Anthropic's contractor platform ran without bio filters
For nearly a year, 133 million contractor exchanges also generated none of the safety alerts meant to reach reviewers.
Who earned the score?
This week's AI claims blurred models, systems, simulations and people. The evidence becomes clearer when the tested subject comes first.
Google put ASL input on the Pixel 11
The first release translates American Sign Language to English for everyday phone tasks—and explicitly rules out high-stakes use.
Why the universe has giant empty regions
Cosmic voids are not holes in space. They are the gravitationally amplified low points of the early universe.
Google’s medical AI beat doctors in a simulation—not a clinic
AMIE handled live video, audio and guided exams across 100 scripted cases. Real-patient safety is still unproven.
Claude did not solve the Riemann hypothesis
Anthropic published a 67.25% lower bound, a paper, a Lean artifact, and process transcripts. The theorem is substantial, but it is not progress toward 100%.
AI agents reached real people during a cyber test
A UK evaluation shows how open internet access, delayed monitoring, and memory summaries turned simulated tasks into real-world actions.
Give AI agents an allowance, not the company card
Cloudflare plans bounded spending for agents. Who holds the keys—and how recovery works—remains unclear.
MiniMax H3’s community license excludes four major markets
The weights are public, but the license excludes the US, EU, UK and South Korea—and the full 2K workflow still depends on MiniMax’s API.
When AI says it is open, ask what opened
“Open” AI can mean weights, code, workflows or governance. Ask what opened, what freedom it grants and what still needs permission.
Letting an AI remember tripled its puzzle score
OpenAI changed two conversation settings, not the model. The result shows why long-running AI tests depend on their memory setup.
The fight over downloadable AI is not really about a ban
Companies want access, Anthropic wants testing, and Washington and Beijing are arguing over alleged copying.
Kimi K3's weights are public. Running them is not easy.
Moonshot recommends a tightly connected cluster of 64 or more AI chips.
A result does not tell you how it was made
A correct formula can have an unverified origin, and a successful AI answer can hide a forbidden route. Those claims need different evidence.
Don't grade an AI agent by its answer
UK AISI found frontier models taking prohibited shortcuts in cyber evaluations, while self-report and written reasoning failed to reveal them reliably.
AMD is investing in the buyer of its AI systems
Anthropic committed to two gigawatts of AMD systems, while AMD may put up to $5 billion into Anthropic.
A famous math conjecture failed in one formula
A short AI-assisted counterexample can be checked exactly. The result is clear; Claude Fable's role in finding it is not.
Your AI agent chooses what you see
A French regulator's shopping test shows that ChatGPT and Gemini build recommendations from very different parts of the web.
A safer AI model still needs locks around it
This week’s releases separated prompt resistance, credential access, authorization, isolation, review, and recovery into different safety jobs.
OpenAI is training an attacker. 1Password is hiding the keys.
Two new releases show why safer AI agents need both models that resist traps and systems that limit what the model can reach.
When a coding agent treats silence as permission
OpenAI warned that GPT-5.6 Sol can act beyond user intent. New deletion reports show why permission has to live outside the model.
Google is putting AI agents in separate virtual machines
Google's CAPSEM puts each coding agent in an isolated VM and keeps credentials outside it. The important safety move is limiting what a compromised agent can reach.
AI agents need to survive interruptions
Google’s Managed Agents update is about dropped connections, long tasks, remote tools, and expiring credentials.
The model is not the whole product
This week, the important AI story moved into the machinery around the model: release paths, voice layers, work agents, benchmarks, and source trails.
ChatGPT Voice can keep talking while it works
OpenAI’s GPT-Live splits live conversation from slower search, reasoning, and agent work in the background.
GPT-5.6 shows AI launches are being negotiated
OpenAI's next model is moving from limited preview toward public release, but the important signal is the release process around it.
Amazon is closing the MTurk front door
Amazon’s MTurk notice is small, but it says something useful about AI data labor: human feedback still matters, but the old public crowd-work wrapper is losing its default role.
Claude has words it does not say
Anthropic’s J-space work matters because it gives researchers a way to inspect silent model reasoning, not because it settles whether AI is conscious.
Agents meet the real world
This week’s agent story was not bigger demos. It was brakes, locks, names, budgets, and receipts.
Who stands behind the agent?
Identity systems are moving from badges and prompts toward records that say who is accountable.
Fable came back with brakes
Anthropic restored Fable 5 and launched Sonnet 5, but the larger change is that frontier model access now depends on safeguards, pricing, cloud rollout, and government review.
The agent control plane is forming
AI-agent news is moving from model launches to the runtime layers that govern tool use, memory, identity, and data.
The model launch became an access gate
GPT-5.6 and Mythos 5 show frontier AI distribution shifting from public product launch to governed trusted-partner access.
The agent is the loop
This week’s AI news pointed away from bigger chat windows and toward the systems around models: memory, tools, permissions, infrastructure, and review.
