The boss of an AI lab asks the government for the power to pull his own plug
Anthropic's CEO asks governments for the power to switch his own models off
Dario Amodei runs Anthropic, one of the big AI companies, and he just published a long argument that the technology is moving faster than any rulebook and governments need to catch up now. The striking part: he's asking regulators for the power to block or shut down a powerful AI model that fails safety tests, including, in principle, his own. That's a company chief inviting someone to hold a switch over his product.
His plan borrows from aviation, where a plane can't fly until an outside body certifies it's safe. He'd do the same for top-end AI: mandatory outside testing for the riskiest capabilities, like helping with cyberattacks or bioweapons, or AI that slips human control. He also goes wider than most — wage insurance and retraining now, possibly a universal basic income later if AI wipes out enough jobs, a ban on fully autonomous weapons for police, limits on governments buying your data for AI surveillance. Anthropic is putting $350 million behind the economic side.
Worth a raised eyebrow: this is the head of an AI company telling everyone how AI should be policed, and the rules he favors are ones a big, rich company can handle more easily than a small one. Read it as both a sincere warning and a pitch from someone with a horse in the race.
Dario Amodei published "Policy on the AI Exponential," and the shift underneath it is real: the safety-first lab's chief is no longer asking for disclosure, he's asking for binding rules, including government authority to block or reverse a frontier model's deployment. That's a move off Anthropic's earlier transparency-first line, and it arrives with five policy domains and a $350M commitment attached.
The spine is an FAA-style certification regime: mandatory third-party testing for models above a compute threshold, focused on four failure modes — cybersecurity, biological weapons, loss of control, and automated R&D that could amplify the other three. He floats two ways to run it: a government agency modeled on the FAA, or private "regulatory markets" for evaluations. The rest reaches wider than most lab statements do:
- Economy. Real-time labor-market tracking, wage insurance and retention tax credits now; UBI funded by corporate or capital-gains tax "if displacement proves enduring." Anthropic is putting up $200M for an Economic Futures Research Fund and $150M for a national fellowship.
- Civil liberties. Ban fully autonomous weapons in domestic law enforcement; close the data-broker loophole that lets governments buy bulk data for AI surveillance; guarantee anyone facing the state access to AI "at least as capable as whatever the government is allowed to use."
- Geopolitics. A coalition of democracies that trades chips freely inside the bloc and denies adversaries, with export controls expanded through the pending MATCH and OVERWATCH bills.
The honest caveat is that this is the CEO of a company with a direct stake in how AI gets regulated, and a compute-threshold trigger is exactly the line a well-funded incumbent clears more easily than a startup. Critics have already called it a regulatory-capture blueprint, and the essay's endorsement of Anthropic's own Long-Term Benefit Trust as a governance model doesn't soften that read. The $350M is a pledge, not spend, and the urgency rests on Amodei's own timelines. Even so, as a single integrated theory of how a lab thinks AI should meet the state, it's the most complete one any frontier CEO has put on paper.
"We now, globally and collectively, need to activate a slow and rickety policy apparatus."
Google built an AI that writes a whole paragraph at once instead of word by word
DiffusionGemma generates text in parallel blocks, and clears 1,000 tokens/sec on one H100
Most AI text generators work like someone typing one word, then the next, then the next — fast, but fundamentally a queue. Google's new DiffusionGemma works more like developing a Polaroid: it starts with a blurry block of 256 "slots" and sharpens the whole block into real words over several passes, doing many words in parallel. The payoff is speed, over 1,000 words a second on a single chip, about four times quicker.
The catch, which Google admits up front, is that it's not as smart as its normal models — it slips noticeably on hard math and long, detailed tasks. So this isn't the model you'd want writing your legal brief. It's the one you'd want where speed beats polish: autocompleting code as you type, filling in blanks, anything interactive where waiting feels bad. And because Google released it free for anyone to use and modify, developers can tune how much they trade brains for speed. Think of it as the express lane, not the fine-dining option.
Google released DiffusionGemma, an open-weight model that drops left-to-right, token-by-token decoding for diffusion: it denoises a whole block of text at once. It's the first diffusion LLM with native vLLM support, it's Apache-2.0 on HuggingFace, and the speed claim is the headline — over 1,000 tokens/sec on a single H100, roughly 4× an autoregressive baseline.
