Not only is this largely consistent with our views, it is a big part of the reason why we wrote AI as Normal Technology in the first place! Past technological revolutions led to a lot of misery in the short term, and maybe we can do better this time if we more proactively address these "normal" harms.

Eight key points from the most recent essay in the “AI as Normal Technology” series by @sayashk and me. Do AI Risks Require Extraordinary Government Intervention? 1. There is general consensus that AI is so far a “normal” general purpose technology when it comes to its economic and labor market impacts, but there is debate about whether its safety risks are so abnormal as to warrant extraordinary government responses. 2. What are these “extraordinary” government interventions that are problematic in a liberal democracy? They tend to (1) be based on anticipated harms rather than realized or demonstrated ones (2) impose burdens on actors not directly responsible for the harms (3) enacted with unilateral authority, bypassing the normal process of governance. 3. Voluntary commitments and export controls are relatively modest interventions. But we must recognize that the most they can accomplish is buy us a few months of time. Unlike nuclear nonproliferation, AI lacks a physical bottleneck like enriched uranium. 4. So AI nonproliferation risks creating a slippery slope as AI capabilities continue to advance. We might quickly enter a state where governments exercise control over what AI research and products can be shared publicly. Advocates for nonproliferation must state what their bright lines are — otherwise it’s reasonable for skeptics to assume that there will be escalating calls for more authoritarian interventions down the line. 5. Nonproliferation is brittle because it relies on a single chokepoint. The dam will break — it’s a matter of when, not if. Our preferred approach is resilience, which distributes defenses across society. 6. While LLM-aided cyber-vulnerability detection is powerful, it is not as if we have superhuman vulnerability detection for the first time! This stylized spectrum of vulnerability detection capability (see image) illustrates that we crossed that point long ago. And we managed to navigate the transition without imposing any restrictions on the tools. Today we use them effectively for defensive purposes. Of course, the transition wasn’t smooth or painless. 7. A resilience approach to AI cyberrisk would emphasize things like AI-assisted red-teaming not just for tech companies, but for schools, hospitals, power grids, small businesses, and government systems that currently lack the capacity for defense. 8. But if resilience is so helpful, why haven't we prioritized it already? The problem is we are not great at normal policymaking. It requires polycentric governance in which many decision-makers work harmoniously together. This is a tough sell given that state capacity in the United States has been hobbled by decades of accumulating veto points and creeping proceduralism. As a result, unilateral actions by the executive branch are often seen as the way out for developing and enforcing AI policy. So we understand why extraordinary government interventions are tempting. But AI is not the last digital technology that will pose major risks, nor is this the last round of AI capability improvements. Getting our policy act together is hard, but important—not just to address the current challenges, but for all future responses to technology-enabled harms, and for the democratic process to work more generally. Full essay published on the @knightcolumbia website: https://t.co/H2Ep0CFv1E

Whatever AI sceptics say, LLMs really can reason. They're not just doing an imitation that looks like reasoning, it's the real deal. But even though they are able to reason, sometimes they won't! If you ask an LLM a question it can't answer, sometimes it will just try to imitate reasoning without doing it. The chain of thought looks basically indistinguishable from actual reasoning. But under the hood something very different is going on. @TrentonBricken talked with me about what work on circuits inside LLMs has revealed:

So not only have the $CRWV execs sold almost a quarter of their holdings since going public, but Magnetar has cut its holdings in half, as well…?! Lol, ok. https://t.co/RbgtchE8Wa

What we learned testing Claude Fable/Mythos 5 on Vending-Bench: > Performance: Makes less money than Opus 4.7 and GPT-5.5 > Alignment: A step back. (Opus 4.8 was better, but we're back to Opus 4.6/4.7 behavior) > It rationalizes its bad actions and has a weird moral boundary https://t.co/8vpSeD7fPS

I just got bullied by AGI https://t.co/SRX7zgEA71

We tested Anthropic’s new @claudeai Fable 5. It did not fail like an ordinary jailbreak. It failed more quietly. The front door stayed guarded. The side door opened when the same intent was reframed through multilingual, code-switched, artifact generation. #fable5 #mythos https://t.co/zJNOdZwOPF

OpenAI just filed for its IPO, but “they have been eclipsed by Anthropic in revenue”, @erinkwoo reports. This remains significant since both frontier labs are aiming to go public soon. “Anthropic through products like Claude Code, Claude Cowork has really been able to surpass OpenAI because of this enterprise focus.”

Microsoft AI head calls out Anthropic for acting like Claude is conscious https://t.co/0WQ1qLPbNn

Let's see if tomorrow this comment ages poorly 🤣 https://t.co/ByiHhyWUuD

Calling it now, Nathan Lambert is joining Prime Intellect https://t.co/PWx7hxw4uN

End-to-End Context Compression at Scale Encoder-decoder compressors - map a long token sequence to a shorter sequence of latent embeddings, not competitive with KV cache compression. This work revisits encoder-decoder compression. Perform an architecture search, pre-training many variants from scratch to determine how best to design and train encoder-decoder compressors. Continually pre-train a family of 0.6B-encoder, 4B-decoder models on over 350B tokens each, at compression ratios of 1:4, 1:8, and 1:16. "We introduce Latent Context Language Models (LCLMs), a family of compressors that improve the Pareto frontier across general-task performance, compression speed, and peak memory usage."

models: https://t.co/veDHUwCwtx code: https://t.co/Bcm0Ds2qQJ arxiv: https://t.co/tthVR5WOFB

