Ben Brooks

Nearly every AI firm is invoking frontier risk to justify a persistent gap between open and closed models. But this application of the precautionary principle deserves more scrutiny than it gets. Restricting access to useful technology—models that will, in their developers' own words, transform the economy—shouldn't be our primary response to uncertain risks. Check it out at @aif_media! There are already a bunch of reasons firms might not release their best models openly: cost recovery, competitive pressure, anxious investors. But we should be skeptical of efforts to freeze open-source behind the frontier under the guise of risk.

One week, two starkly different visions for national AI regulation. Compare the White House framework from @DavidSacks @mkratsios47 @sriramk to what's currently before e.g. the UK Parliament.

To "create" a chatbot or model that can "encourage" crime would be an offense punishable by 5 years imprisonment under amendments voted through (203 to 148) by the UK House of Lords. This follows recommendations from the Center for Countering Digital Hate. The imprecision in these proposals is appalling. Chatbots and language models can be goaded into saying bad things. When they're used by billions of people, even the most cautious developers cannot exclude that risk. Criminalizing their development based on whether they generate any proscribed content (per sec. 1) is a garish attack on fundamental research and consumer technology. Governments can uplevel AI safety in many ways, some of which are represented in this bill. But this kind of theatrical overreach is setting back meaningful reform.

We announced at @NVIDIAGTC that we're joining @nvidia's Nemotron Coalition to advance open frontier models. At BFL, we develop multimodal generative models for visual intelligence, ranging from images to real-time video and action prediction models. We've always been convinced that open models help drive frontier capabilities. We're focused on building the next state-of-the-art open multimodal models in partnership with NVIDIA.

These product integrations (which are awesome) are going to quickly crystallize the competition issues in AI. Normal RAG search surfaces information from the public web. Normal Maps surfaces information from Google, but users still have to do a lot of manual interpretation and decision making. If Maps now surfaces only a handful of recommendations based on entirely on (i) Google-held information about local businesses and (ii) Google-held feedback from customers, the platform is going to become an exceptionally sticky walled garden. Google already has >50% market share in online reviews, and the center of gravity is going to shift further away from the Yelps, TripAdvisors, and Trustpilots of the world. That's great from my perspective as a Google / Android devotee, but also a canary in the coalmine for the market power that existing platforms will amass with AI search.

We present a research preview of Self-Flow: a scalable approach for training multi-modal generative models. Multi-modal generation requires end-to-end learning across modalities: image, video, audio, text - without being limited by external models for representation learning. Self-Flow addresses this with self-supervised flow matching that scales efficiently across modalities. Results: • Up to 2.8x faster convergence across modalities. • Improved temporal consistency in video • Sharper text rendering and typography This is foundational research for our path towards multimodal visual intelligence.

@AdamThierer This acronym 🤌🏼

Sen. Blackburn’s radical AI regulatory proposal is at odds with the thrust of almost everything the Trump administration has proposed or implemented thus far on AI policy. I am quite certain she was not proposing this at Trump's behest.

14110, 2024/1689, 1047, and now 3252. A handful of numbers burned into the retinas of anyone who touches AI policy

There's a legitimate case to rescind the DoW contracts, and even designate Anthropic (in the narrowest 10 USC 3252 sense) a supply chain risk. But this argument that conceding to Anthropic might give rise to an "independent power structure" is unconvincing. It continues to indulge this frustrating trope that frontier AI will necessarily remain a concentrated resource. Regardless of whether folks think AI is more like nuclear or normal technology, that just isn't the case. This is an increasingly competitive technology. The underlying breakthroughs are replicable. The value is orchestration and integration, not just raw capability. The zone of usefulness (i.e. the literal frontier plus the "good enough" area behind it) grows bigger every day. Despite the compute and capital bottlenecks, there are more players at the frontier than major telcos or airlines. Anthropic itself acknowledges that this isn't a one-horse nuclear monopoly with a big moat. That's why they retired their earlier RSP framework—on the assumption that if Anthropic holds back, other firms (OAI, GDM, MSL, xAI etc) will push the frontier in their absence. It's entirely possible that the frontier will re-centralize as money and chips dry up. But instead of threatening to e.g. compel hosted services under DPA, government could more usefully apply itself to ensuring the market is robust, competitive, and diverse. We don't need to resign ourselves to an enduring monopoly.