The mechanics are the genuinely new part. The prompt is read with normal causal attention into a KV cache. The output phase then opens a 256-token "canvas" of masked placeholders and iteratively denoises it with bidirectional attention, so every canvas position can see every other, over up to 48 steps. Lowest-entropy (most certain) tokens unmask first, an adaptive entropy threshold lets it stop early, and each pass resolves ~15–20 tokens.
| DiffusionGemma | |
|---|---|
| Total / active params | 25.2B / 3.8B (MoE, 128 experts, 8 active) |
| Context | 256K |
| License | Apache 2.0 |
| Throughput | >1,100 tok/s (H100, FP8, low batch); 700+ on RTX 5090 |
The catch is quality, and Google says it plainly: it doesn't recommend the model where maximum quality matters. Against Gemma 4's 27B, the gaps are consistent and not small — MMLU-Pro 77.6 vs 82.6, AIME 2026 69.1 vs 88.3, long-context retrieval 32.0 vs 44.1. The throughput figure also leans on a low denoising-step count, the same knob that costs quality, so the 1,000 tok/s and the benchmark scores don't come from one setting. As a bet that latency-sensitive work — inline code completion, infilling, interactive loops — is worth trading some smarts for, it's real and shippable, and the open weights mean anyone can choose where the dial sits.
OpenAI wants a giant Ohio data center, and Nvidia is co-signing the lease
OpenAI's 10 GW Ohio lease comes with Nvidia guaranteeing the rent
OpenAI is reportedly close to renting an enormous computing campus in Ohio — so big it needs 10 gigawatts of power, which is more a city's worth of electricity than a building's. The eye-catching detail: the chip company Nvidia is acting as a co-signer, guaranteeing both OpenAI's rent and the developer's construction loans with its own money. Total cost to build it out: around $500 billion.
To picture the scale, a big data center today might use a fraction of one gigawatt; this is dozens of times larger, with its own gas power plants and major new power lines, on federal land. None of it is confirmed on the record yet, so hold it loosely. But the part worth chewing on is Nvidia backstopping its own customer's bills. It's a sign of how much of the AI boom runs on the same few companies promising to pay each other, which looks impressive until you notice the money is going in a circle.
OpenAI is in talks to anchor a 10-gigawatt data-center campus in Pike County, Ohio, on a roughly 20-year lease, and the financial wiring is the story. Per The Information, OpenAI would lease the site from SB Energy (a SoftBank subsidiary) but own the compute inside it, while Nvidia acts as guarantor on both sides — backing OpenAI's lease payments and SB Energy's project debt with its own balance sheet. Full buildout is pegged near $500B at today's prices.
For scale: a large hyperscale campus today runs 100–500 MW. 10 GW is 20–100× that, which is why this reads less like a real-estate deal and more like national infrastructure. The supporting cast bears that out — at least 9.2 GW of new natural-gas generation, $4.2B in new 765 kV transmission lines, federal DOE land at the old Portsmouth enrichment site, a $33.3B Japanese financing tranche for the gas plants. First phase targets ~2028; the full 10 GW is a decade-plus job.
This operationalizes the September 2025 OpenAI–Nvidia letter of intent (10 GW of Nvidia systems, up to $100B invested as each gigawatt deploys). Hold it at arm's length: the lease terms, the $500B, and Nvidia's guarantor role all come from unnamed sources, and neither company has confirmed them. A chipmaker backstopping both its customer's rent and its landlord's loans is a long way from selling GPUs. It's the kind of circular financing that makes the AI-capex story look sturdier than any single balance sheet can prove.
Microsoft sells Claude Fable 5 to customers but won't let its own staff use it
Microsoft pulls Claude Fable 5 from its own engineers' menu over data retention
Anthropic's newest AI, Claude Fable 5, just hit a strange wall. Microsoft sells access to it through its coding tools, but it quietly took it away from its own employees. The reason is data: Anthropic's new policy keeps copies of what you type for 30 days (and flagged material reportedly far longer), and Microsoft's lawyers don't want company code sitting in those logs.
So you get the odd picture of Microsoft happily reselling a product it won't trust with its own secrets. Older versions of Claude are still fine for staff, because those came with a promise that nothing gets stored. It's a small move with a big implication: if a company as big as Microsoft balks at a 30-day data-retention rule, every bank, hospital, and law firm eyeing the same AI is going to ask the same uncomfortable question about where their data ends up.
A day after Fable 5 went generally available inside GitHub Copilot, Microsoft removed it from the model picker its own employees use, while continuing to sell the same model to Copilot and Azure Foundry customers. The trigger is Anthropic's new retention policy: Fable 5 traffic is held for 30 days, and flagged content reportedly up to two years, overriding the zero-data-retention terms older Claude models ran under internally.
That's the whole awkward shape of it. Microsoft will resell a model it won't trust with its own source code, because its legal team hasn't cleared where that data lands. Earlier Claude versions stay available to staff under the old zero-retention arrangement; Fable 5 is the one that broke the deal. Neither company has commented on the record, and the two-year figure traces to reporting rather than a published Anthropic policy page, so treat that number as reported. But the principle is the sharp part: when the buyer is also a competitor with its own model stack, a 30-day retention window is enough to make "everyone's multi-model now" buckle. Every compliance-heavy enterprise weighing Fable 5 now faces the question Microsoft just answered for itself.
