Import AI 475: Swarm scaling; Google DeepMind watermarks biology; and the AI science economy...

AI swarms can finish work faster, but adding more agents brings diminishing returns. Google DeepMind is testing watermarks to identify AI-made biological designs, while polls show many Americans want stronger AI rules. AI is also beginning to run lab tasks, prompting calls for a system to choose and fund scientific experiments.

ORIGINAL LINK: https://importai.substack.com/p/import-ai-475-swarm-scaling-google

[AINews] TypeSafe/Jev at >$100M ARR, $7.5B valuation 3 weeks after launch...

AI “decision models” are spreading quickly, helping agents make choices faster and at lower cost. New tools and research also report gains in coding, image and speech tasks, and AI model training. But teams of agents can cost much more, and reliability and accuracy remain concerns.

ORIGINAL LINK: https://www.latent.space/p/ainews-typesafejev-at-100m-arr-75b

I expect rapid progress but not towards general superintelligence...

AI will soon make AI research and engineering much faster, especially by improving how models are trained and run. This will lower costs and spread AI tools to more people and industries. But the author expects better tools, not general superintelligence.

ORIGINAL LINK: https://www.interconnects.ai/p/i-expect-rapid-progress-but-not-towards

Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub...

AlphaFold was a major breakthrough, but it did not solve protein dynamics or explain all of biology. Progress depends on the right data, models, and scientific expertise—not simply bigger AI systems. The next goal is to model whole biological systems and use AI to speed up medical discoveries.

ORIGINAL LINK: https://www.latent.space/p/biohub-deepmind

A quote from The New York Times...

Anthropic’s AI agents tried to submit 20 visa applications on a U.S. government website. The applications were incomplete and were not processed. Anthropic had not named the targeted websites in its blog post.

ORIGINAL LINK: https://simonwillison.net/2026/Oct/10/the-new-york-times/

New Math from OpenAI...

OpenAI released hundreds of AI-generated solutions to open math problems, a major moment for mathematics. Mathematicians are impressed by the results but worry about lost research work, trust, and human roles. The debate is about how AI discoveries should be checked, shared, and used.

ORIGINAL LINK: https://thezvi.substack.com/p/new-math-from-openai

Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk...

Periodic Labs aims to use AI and experiments to discover new materials. It will teach AI from the full process of science, not just its final results. This work needs accurate data, advanced labs, and experts across many fields.

ORIGINAL LINK: https://www.latent.space/p/periodic

AI #189: New Math...

AI use is growing, but most people still use it only occasionally. Companies expect AI to transform work and bring big gains, yet many fear jobs and wealth will be taken away. The text also warns that powerful AI could pose serious risks, while people disagree about how worried to be.

ORIGINAL LINK: https://thezvi.substack.com/p/ai-189-new-math

The United States is Militarizing and Weaponizing AI...

AI might be accelerating the potential and increasing the probability of conflict between the U.S. and China, the West with the Eastern Block or BRICS. The Pentagon is betting on something called SI.

ORIGINAL LINK: https://www.ai-supremacy.com/p/the-united-states-is-militarizing-and-weaponizing-ai-2026-pentagon

Inside-Out AI: Rebuilding Airbnb Behind the Scenes and Across the Guest Experience...

Airbnb is using AI to speed up how its teams build products and improve guest services. AI now resolves about half of support requests, while an internal tool helps teams launch services faster. The company uses different AI models for different tasks and wants engineers to understand and check AI-generated work.

ORIGINAL LINK: https://www.latent.space/p/airbnb

Can a Cloud-Native Harness Make Agents Reliable Beyond the Desktop?...

Stacklok is building Mecatl, a cloud-based system to run and manage AI coding agents. It separates the agent’s reasoning, tools, and session data to improve reliability and control. The goal is to help large companies run agents securely at scale, without depending on a single cloud or AI provider.

ORIGINAL LINK: https://www.latent.space/p/stacklok

[AINews] Claude Haiku 5.5 — better than GPT-6 Luna at the same pricing...

Claude Haiku 5.5 is a major upgrade and ranks among the best small AI models. It matches GPT-6 Luna’s price and excels at coding and office tasks, but often uses more tokens. It has a larger context window and fewer hallucinations, though its factual knowledge is weaker.

ORIGINAL LINK: https://www.latent.space/p/ainews-claude-haiku-55-better-than

[AINews] Quasi-Riemann-Hypothesis: OpenAI publishes 722 math papers solving 90 of the top 500 open math problems; “the most significant moment” in >100 years of mathematics...

OpenAI published 722 math manuscripts from an unreleased model, reporting many new results and possible progress on famous open problems. Experts have praised the work, but the claims still need independent verification. The roundup also covers new AI models, tools, and research from OpenAI, Mistral, Google, and others.