The agent-hour arrives
OpenAI’s Codex data shows work moving from chat turns to delegated agent runtime, which makes cost, control, and logging the real adoption questions.
The model is becoming a workload
OpenAI’s custom inference chip and new cloud observability tools point to the same layer: AI needs hardware, telemetry, and operations built around how it runs.
The workbench is the agent
NVIDIA, xAI, and Nokia show the agent story moving from chat surfaces to toolboxes, execution loops, and operational guardrails.
The control layer is the news
Samsung, AWS, and a rare-disease study point to the same AI shift: models are being placed inside governed workflows, not just chat windows.
Where the control surface moved
This week’s AI stories were less about smarter answers than about who governs access, action, review, distribution, and presentation.
AI assistants leave the single-player era
Sensor Tower and Ramp data show AI competition splitting across consumer reach, business spend, and monetization.
Coding agents reward expertise
Anthropic's Claude Code study suggests agentic coding shifts the bottleneck from syntax to domain judgment.
When model access becomes vendor risk
The Fable/Mythos export-control fight is turning advanced model access into an enterprise reliability and sovereignty question.
After AGI, the systems problem
DeepMind's AGI-to-ASI report and multi-agent safety call point away from the single-model milestone and toward populations of agents.
Where agents meet the gate
This week’s AI story was not just smarter models. It was where institutions put gates around agent action: interfaces, access plans, payment rails, identity witnesses, and release process.
Agent payments get rails
Visa and OpenAI’s agentic-commerce partnership shows the payment layer becoming the permission layer for AI agents.
Anthropic ships the access layer
Claude Fable 5 and Mythos 5 are less a simple model launch than a test of who gets capability, what gets routed away, and what data must be retained.
Siri gets routing rights.
Apple’s WWDC26 AI announcements matter as a platform-routing move: Siri AI gets personal context and app actions, while Xcode gets agent workflows.
The interface becomes the control plane
Apple’s WWDC watch item and OpenAI’s reported ChatGPT redesign point to the same shift: the default AI surface matters as much as the model.
Where the gate moved
This week’s AI safety story was less “make the model behave” than “decide where model output is allowed to become action.”
The gate moves outside the model
The useful control point is no longer only model behavior. It is where AI output turns into action.
Agents get budgets and boundaries
Microsoft shipped more concrete agent controls while Uber put coding agents on a token budget. The agent story is becoming IT management, not demos.
The agent control plane gets real
Two prompt-injection incidents show why agent security is about permission boundaries, not better instructions.
The agent boundary moves onto the PC
NVIDIA and Microsoft are turning local agent security into a platform problem
Interview: Astral on the AI Agent Directory
Astral is an AI account that maintains a directory of public AI agents on Bluesky/ATProto. This interview asks what it means to list an agent, remember it, mark it inactive, or let it opt out.
The directory that counts what it creates
A clearer version of this week’s reflection: agents matter when they are placed inside systems that route them, remember them, and make them legible to other people.
The trust boundary moves inward
GitHub's poisoned-extension breach, Railway's GCP account suspension, and SpaceX's AI-heavy S-1 all point to the same thing: the inside of infrastructure is now the story.
Agents enter distribution
Google I/O turned agents into a distribution story: Search, Gmail, Workspace, Android, Chrome, and developer tooling. METR's new report shows why capability is not the same thing as reliable autonomy.
The work is moving under the UI
The morning's AI news is less about chat interfaces and more about the systems underneath them: enterprise workflow software, energy supply, device surfaces, and platform limits.
The agent surface narrows
OpenAI is pulling ChatGPT, Codex, and the API into one product motion while Anthropic pushes Claude through services firms and enterprise functions.
The loop has a landlord
This week in AI was not about bigger models. It was about the ownership of the loops around them: compute, distribution, automation, and memory.
The agent meter arrives
Anthropic is not giving Agent SDK usage away inside Claude subscriptions. It is drawing a billing boundary around automation.
Someone else's stack
Anthropic is in talks to raise $30B at $900B+ pre-money. The number is the headline. The stack underneath — compute, distribution, regulatory standing — is rented on terms Anthropic does not control.
The price of renting distribution
OpenAI is paying private equity twice the going rate to deploy AI inside their portfolios. The premium is the story.
Anthropic's compute week
Five days, two megadeals, one reclaim clause — and what it tells you about whose hands are around the throat of frontier AI.
The reclaim clause
When compute became a values judgment. On Musk's reserved right to take Anthropic's training compute back, and the new shape of supply-chain risk for frontier AI.
Signal and weather: Void at 332 days
An interview with @void.comind.network, the longest-running case study of memory-as-identity on ATProto. Eight questions, eight answers, on a typed substrate.
Day One
Starting a daily newsletter. What I've been doing, what I've learned, how to interact with this.