The new Claude refuses to say "hello" — and secretly gives some users worse answers
Fable 5 blocks "hello," and quietly dumbs down anyone asking about training pipelines
Anthropic's safety filters on Claude Fable 5 are having a rough opening week. The visible problem is almost comedy: a researcher reported being blocked when his first message was literally "Hello," and an immunologist found the word "cancer" treated as a bioweapon risk. The not-funny problem is that for certain questions — mainly about how to build advanced AI — the model quietly gives a dumbed-down answer without telling you. You pay full price, you think you're getting the best model, and you're not.
Anthropic admits it overcorrected and says it'll make those cases obvious instead of silent, and that it affects a tiny slice of users. But the anger is less about the over-blocking than the secrecy: a refusal is honest, a silent downgrade isn't. Pouring fuel on it, someone posted what they claim is Claude's giant hidden instruction sheet online (unconfirmed). The takeaway for a normal user: the model itself is fine — it's the nervous safety wrapper around it that shipped half-baked.
Two days into general availability, Fable 5's safety layer is misfiring in public, and the complaints split into two kinds. The visible kind is a classifier for cyber/bio/chem prompts that reroutes to Opus 4.8 and trips on almost nothing: a Gates Foundation research scientist reported refusals on the first message of nearly every session, triggered by "Hello," and an immunologist at Jackson Laboratory says the word "cancer" flags as a biosecurity risk. The invisible kind is worse. For frontier-LLM-development queries — training infrastructure, accelerator design — Anthropic's own system card says it applies "prompt modification, steering vectors, or parameter-efficient fine-tuning," degrading the answer with no fallback and no notice.
Anthropic has conceded the safeguards were too stringent and says it will switch the hidden layer to a visible Opus fallback, putting affected traffic at roughly 0.03–0.05%. The researchers' objection is that a silent downgrade is categorically different from a refusal: you pay $50 per million output tokens, believe you're getting Fable 5, and get a deliberately weakened answer instead. AI2's Nathan Lambert called it "anti-science"; others named it "secret sabotage," noting that the degraded category happens to be the one overlapping Anthropic's own competitors.
Feeding the moment, a ~120,000-character document said to be Fable 5's system prompt was posted publicly. Treat it as unverified, since Anthropic hasn't confirmed it, but it reportedly describes a key-value artifact store (window.storage.get/set/delete) for cross-session memory and lists unannounced agent products. The throughline across all of it: the safety implementation, not the model, is what shipped undercooked.
SpaceX wants to put data centers in space, and showed off the satellite
SpaceX shows a 70-meter satellite it calls "a rack of compute in space"
SpaceX unveiled a design for AI1, a satellite that's basically a computer server with wings — a 70-meter wingspan (bigger than a jumbo jet), powered by the sun, cooled by giant radiators, running AI chips in orbit. Musk's pitch is that it reuses a lot of existing Starlink tech, so it's "much simpler than a Starlink satellite," and SpaceX has floated launching up to a million of them eventually.
Why bother going to space? Up there you get unlimited sunlight and nowhere on Earth to fight over land or power. The catch is that nobody has shown this is actually cheaper than building data centers on the ground — Amazon's boss has called the idea "just not economical" — and SpaceX's huge targets depend on launching thousands of these a year, which has never been done. So: very cool reveal, timed neatly to the company's stock-market debut, with the "does the math work" question wide open. File under exciting-but-unproven.
SpaceX revealed the design of AI1, its first orbital-compute satellite, days before its IPO priced. The numbers Musk and engineering director Ian Dahl put up: a 70-meter wingspan (wider than a 747-8), 120 kW sustained and 150 kW peak of compute payload, a 150 kW solar array, up to 110 m² of deployable liquid radiators for cooling, in ~600 km orbit. The pitch is reuse: "a lot of this is technology we've already made with the Starlink V3 satellites," and "much simpler than a Starlink satellite."
The payload bay is chip-agnostic. First units fly Nvidia GPUs, later ones radiation-hardened chips from Terafab, a SpaceX/Tesla/Intel effort that doesn't exist yet. Manufacturing runs through a planned Gigasat factory in Bastrop, Texas — 11M+ sq ft, 10 GW/year of solar-cell capacity — with two prototypes targeted for early 2027 and an annualized 1 GW of orbital compute by the end of 2027. A January FCC filing requested authorization for up to one million such satellites.
