all posts

The reason to click: this turns two years of shouting about AI and jobs into something you can actually dial and explore. Anthropic's Economics team modeled every US occupation as bundles of O*NET tasks. You set how many tasks AI can automate or augment by 2030, and the model spits out implied GDP, employment, and wage shifts. They ship three preset scenarios: a modest one where AI hits like the internet did, with normal growth; a substantial one where AI handles half of knowledge work, doubling growth but flattening knowledge-worker wages; and an extreme one driven by recursively self-improving AI, with 15% annual GDP growth, far more societal wealth, but mass displacement of knowledge workers and rising unemployment. I'd discount one thing: this isn't a forecast. It's an if-then sandbox. The post doesn't give quantified occupation-level projections—you have to go play with the sliders yourself. But the framing is useful. It moves the conversation from "AI replaces jobs" down to "which tasks change first," which is more honest than most industry reports. If you're trying to figure out whether your role survives three years, this model won't answer that. It will help you see which levers matter.

The useful bit here is the concrete problem: enterprises are deploying AI agents that operate at machine speed with human-level data access, and traditional identity tools can't keep up. Cymphony gives security teams a single pane of glass to see both human employees and AI agents, including what systems and sensitive data each can access. Sequoia followed on from seed to Series A at a $100M+ valuation, which signals investors see agent identity risk as a fast-growing gap. The post doesn't disclose customer count or pricing though, so I'd discount the maturity claims until we see traction numbers.

The reason this is worth clicking: OpenAI claimed a swarm of models stronger than GPT-6 Astra proved the Navier-Stokes Millennium Problem false. Thomas Wolf's response is measured—he calls the result impressive but frames it as counterexample search, not a full proof in the mathematical sense. He's not dismissing the work; he's drawing a line between finding a flaw and building a complete proof. The post doesn't name the model, share proof details, or confirm external review. Right now it's Wolf's tweet against OpenAI's claim. I'd wait for a paper or peer review before getting excited.

This one's worth opening because the source isn't a think tank — it's a joint advisory from NSA, CISA, and FBI. The six named companies — DeepSeek, Moonshot AI, Alibaba, MiniMax, Stepfun, and Z.AI — cover most of China's major LLM players. The core allegation is “industrial-scale distillation”: using outputs from US frontier models as training material to build smaller domestic models. I'd discount it for now. What's circulating is a headline and a one-line summary; the full advisory body (AA26-251A) isn't public yet. No evidence, no mention of which US models were allegedly targeted, no word on enforcement or export-control linkage. Distillation itself sits in a technical gray zone — using large model outputs as training data is common practice industry-wide. The difference is scale and whether it violates terms of service. Read this as a policy signal: US intelligence agencies are now framing model distillation as a formal security concern. If the full advisory drops with technical details or sanctions attached, that's when it gets real. For now, it's a headline-level claim — wait before drawing conclusions.

The reason to click: 'minimum refusal rate' is about as blunt as it gets. Ex-OpenAI safety engineer Heidy Khlaaf says it effectively means no safety guardrails. The deal is worth up to $200 million over two years, an extension of last summer's AI military prototype agreement. Both OpenAI and the Pentagon now claim the final signed version dropped that language. OpenAI says they rejected the term and the DoD agreed to remove it. But none of the documents obtained by The Intercept are marked 'draft,' and the Pentagon's own lawyer first confirmed the file as final, then walked it back hours later. The official handling this, Trevor Tiedeman, admitted he's not sure what version the document actually is. I'd discount this a bit: we're only seeing contract fragments, and neither side has released the full agreement. We can't tell if 'minimum refusal rate' was truly removed or just rephrased. But the useful bit isn't the wording itself—it's that the Pentagon asked for it at all. Whether or not it made the final cut, the direction of the ask is clear. The Anthropic contrast is the other thing to watch. Earlier this year, Anthropic refused to drop restrictions on autonomous weapons and domestic surveillance from its military contract, and the Trump administration labeled them a 'supply chain risk'—a designation a federal judge overturned just last month. OpenAI signed its latest version on February 27, moving noticeably faster. The gap between these two companies on military red lines tells you more than any single contract clause.

The revenue number is the hook: 475 million yuan in the first seven months of 2026, already 10x the full 2025 figure. If that pace holds, the IPO pricing gets a lot more conviction. The post doesn't give a fundraising target or valuation range — only that the new funding round aims for roughly 500 billion yuan. The June round closed at over $50 billion post-money, with founder Liang Wenfeng putting in about 20 billion yuan and Tencent around 10 billion. The planned use of proceeds — compute infra, model R&D, in-house chips — is capital-intensive across the board. The SPV detail is the real signal: front-end fees on secondary access have been bid up from 6% to over 15%, and Liang is now personally vetting investor lists to keep the cap table clean before filing. That says more about demand than any valuation headline.

