already use.
Congratulations to the Liquid AI team. Proud to back them, and excited for where on-device AI goes next.
You can read the full piece here: https://t.co/Y2U9NIUR4q
This week’s midterm results have rattled some nerves about the future of AI infrastructure spending, but beyond politics, there's an interesting signal here. The buildout is finally being asked to justify its economics.
The worry itself isn’t new. The market has been uneasy for a while about how much is being spent on AI infrastructure, and how much of that spending is fueled by debt. The elections didn’t create that concern, only resurfaced it.
This isn’t a referendum on whether AI is here to stay. But it does mark the moment capital discipline enters the picture: the market is increasingly asking not how big the buildout can get, but whether the spending produces durable returns.
That pressure tends to reward the companies that do more with less, rather than the ones racing to spend the most. Constraint is often where real advantage gets built, and a slower, more scrutinized buildout favors efficiency over sheer scale.
The market is starting to favor the companies that spend well over the ones that simply spend big, and that’s the shift worth watching.
Last month, @OpenAI published hundreds of math papers claiming an AI model had cracked over 300 hard problems.
The initial reaction to the news was mixed, but to me, the development raises a more fundamental question than how AI should be used in mathematics: why do we pursue these proofs in the first place?
The field’s most famous unsolved problems have resisted researchers for decades, and new progress is valuable. But solving them isn't about collecting trophies. Often the methods invented in pursuit of a proof matter as much as the conclusion.
Calculus did more than settle questions about motion; it gave scientists a language for change. Probability did more than answer puzzles about chance; it created a way to reason about uncertainty. Great mathematics doesn’t simply resolve problems. It changes which problems we are even capable of asking.
That is why counting AI-generated proofs is the wrong measure of progress. Proofs can establish that something is true without illuminating why. They only become useful knowledge once people verify them, connect them to what’s known, and build upon them.
The opportunity here isn’t to feed machines a longer list of famous conjectures. It’s to put mathematicians and scientists in a position to ask bigger questions.
I explored the evolving role of AI in mathematics in “Beyond Proof,” my latest piece for the @Principal_VC newsletter. Read it here: https://t.co/PMij8ytP4d
This week @Principal_VC celebrated a milestone: our move into a new Palo Alto office, with an Office Opening Reception that brought together leaders in AI and friends of the firm.
The evening centered on conversations with founders we’re proud to back, moderated by @WIRED Senior Writer Paresh Dave (@peard33).
We were lucky to hear from Daniela Rus, Director of MIT’s Computer Science and Artificial Intelligence Laboratory (@MIT_CSAIL) and co-founder of our portfolio companies Venti Technologies and @LiquidAI, and from @TimKentleyKlay, co-founder of @zoox and founder of our portfolio company HYPRLABS (@hypr).
Both spoke to the same shift: the developments in physical AI, intelligence that doesn’t just generate answers on a screen but operates in the unpredictable conditions of the real world with clarity and precision.
It’s one of the changes we’re watching closely as we look to what’s next in the AI trade, and these discussions were a fitting way to mark the start of a new chapter for PVP.
Thank you to Paresh, Daniela, and Tim, and to everyone who joined us. Glad to be in our new home, and looking forward to growing and building here.
The next stage of AI isn’t a better chatbot. It’s intelligence that can operate reliably in the physical world, where conditions are messy and mistakes are costly.
We’re starting to see what that looks like. @Principal_VC portfolio company Venti Technologies recently announced its first fleet of driverless trucks in an American rail yard, moving shipping containers in live operations.
This isn’t a leap out of nowhere. It builds on the work Venti has already done globally: its driverless vehicles have moved more than 360,000 containers over 500,000 miles, positioning themselves within an inch of their targets.
The implications are massive. Supply-chain inefficiency is an estimated $600 billion problem every year, and Venti’s trucks can cut transportation costs 40 to 70 percent.
This is what it looks like when AI stops generating answers and starts moving goods in the real world.
Congratulations to the Venti team. We’re proud to back them, and excited for what’s next.
The most common AI tools generate answers by writing text, like ChatGPT or Claude. But a lot of real-world tasks don’t need a written answer. They need a decision.
That’s the idea behind a new kind of AI. Where a generative model writes out a response step by step, a decision model makes the call directly. For the simple, repeated decisions businesses make millions of times, that’s faster and cuts costs.
@Principal_VC portfolio company @liquidai released models to meet this need last week, built to make fast, structured decisions, and small enough to run locally on devices, where the decision matters most.
It points to a shift I keep coming back to: as intelligence becomes abundant and cheap, the advantage moves from having the biggest model to deploying the right one, purpose-built, efficient, and right at the point of action.
A new federal task force will study AI's risks and define the government's role in managing it. The news gets at a question I've been thinking about for a while: who gets to set the terms for how this technology is allowed to behave.
I've argued before that this shouldn't rest with the companies building these systems alone. The next version of the question isn't who holds that authority, but how the players involved, like companies, government, independent experts, and the public, share it.
The framing of these efforts usually lead with the US staying ahead in the AI race, and that's valid. But being first isn't the same as getting it right.
When the goal is defined entirely as winning, it can crowd out the harder questions, like how we build the kind of regulatory ecosystem that supports the advancement of this technology and holds it accountable at the same time.
Because "safe" isn't something any single actor declares. Not one company, not one task force. It emerges the way robust systems do: from many participants and checks working together.
A step toward some government oversight is welcome after years of largely leaving the question to the companies themselves.
But the goal isn't to move responsibility from one set of hands to another. It's to widen who shapes the terms, so the people who will live with these systems have a real say.
When Moonshot AI released Kimi K3 (@Kimi_Moonshot), an open model rivaling the top U.S. systems at a fraction of the cost, it resurfaced a question about the distribution of power in the AI age: will it stay with a few companies, or reach the many?
The debate often gets flattened into a binary — for open-weight AI, or against. But this isn't really about that. It's about what scientific progress actually requires.
AI was built on open science: public funding, shared datasets, research published for anyone to build on. Advances made by @GoogleDeepMind's AI system AlphaFold in predicting the shapes of proteins show us what this openness can unlock.
In publicly sharing its methods and code, DeepMind let more than three million researchers across over 190 countries build on the work, using it to study disease and develop medicines that advance humanity.
Open weights let AI research spread the same way. Without that openness, a handful of companies could control the most capable systems and how much the rest of the world is allowed to learn about them.
Still, there are real risks. Once weights are released, they can't be recalled. Safeguards can be stripped away, and powerful capabilities can be repurposed for harm. This is exactly why researchers, policymakers, and developers need to work in partnership to create the guardrails that will advance this technology and protect our communities.
I wrote all about the case for open-weight AI in my latest for the @Principal_VC newsletter. Read it here: https://t.co/9ZYzn7U26a
AI's biggest bottleneck right now isn't chips or models.
It's permits, power, and land.
@nvidia's earnings showed demand isn't the question anymore. The hard part is building fast enough in the physical world to meet it.
I sat down with @BrianSozzi on @YahooFinance's Opening Bid to get into it.
We touched on the opportunity for a better AI buildout that serves companies and communities, the global demand the "AI is peaking" crowd is missing, the real lessons from the Hugging Face hack, and more.
Watch the full conversation on the @PrincipalVC Substack here: https://t.co/BBvGxByo85