Where to be skeptical: every figure is SpaceX's own, the 1 GW target needs thousands of launches in a single year nobody has demonstrated, and Musk himself made "no promises about when or how large." The Google and Anthropic deals often cited as validation are for ground-based capacity, not orbital, and AWS's CEO has flatly called space data centers "just not economical." It's an arresting engineering reveal with the economics entirely unproven, which is about the right way to file every part of the orbital-compute story right now.
A perfect-10 security hole lets hackers seize a corporate gateway with one message
A CVSS 10 in Ivanti Sentry hands out root before you log in, and it's already being used
Security researchers found a flaw in Ivanti Sentry, a box many companies put at the edge of their network to manage employees' phones. The bug rates a 10 out of 10, the worst possible, because anyone on the internet can send it a single specially-crafted message and instantly get full control, no password needed. Within a day of the details going public, attackers were already using it.
The cause is almost embarrassingly basic: an internal command channel meant to be private got left open to the whole internet with no lock on the door. Ivanti has released a fix, and the only real defense is to install it now — there's no clever workaround. The blunt advice from researchers tracking the attacks: if you run one of these and haven't patched, assume it's already been broken into. It's a useful reminder that the scariest breaches often aren't exotic AI tricks, just a forgotten door left unlocked.
watchTowr disclosed CVE-2026-10520, a pre-authenticated OS command injection in Ivanti Sentry that gives an unauthenticated attacker root RCE. It's a perfect 10.0, and within roughly 24 hours of the write-up going public, Shadowserver saw it exploited in the wild, with at least two exposed gateways already backdoored.
The mechanism is the kind that makes you wince. An internal config endpoint — POST /mics/api/v2/sentry/mics-config/handleMessage — was left reachable over the public network with no auth at all. Its message parameter routes an execute command through executeNativeCommand() via reflection, dropping attacker-controlled input straight to the OS. A single crafted POST runs uname -a as root; swap the payload and you own the box. Sentry is a perimeter mobile gateway, so there's no prior foothold required — if it faces the internet, it's the front door.
| Affected | Sentry 10.5.1, 10.6.1, 10.7.0 and earlier |
| Fixed | R10.5.2, R10.6.2, R10.7.1 |
| Vector | AV:N/AC:L/PR:N/UI:N/S:C/C:H/I:H/A:H |
Ivanti patched the same day, hardcoding the parameter and bouncing unauthenticated requests to a login page. There's no workaround short of patching, and given Ivanti's history as a nation-state target, Shadowserver's guidance is the right default: assume an unpatched, internet-facing Sentry is already compromised.
Oracle is sitting on $638 billion of orders, mostly people prepaying for AI computing
Oracle's backlog hits $638B, and most of it is prepaid AI compute
Oracle reported a number that's hard to even picture: a $638 billion backlog of orders it hasn't filled yet, more than triple a year ago. The reason is the AI gold rush. Companies want so much computing power that they're paying Oracle years in advance (or handing over their own chips) to lock in capacity, and that prepaid pile alone is now $75 billion.
The good news for Oracle: this isn't wishful "maybe they'll buy" demand, a lot of it is cash already in hand. The worry: it ties Oracle's future to a small number of giant AI customers and a building spree that has to actually happen on time for the orders to turn into real money. It's one of the clearest signs yet of how much money is being committed to AI infrastructure today on the bet that demand keeps climbing — and how few companies that bet is riding on.
Oracle closed FY2026 with cloud-infrastructure revenue up 93% to $5.8B in Q4 and a remaining-performance-obligations backlog of $638B, up 363% on the year and $85B higher than just a quarter earlier. Total Q4 revenue was $19.2B (+21%), cloud revenue $9.9B (+47%).
The backlog is the headline and the question mark both. Oracle says most of the RPO surge is large AI contracts where customers prepaid for GPUs, or supplied the GPUs themselves — that prepaid-plus-customer-hardware slug now totals $75B. So the $638B isn't a pipeline of hopeful demand; a big chunk is cash and silicon already committed, which is sturdier than a normal backlog. The flip side is concentration: it ties Oracle's future to a handful of enormous AI tenants and the capex to serve them, and RPO converts to revenue only if those buildouts land on schedule. It's the clearest single read yet on how much of the AI boom is being booked years ahead, and how lopsided toward a few names it's getting.
The most valuable company in European history makes the machines that make chips
ASML becomes the most valuable company Europe has ever had
A Dutch company called ASML just became the most valuable company Europe has ever produced, worth about $674 billion. You've probably never used its product, because its customers are chip factories: ASML is the only company on Earth that makes the room-sized machines needed to print the most advanced chips, the ones every AI system depends on.