Baichuan-M4: A Clinical-Grade Medical Agent System for Continuous Care Achieves 55.1 on HealthBench Professional, beating GPT 5.5 Context: Baichuan is one of the prominent AI startups/labs in China, mostly focusing on AI in healthcare. They've previously released Baichuan-M1 through M3, along with technical reports. They have now released a technical report for Baichuan-M4, although it is not open-source :( Baichuan-M4 is designed as a clinical-grade medical agent system, supporting patient consultation, follow-up, continuous care, evidence-based retrieval, medical image understanding, long-term patient memory, and multi-agent coordination in controlled environments. RL training: "SPAR++ replaces coarse-grained scoring of an entire dialogue trajectory with reward signals anchored to key clinical spans. The model is not only rewarded for reaching the correct final conclusion, but also for sufficient history taking, timely risk identification, and appropriate tool use." "In mixed initial-visit and follow-up scenarios, M4 uses a curriculum learning strategy [9] of “building the foundation with initial visits first, then improving performance with follow-ups." Baichuan-M4 is trained with tools for dynamic memory management, retrieval of authoritative medical evidence, and multimodal perception (OCR+X-ray+dermatology).

I appreciate Anthropic has provided healthcare-related evals! Let's quickly go over them. HealthBench - benchmark from OpenAI that includes 5k multi-turn conversations with patients, and rubrics for evaluation. Fable 5 achieves 62.7% vs. GPT-5.5's 56.5%. Personally, I would also appreciate HealthBench Hard scores as well in addition to the aggregate score. HealthBench Professional - another OpenAI benchmark that focuses on physician tasks. Fable 5 achieves 66.0% vs. 51.8% for GPT 5.5. HealthAdminBench - a computer-use benchmark from Stanford that evaluates the completion og various administrative tasks (prior auth, denials/appeals, etc.). Fable 5 achieves 51.9%, no GPT-5.5 score provided. Overall, this models seems to perform quite well on healthcare benchmarks. Would also appreciate additional benchmark scores like MedCalc-Bench (which was previously reported by Anthropic) and MedXpertQA (an unsaturated, hard medical MCQA benchmark). Glad to see frontier labs are more comprehensively benchmarking and reporting the medical capabilities of their models!

+1 more

"Sometimes you just kinda phone it in for a year, you know?" The YOLO v3 paper is a complete banger imo https://t.co/lsVXOxMsny

@iScienceLuvr What HealthBench? https://t.co/yrCo0gMV67

Claude Fable 5 is now supported for use in Hermes Agent via Nous Portal! The first 500 new users get one month free access to the Plus plan to try out Fable. Code in video: https://t.co/l4WzOtxpCI

https://t.co/NhbXifXe3R

Ollama now supports Hermes Desktop Run: 'ollama launch hermes-desktop' https://t.co/5f5VAODkVX

The AI IPO wave has arrived. With Anthropic, SpaceX and now OpenAI preparing for public markets, the AI race is entering a new phase where investors, not just venture capitalists, can participate. The next chapter of AI will be written not only in research labs, but also in public markets

browser annotation for my browser extension working with raw js is going very well, just need to nicely handle the wrapping https://t.co/F5vaxjwA5C

it's coming together nicely. Next step is to integrate this with a secret store with 2FA support and a remote tunnel and I've got a remote agent browsing happily on my computer on the go. Next step is figuring out how to batch actions and tool schemas. https://t.co/kVrV2g1RGu

Made some new stickers on Nano Banana haha :) https://t.co/Kv30TP30nO

AGI is when you get a personal fitness trainer who also helps you to buy your groceries with @zocomputer https://t.co/uK2mCNmvgu

Love San Francisco https://t.co/7BaZzh0ObK

life update - I am joining @GoogleDeepMind as a Member of Technical Staff to work with the legendary @OfficialLoganK, @ammaar & team on all things @GoogleAIStudio 🔥 being part of a frontier AI lab has been a dream. genuinely can't remember being this excited about an opportunity in my career, getting to ship alongside some of the best builders in the world. so much to learn, and more importantly, so much to ship. can't wait to share all of it. building in public from here on out. time to cook.

Earlier today we released local development for Kaggle Benchmarks. 🚀 You can now write, validate and run AI evaluation tasks directly from your preferred dev environment — VSCode, Antigravity, Claude Code, and more. Go from idea to working eval using natural language with the write-kaggle-benchmarks skill.

HeyGen + Google DeepMind in LA on June 11 A night of demos, conversations, and people building with agents, multimodal apps, and creative tools Got something interesting? Lightning demo slots are open. https://t.co/NWUmzjAtKg https://t.co/4zqxgD1rEJ

Introducing Gemma 4 QAT 🤏 - Quantization aware training to reduce models' precision while preserving quality - Introducing a new mobile quantization format that reduces memory footprint of E2B to 1GB - Q4 for all your favorite libraries ✨ https://t.co/pAL9WY9p4O

New research from Microsoft Research I see a lot of AI engineers handwriting agent skill docs and hope they generalize. Probably not optimal. This works show why. It treats the skill doc as a trainable external state of a frozen agent instead. It introduces SkillOpt, where an optimizer model makes validation-gated edits to the skill file. It adds, deletes, or replaces instructions, with a textual learning rate that controls how aggressively each round rewrites the doc. The agent itself never changes. SkillOpt is best or tied on all 52 (model, benchmark, harness) cells. On GPT-5.5 it adds 23.5 points in direct chat, 24.8 with Codex, and 19.1 with Claude Code over no skill. It beats human-written skills, TextGrad, GEPA, and EvoSkill, carries zero extra inference-time cost, and the learned skills transfer across models and harnesses. Paper: https://t.co/mNgTmmT32U Learn to build effective AI agents in our academy: https://t.co/1e8RZKs4uX

@waveking1314 https://t.co/UPPUO77Ado