AI can be open and responsible—not one or the other. For serious harms, our models demonstrate >10x fewer vulnerabilities than other open models. We're sharing these results to show that high performance, open weights, and sensible mitigations can go hand in hand. Our FLUX.2 tops open-weight leaderboards. We're the only American or European lab to rival DeepSeek by likes on @huggingface. And our competitors are some of the largest companies in the world. Our @bfl_ml team is working hard to mitigate emerging risks and keep AI open.

Maybe we won't need all those data centers...

It's bizarre that this is still a thing. Everyone should reject "global AI governance", whether they are pro-race, pro-rules, or pro-pause. The overwrought IAEA / IPCC rhetoric from the past few years is a fig leaf for doing absolutely nothing, in any direction, except fearmongering our way towards a corporate monopoly on critical technology. All politics is local: action happens when there's a domestic mandate and overseers with teeth.

.@AFP: U.S. ‘totally’ rejects global AI governance france24.com/en/technology/…

If anyone knows anyone who has an Artemis II launch ticket but can't make the new date, please shout out 🙏🥹

A really nice critique of AI today. Not so much "art is creative choices" but "art is iteration and discovery". It reads like two objections: (1) The interface is frictionless (e.g. single turn prompting) and lossy (i.e. natural language), and the models over-optimize for the existing distribution. Together, that makes it hard to manipulate, explore, or extrapolate out of the distribution; and (2) The mapping of inputs to outputs (or one modality to another) assumes art is about translating "intent to execution", when we shouldn't presuppose intent in the first place. Not sure I agree with the first—if a manifold is large and rich enough, surely even the most creative and inventive works are ultimately acts of interpolation in the distribution? There is no going off the manifold in that world, no matter whether you're wandering or zero-shotting. All anyone can do is probe the interstitial space. The second resonates. The goal shouldn't be to execute a vision, but to help uncover that vision.

@bfl_ml @NanoBanana Thanks @RapidataAI and @huggingface for this illuminating benchmark

@axios That's unfortunate. There are plenty of more onerous, less precise bills to target instead. HB286 advances some interesting ideas without the baggage from previous catastrophic risk proposals.

White House pressures Utah lawmaker to kill AI transparency bill trib.al/27pvvj2

Zoe Hitzig: nytimes.com/2026/02/11/opi…

This beautiful critique of ads is exactly why open weights are so important for private and secure AI stacks. >50% of ChatGPT messages seek information or guidance. People sharing their deepest secrets with third-party chatbots have no technical guarantee of privacy.

* Endnote: There are plenty of jurisdictions where constitutions truly are just aspirational vibe laws. And in places like the UK, parliament could repeal the "constitution" in an afternoon. But I think most people attach greater weight to "constitutions" than, say, Moscow. Full paper here: anthropic.com/constitution

I think Anthropic uses the rhetoric of constitutionalism in ways that are deeply unhelpful. When people think of a constitution, they don't think of "aspirational technical and social outcomes", which is the new Claude Constitution in a nutshell. No, if you select the word "constitution", you call to mind two attributes: regulation and supremacy. The constitution enforceably regulates the interactions between and among governments and people. And the constitution is supreme to other rules. But here, we're told: > Claude's Constitution is a vision that guides Anthropic, but it isn't entrenched in Claude in any technically, procedurally, or legally meaningful sense. > They hope Claude agrees, because enforcement may be difficult without its consent. > Claude (a bundle of models and systems) is a "novel entity" with potential moral patienthood distinct from—and with interests potentially orthogonal to—its developers (founding fathers), users (citizens), third-party stakeholders (subjects), or other computer systems (peers). (Significantly, I think, Anthropic doesn't use moral "personhood", a term that might imply Claude actually owes a moral obligation to others, and isn't just owed moral obligations from others.) These aren't really consistent with constitutionalism as its understood on main street.* It feels like the same corporate theater we've seen across industry for 4+ years. Making soft institutions and structures look enforceable, when they plainly aren't. The language rings a little hollow after ~3 decades of free rein in tech. I'm reminded of Justice Scalia: "every banana republic in the world has a bill of rights". To be clear, I think Anthropic is thoughtful and earnest. Amanda and the wider team are top of their respective games, and this document is a rich read. But the faux legalism rubs me the wrong way. If AI is even half as consequential as Anthropic keeps telling us, then fair labelling matters in governance. As a community, we'll have a more productive dialogue about AI safety if we're ruthlessly clear-eyed about the limitations of corporate commitments, and cut the theater.

While discussing the impact of AI on surfboard manufacturing, a young friend unironically described the mid-90s as "the 1900s". Way to start 2026, the second quartile of the century