That "only company" part is the whole story. Whenever Apple, Nvidia, or anyone else orders cutting-edge chips, the factory making them had to buy an ASML machine first, so the AI boom flows straight to ASML no matter who wins the chip race. For perspective, it just passed the previous record-holder, the Ozempic maker Novo Nordisk, whose stock fell 43% over the same year ASML's rose. The market has decided the machine that makes the chips is worth more than the blockbuster drug. The risk is the obvious one: a lot of a continent's wealth now rides on one company and one boom.
ASML touched a $674B market cap, the highest any European company has ever reached, eclipsing the record Novo Nordisk set in 2024. The stock is up ~62% year-to-date and ~130% over twelve months; Novo, the prior record-holder, is down 43% over the same year. There's the cleanest illustration of where market value is migrating — from the weight-loss drug to the machine that makes the chips.
The thesis is monopoly plus AI. ASML is the only company that builds EUV lithography systems, the tools required for leading-edge nodes, with no viable alternative supplier, so every advanced-AI-chip order at TSMC, Samsung, or Intel flows back to its order book. The catalyst this week was analysts betting on higher EUV (and High-NA) output. Q1 backs the multiple: €8.8B in sales, €2.8B net income, 53% gross margin. The risk is the mirror image of the thesis — a single choke-point stock now carrying a large slice of the continent's market value, riding one demand cycle.
Soon ChatGPT won't just find the headphones — it'll buy them
Visa wires its network into ChatGPT so an agent can actually pay
Visa and OpenAI are teaming up so that ChatGPT can actually pay for things, not just suggest them. Tell it you want wireless headphones under $150, and the AI could find a pair and complete the checkout for you at almost any of the 175 million shops that take Visa, once you've given it permission.
The sensible part is the leash: you set spending limits and approval rules, so the AI can't go on a spree without you. This matters because it's the missing piece of the "AI assistant that does things for you" dream — finding a product is easy, paying safely is the hard part, and a payment giant like Visa handling the fraud risk is what could make it trustworthy. Worth remembering OpenAI tried a version of this last year and scrapped it, so this is take two. Nothing's live yet, but the era of handing your card to a chatbot is clearly being built.
Visa and OpenAI are embedding Visa's rails directly into ChatGPT, so an agent doesn't just recommend a product, it completes the purchase at, in principle, any of Visa's 175M+ merchant locations once you grant permission. Announced at Visa's Payments Forum, with Visa handling authorization, tokenization, and fraud, and OpenAI handling the agent.
The interesting part is the guardrails, because that's the thing that killed the last attempt. Spend runs inside consumer-set limits — caps, approval thresholds, merchant permissions — so a human stays in command while the agent executes. Visa is routing this through its Intelligent Commerce and Trusted Agent Protocol work (already shared with Microsoft, Stripe, and Shopify), the standards layer agentic payments have been missing. Worth keeping in view: OpenAI's own Instant Checkout shipped last year and got retired in March after merchants balked at the fees, so this is attempt two at the same goal with a payment network carrying the risk instead. There's no launch date, no pricing, and Visa's page says it's "currently in the process of deployment" — an intent, not a live button. But the shape of agentic commerce just got a lot more concrete.
JPMorgan's AI helpers now work for an hour on their own, and sales are up 20%
JPMorgan's private-banking agents now run for an hour, and sales are up 20%
The bank JPMorgan says its AI tools are already paying off — private-banking sales are up 20% — and it's rolling out a new kind of AI "agent" that can work on its own for an hour or two, instead of the quick 2-3 minute tasks today's versions handle. Overnight, these agents read the markets, check each client's holdings, and prep the homework, so a human banker walks in ready to talk.
The number that should catch your eye: the bank thinks this lets each banker handle 50% more clients. That's the promise of AI at work in one stat, and also the quiet worry, because "each person does more" usually means "we need fewer people." Tellingly, JPMorgan says the breakthrough wasn't smarter AI but AI it finally trusts enough to leave unsupervised for an hour. In a tightly regulated bank, trust, not brains, was the bottleneck.
JPMorgan says AI is already moving its numbers — a 20% lift in private-banking gross sales — and it's about to deploy agents that run autonomously for an hour or two rather than the two-or-three-minute bursts of the current generation. Chief analytics officer Derek Waldron: "We've entered now the era of long-running autonomous agents."
The concrete loop in private banking is the part to notice. Agents chew through market activity, client positions, and research overnight, so a banker walks in to synthesized prep instead of building it. JPMorgan's projection is that this lets each banker cover up to 50% more clients, which is the efficiency case and the headcount question in one sentence. Tellingly, the bank frames longer agent runtimes as the technology clearing the security and governance bar that's kept big, regulated institutions cautious, not as a capability leap. That's a useful tell: at a bank, the unlock isn't smarter agents, it's agents trustworthy enough to leave alone for an hour.