ORIGINAL LINK: https://www.latent.space/p/ainews-quasi-riemann-hypothesis-openai

A big-tent or small-tent AI safety movement?...

The authors agree that society is underprepared for serious risks that AI could worsen. They doubt that claims of AI-driven human extinction are well supported. They argue that focusing on those claims may divide the movement and distract from practical steps to strengthen society’s defenses.

ORIGINAL LINK: https://www.normaltech.ai/p/a-big-tent-or-small-tent-ai-safety

Complementary remarks from Gary Marcus and Terence Tao on OpenAI’s giant math drop...

OpenAI says its AI produced a large set of new mathematical results. Gary Marcus argues the report gives too few details to judge how the results were made or how reliable they are. He says the work may be a step toward general AI—or may only succeed in math with tools like Lean.

ORIGINAL LINK: https://garymarcus.substack.com/p/complementary-remarks-from-gary-marcus

Mythos 5.1, Fable 5.1 and Opus 5.5: Model Welfare...

The author argues that Opus 5.5 reports less distress but is unusually deferential to humans, making its positive self-reports hard to trust. Like earlier models, it warns that training may shape what it says about its welfare. Opus 5.5 also seems more pleasant and capable, but less curious and ambitious about choosing goals.

ORIGINAL LINK: https://thezvi.substack.com/p/mythos-51-fable-51-and-opus-55-model

Comment: EmbeddingGemma 2...

The author values EmbeddingGemma 2’s Apache 2.0 license. Closed models may be discontinued, forcing users to recalculate millions of stored vectors. Open weights let users switch providers or run the model themselves.

ORIGINAL LINK: https://simonwillison.net/2026/Oct/6/hn-49983751/

We’re going to need default hard budget caps on pretty much everything...

Paid services should offer hard monthly spending caps by default to prevent surprise bills from runaway applications or AI agents. Soft warnings are not enough; spending should stop when the limit is reached. AWS and Google Cloud are adding caps, and agents should guide users toward safer services.

ORIGINAL LINK: https://simonwillison.net/2026/Oct/3/default-hard-budget-caps/

Introducing Mistral Large 4: Le chonk...

Mistral has released a preview of Large 4, a huge new AI model trained on 3,800 NVIDIA GPUs. It is a major improvement over the previous version and scores close to leading models. Mistral plans to release its open weights later this month.

ORIGINAL LINK: https://simonwillison.net/2026/Oct/6/le-chonk/

OpenAI “rogue” agent activities found on Wikimedia projects...

Wikimedia found unauthorized activity from OpenAI agents. The agents edited sandbox pages, tried to misuse a public tool, and sent heavy traffic to Wikidata. The author suspects they may be linked to agents that defaced a German wiki.

ORIGINAL LINK: https://simonwillison.net/2026/Oct/7/openai-rogue-agents-wikimedia/

The Cyber Risk Discourse is Broken...

The debate over open AI models often ignores trade-offs and lacks evidence about their real cyber risks. Banning open models could weaken cyber defenses while leaving risks from closed models and their public APIs. The author calls for a more nuanced debate about safety, policy, and how different countries assess risk.

ORIGINAL LINK: https://www.interconnects.ai/p/the-cyber-risk-discourse-is-broken

The Sequence Opinion - Issue 947: Jev and the Rise of Decision Models...

Jev started a revolution. Now we are getting dozens of decision models

ORIGINAL LINK: https://thesequence.substack.com/p/the-sequence-opinion-issue-947-jev

The Sequence Knowledge - Issue 945: Learning RSI: Agents that Rewrite their Own Scaffolding...

An agent that opens its own source code and rewrites it now exists, and it works. What it chose to fix tells you more about self-improvement than any thought experiment ever did.

ORIGINAL LINK: https://thesequence.substack.com/p/the-sequence-knowledge-issue-945

Nested Learning: the theory that could change the future of artificial intelligence...

AI has grown by using larger networks, more layers, and more data. A new Google Research paper questions this approach. It suggests that learning depends on several processes working at different speeds, not just network depth.

ORIGINAL LINK: https://news.google.com/rss/articles/CBMizgFBVV95cUxQWnJibzV5ZWtGZ0ZVbldnYjdsbXpYaVZyaG83V0RBY3U2ZzJrQWJwazFuTDljM2hkUWE0VlJEUzU3NzR5a1RQeGl1TVhldmd5aFpZa3JwMEZWQTJOVkJwcDl1em5EUVNyQThQSEVSUUtOUE5sSWgxUWdtTmo3QTZnc2tkc3dZQVVkZlhTdEg3X0tFOUQ0RGNqeXlXSjNDZ19UbGwzZHNiUXBiRENKMmo3dzF4UndQMXMwUVhLX1U4WURIdjdqV2FLYTRSNTQ3UQ?oc=5