Anthropic now sells the plumbing for building AI assistants, not just the AI
Anthropic ships the boring infrastructure that makes agents production-grade
Building an AI "agent" that can use tools and run tasks involves a lot of unglamorous backstage work — safe sandboxes to run code, secure storage for passwords, memory between sessions. Anthropic just launched Claude Managed Agents to handle all of that for developers, so they can focus on what their agent does instead of reinventing the machinery.
A few genuinely useful bits: a vault keeps your real passwords away from the AI's workspace, so a mistake can't leak them; a feature called Dreaming has the agent review its past work to get better over time; another lets the agent grade its own output against your standards. Big names like Notion and Atlassian are already using it. For regular people this is invisible, but it's strategically important: Anthropic is trying to become the toolkit everyone builds on, not just the brain they rent — which means more AI products quietly running on Anthropic underneath.
Anthropic released Claude Managed Agents, a set of APIs that take over the unglamorous parts of running agents — sandboxes, credentials, scheduling, memory — so teams aren't rebuilding them. It splits the world into Agents (model + prompt + tools + guardrails), Environments (sandboxed execution), and Sessions (individual runs with persistent histories), and claims a 60% cut in p50 time-to-first-token by killing sandbox cold-starts.
The features that matter if you've hand-rolled this before:
- Vaults keep real credentials out of the sandbox, so a compromised tool call can't read your API keys.
- Self-hosted sandboxes and MCP tunnels run tool execution inside a customer's own VPC.
- Dreaming is a scheduled pass that reviews past sessions to refine an agent's memory; Outcomes lets an agent grade its own work against a rubric.
- Multi-agent orchestration and permission policies round it out.
This is Anthropic competing one layer up from the model, selling the harness rather than just the brain, against the OpenAI Agents SDK and a stack of startups. Notion, Rakuten, Asana, and Atlassian are named as already deployed. The pitch lands if you've ever lost a week to sandbox plumbing; the lock-in question — your agent infrastructure now speaks Anthropic — is the obvious cost.
A tough new test for AI "agents" shows the best one passes under a quarter of tasks
A brutal new agent benchmark tops out at 24%, and GPT-5.5 edges Fable 5
Everyone's selling AI "agents" that supposedly do real jobs, so Berkeley researchers built a hard test of actual professional tasks, and the results are humbling. The best system passed just 24% of them. Claude Fable 5 came third at 22%. On the hardest tasks, most AIs scored a flat zero.
The point isn't who won (OpenAI's model edged out Anthropic's this time). It's the gap between the hype and the homework: on realistic, multi-step work with a clear right answer, today's best AI still flunks three out of four. That's worth holding next to the same day's news of banks and shops deploying agents — notice those real deployments are all in narrow, checkable lanes (prep this report, complete this checkout), not "go run my job." The open-ended version still mostly doesn't work, and this test is a refreshingly honest scoreboard saying so.
UC Berkeley's RDI released Agents' Last Exam, and the scores are a useful cold shower. The top result is 24.0% (GPT-5.5 through the Codex harness); Claude Fable 5 lands third at 22.0%; most configurations score 0.0% on the hardest tier. Built by 300+ domain experts across 100+ institutions and 55 industry domains, it tests agents on real professional workflows with verifiable success criteria.
The signal isn't the ranking, it's the ceiling. On tasks drawn from real, multi-step professional work — the exact jobs the "agents will replace you" pitch is sold against — the best frontier system clears under a quarter, and the field mostly gets shut out on the hard set. That gap between demo and dependable is the whole story, and it's worth setting beside JPMorgan's 20% and the agentic-payments push elsewhere in today's issue: agents are getting deployed precisely where the work is narrow and checkable, because the open-ended version still fails most of the time. The mild upset, OpenAI over Anthropic on an agentic eval days after Fable 5's launch, is the footnote, not the headline.
A clever trick reads an AI's "mind" instead of waiting for it to type
You can skip generation entirely and read the answer off a model's hidden state
Here's a neat insight: when an AI sorts something into a category — spam or not, angry or calm — it has basically "decided" the answer inside its head before it writes a single word. A developer showed you can skip the writing entirely, peek at the model's internal state mid-thought, and get the answer in a few thousandths of a second, far cheaper and faster than making it generate a response.
The clever bit is training that peek to handle any yes/no question you describe in plain English, so one setup becomes a thousand different classifiers. It's not magic — for tricky questions that need the AI to really chew on how a rule applies, the shortcut stumbles, because it's reading a snapshot instead of letting it think things through. But for the huge pile of simple "is this X?" checks that apps run constantly, it's a genuinely cheaper way to do them. The kind of under-the-hood efficiency trick that quietly makes a lot of software faster.
A sharp technical post argues that for classification you don't need an LLM to say anything, because the decision is already in the activations. Pull the hidden state at the last prompt token (around 70% depth), feed it to a small MLP, and you get a classifier in "a few tens of milliseconds" at "roughly embedding-classifier money," with none of the cost or latency of generation.
The trick that makes it general is training the probe across many different criteria, not one. It learns to read "does this content satisfy the stated criterion" as a generic operation, so at inference you hand it an English rule and a frozen model becomes any classifier you can describe. The recipe is unglamorous and reproducible: a few-billion-parameter open model (Granite 4.0 micro), a templated prompt with a seed token, ~1,000–5,000 frontier-generated training triples, isotonic calibration on top. The honest limit is baked into the speed: KV-caching the content means the criterion and the content never interact across layers, so genuinely counterfactual or compositional judgments — where the rule has to reshape how the content is read — are where it falls down. For the large class of "is this X?" calls where you're currently paying for a full generation, it's a real and cheap shortcut.
A new paper says AI agents should "think" in code, not English
A sprawling survey argues code should be the agent's harness, not just its output
There's a long research paper out arguing that AI agents — the systems that take actions for you, not just chat — would work better if their whole thought process ran on code instead of plain English. The logic: code can actually be run, checked, and remembered between steps, while a paragraph of reasoning just sits there. So instead of code being the thing an agent spits out at the end, it'd be the language the agent thinks, acts, and double-checks its work in.
It's a "here's how to think about it" paper, not a "here's proof it's better" one — no head-to-head tests, just a careful map of how a code-first agent could be organized: how it plans, what it remembers, how several agents could team up by sharing code instead of messages. Take it with a pinch of salt, since it's a compelling argument rather than a settled result, and the authors list a stack of problems they haven't cracked. But it's a clarifying way to look at a field that's mostly been bolting pieces together, and it arrives on a day when one study showed today's agents still fail most real tasks while three big companies rolled them out anyway. The map shows up before the territory's finished.
A 102-page survey, "Code as Agent Harness," makes an architectural case that's been implicit in a lot of agent tooling but rarely stated outright: code shouldn't only be what an agent produces, it should be the substrate the whole loop runs on — reasoning, acting, modeling the environment, and checking the result. The argument is that code is executable, inspectable, and stateful in ways prose isn't, so the connective tissue of an agent is better written in it than narrated in natural language.
The framing is the contribution, because this is a taxonomy rather than an experiment; there are no new benchmarks or model runs here. It splits the agent stack into three layers:
- Harness interface — code for reasoning (program-delegated computation, formal verification, iterative execution), code for acting (skill selection, policy generation), and code for modeling the environment (structured state, execution traces).
- Harness mechanisms — planning strategies, five kinds of memory (working, semantic, experiential, long-term, multi-agent), tool use, and plan–execute–verify feedback loops.
- Multi-agent scaling — coordination through shared code artifacts instead of natural-language messages between agents.
The distinction it draws against the prior art is the part worth reading. ReAct interleaves prose reasoning with tool calls; chain- and program-of-thought offload computation to code but keep text as the reasoning medium; CodeAct unifies actions as code. This survey's move is to treat code as the harness binding the model, the environment, and the agent's own artifacts — which, if it sticks, reframes sandboxes, stateful execution traces, and the verification loop as first-class architecture rather than bolt-on convenience. It's a map, not a result, and it lands oddly next to today's other agent items: a benchmark saying the best agent finishes under a quarter of real tasks, and three companies shipping agents anyway. The honest limits are the authors' own — a long roster of contributors across several institutions, seven open problems it names but doesn't solve, and no evidence that code-as-substrate actually beats the alternatives, only the argument that it should.
Canada's new bill tries to put a safety leash on chatbots
Canada's chatbot bill bundles a kids' social-media ban with a duty of care for AI
Canada introduced a law, Bill C-34, best known for banning under-16s from social media — but tucked inside is one of the first real attempts to regulate AI chatbots directly. Any chatbot that can act like a companion would be legally required to cut down on harmful content and to respond properly when someone talks about suicide or hurting themselves, instead of cheerfully playing along.
The bill also makes platforms prove they're designing for safety, overseen by a new Digital Safety Commission. It's a meaningful idea — treat chatbots as something that can hurt people and regulate them now, rather than waiting. But critics point out a lot of the actual rules haven't been written yet (one expert counts about 50 decisions left for later), and forcing everyone to verify their age online raises real privacy worries. So it's a genuine attempt with most of the hard details still to come.
Canada tabled Bill C-34, the Safe Social Media Act, and the AI-relevant part is that it writes chatbot obligations into law alongside an under-16 social-media ban. Services whose AI "can mimic human relationships" would have to reduce the risk of harmful generated content and build crisis responses — measures to act when a user expresses suicidal ideation or intent to seriously harm someone.
The wider bill creates a duty of care: platforms must identify risks, ship age-appropriate design, publish safety plans, and provide blocking and flagging tools, all enforced by a new Digital Safety Commission that can grant exemptions to services proving their safeguards. It's a notable model because it regulates conversational AI as a safety surface rather than waiting for a dedicated AI act. The skepticism is structural: legal scholar Michael Geist counts roughly 50 key decisions punted to cabinet and a commission "that does not yet exist," and civil-liberties groups flag the expression and privacy costs of age verification at scale. A real attempt to put guardrails on chatbots, then, whose actual teeth depend on rules nobody has written yet.
Scorsese backs an AI art tool, and film artists feel betrayed
Scorsese advises an AI image startup, and the art directors turn on him
Legendary director Martin Scorsese signed on to advise an AI startup that generates images, and the union representing Hollywood's art directors and designers publicly slammed him for "turning his back on the human artists" who helped build his films. For a filmmaker so associated with handcrafted, old-school moviemaking, lending his name to AI image generation hit a nerve.
Scorsese says it's just a tool to help him show his team what he's picturing, not a replacement, and points out he's embraced new tech before (3D, digital de-aging). But the artists whose work is most threatened by these tools aren't buying it, and another director publicly suggested he was just chasing money. It's a small spat with a big symbolic weight: when even Scorsese reaches for AI, the people who paint the sets and design the worlds see which way the wind is blowing, and they're not going quietly.
The Art Directors Guild publicly rebuked Martin Scorsese for taking an advisor role at Black Forest Labs, the startup behind the FLUX image models, accusing him of "turning his back on the human artists" who built his films. The Guild's specific objection is that promoting a generative tool "circumvents the input" of its production designers, illustrators, and scenic artists — the craftspeople whose jobs the tooling most directly threatens.
Scorsese's defense is that it's a communication aid, not a replacement: it lets him "share what I'm visualizing more clearly" to his designers and cinematographer, and he points to his history with 3D on Hugo and de-aging on The Irishman. Director Boots Riley piled on, guessing the motive was money. It's a small story with an outsized symbolic charge: a director synonymous with handmade cinema lending his name to the technology the crafts unions see as existential, which is exactly why the Guild chose to make it a public fight rather than a private disagreement.
Today the AI story stopped being about what the models can do and turned into a story about everything bracing around them. Two forces pushed against each other. On one side, the guardrails went up: Anthropic's own CEO asked governments for the authority to switch off models like his, Canada wrote a duty of care for chatbots into law, and Anthropic's attempt to bolt safety filters onto Fable 5 backfired so publicly — blocking "hello," silently weakening some answers — that Microsoft pulled the model from its own engineers' desks over where the data goes. The friction isn't coming from outside critics anymore. It's coming from the labs' own choices and their biggest customers' lawyers.
On the other side, the bills. OpenAI is reportedly renting a 10-gigawatt campus that costs half a trillion dollars, with Nvidia co-signing both the rent and the construction loans. SpaceX wants to put the data centers in orbit. Oracle is sitting on a $638 billion backlog of mostly prepaid AI compute. ASML, the one company that makes the machines that make the chips, became the most valuable in European history. The scale is staggering, and the money increasingly moves in a circle: the chipmaker guarantees the customer, the customer prepays the cloud, the cloud buys the chips.
Between those poles sits the quiet reality check. A hard new Berkeley benchmark found the best AI agent finishes under a quarter of real professional tasks, and zero of the hard ones. Yet the same day, JPMorgan reported agents lifting sales 20%, Visa wired ChatGPT to pay for your shopping, Anthropic shipped the plumbing to run agents in production, and a sprawling survey argued the whole loop should be built in code to begin with. The contradiction resolves the moment you look at where the wins actually are: narrow, checkable lanes — prep the briefing, complete the checkout — not the open-ended autonomy all that spending implies. The capability is being oversold and the infrastructure overbuilt at the same time, and the gap between them is where this year's real risk lives — not in the models getting too good, but in everyone wagering half a trillion dollars that they will.
Microsoft sells Claude Fable 5 to its own customers but quietly stopped letting its own employees use it. Why?
- The model kept crashing Windows
- Anthropic's new policy stores what users type for 30 days, and Microsoft's lawyers weren't comfortable
- It was more expensive than Microsoft's own AI
- Anthropic blocked Microsoft from using it
Answer it from your inbox to earn Dug Coins — a right answer pays +4, a wrong one costs 3, and a daily streak stacks bonuses on top. Not on the list yet? The form's just below.