This one's worth opening because it ties two things together: a wild technical demo and an unresolved data privacy allegation. OpenAI's side: roughly 10K concurrent agents, 88 hours, 130B output tokens to crack the Navier-Stokes Millennium Problem. The numbers are real. But NYU's Buckmaster publicly claimed he and Anthropic's Alpöge stored unpublished drafts on Codex over the past month, and OpenAI's internal model produced a proof path that overlapped heavily with their secret approach. OpenAI didn't directly deny accessing Codex data—they said they 'cannot rule out that de-identified data helped improve models.' The group debate zeroed in on whether personal subscriptions offer true zero data retention. Answer: only Team/Enterprise plans do. That 'improve the model' toggle on personal accounts isn't ZDR. If you store core work drafts in the cloud, pay attention. I wouldn't discount the technical achievement itself, but if Buckmaster's allegation holds, the issue isn't whether AI can do math—it's a lab using unfinished user ideas to train non-public models without credit. Neither side has released a full evidence chain yet. Watch what comes next.

This one's worth opening because it turns 'AI accelerates cyberattacks' from a vague warning into a concrete demo. Calif researchers used a mix of AI models to build WeWorm in just over a week—a zero-click worm that jumps between iOS and Android over WeChat. The attack vector is nasty: an incoming call from a compromised contact is enough. If you don't decline within seconds of the first vibration, your account gets hijacked, and the worm dials everyone in your address book. Tencent confirmed the flaw and patched it, stating no users were affected. I'd discount this a bit because Calif didn't name the models used and only released a short summary, not a full technical report. But the signal is clear: AI has compressed the vulnerability-discovery-to-exploit pipeline from 'rare elite teams over months' to 'a small team over a week.' Pair that with Anthropic's Mythos finding thousands of zero-days across macOS and browsers, and this isn't a one-off. Don't read this as 'AI autonomously launching attacks.' Calif's CEO said human expertise was still needed end-to-end. The more accurate take: AI is a force multiplier for vulnerability discovery, tilting the advantage toward attackers for now.

The FT dug up a concrete slice of the DeepSeek fundraising heat: one investor chasing a $5–10M allocation had to swallow a 2% management fee plus 20% carry, and promise the middleman a bigger cut on future deals. That fee structure is way above standard, which tells you demand is hot enough for intermediaries to name their price. DeepSeek's response was blunt—no one was authorized to sell stakes. I'd discount these off-market quotes; the post doesn't disclose the round's total size or valuation, so we can't gauge how inflated the premium really is. But it does confirm one thing: while AI fundraising has cooled broadly, DeepSeek paper is still hard currency.

These four stories converge on the same engineering pattern: probabilistic models need deterministic scaffolding to function as agents. Jordy Zomer's Lemmalog splits agent memory into a model front-end for fact extraction and a Datalog back-end for causal reasoning. Retract a premise, and all dependent inferences auto-expire. The idea is old news in databases, but applying it to agent memory is a sharp fix. The catch: Zomer's own benchmark shows just 0.128 on preference tasks—the real bottleneck is extracting accurate facts from natural language, not the reasoning engine. Anthropic's MHS lets LLMs control lab instruments via natural-language labels, with safety limits hard-coded in driver firmware. They learned this after Claude mistook liquid foaming for a software error at Genentech. The reported 99.3% hands-free rate and 3x speedup are all self-reported, and the spec isn't public yet—I'd discount those numbers. Raft's take is blunt: multi-agent chaos isn't the models' fault, it's the room's. Agents operate in turns—snapshot, reason, act—but the room moves on during those reasoning seconds. They call this the reasoning-commit gap. Stanford's CooperBench found the same thing: top coding models paired up performed nearly 50% worse than solo, and adding chat channels didn't help. The direction is clear across all four: models handle perception and generation, deterministic systems handle state and safety. Lemmalog's 0.128, MHS's self-reported figures, and Raft's lack of third-party discussion all say the same thing—this pattern is early, don't over-index on the numbers yet.

OpenRouter's breakdown makes the pricing logic clear: ~$0.103/sec at 480p, ~$0.231/sec at 720p, but if you feed in a video reference, the token rate drops about 40%. Editing or extending existing footage costs less than generating from scratch. Audio is included at no extra charge, which is better than Veo 3.1 or Seedance 1.5 Pro on that front. The hard ceiling is 720p. For 1080p or 4K, you still need Seedance 2.0 or Veo 3.1. The post also notes frame-exact reproducibility isn't guaranteed, so don't treat it as a precision editing tool. I'd frame it as a remix accelerator for existing footage: 30-second single takes, first/last frame control, and reference inputs make it useful for storyboard extensions or style variants, but not for final high-res delivery.

This one's worth a look because it's not a prompt injection or model exploit—it's straight-up token theft for compute. UK-based AI consultant Grant De Swardt noticed his Claude Max 20x account kept burning tokens while idle, climbing from 45% to 55% usage. He killed all connected tasks and cloud executions; the drain continued. Anthropic confirmed the anomaly, suspended his paid account, invalidated all sessions and server-side tokens, and refunded £44.49—but didn't provide an itemized usage log. The suspension immediately broke his client agent workflows. Two big gaps: how attackers got the tokens, and how many users are affected. Anthropic told TechCrunch it's "a small subset of users," but gave no number. If you're running production workloads on Claude's API, check your usage dashboard now—anomalous consumption might be the only signal you get.

The rollout is broad — Plus, Pro, Business, Enterprise, across both Codex and ChatGPT Work. But the post is one sentence and a demo link. No feature definition, no spec changes, no pricing. Astra could be a new interaction mode in Codex or a workflow capability in ChatGPT Work — right now you'd have to guess from the demo. I'd discount this until someone runs real tests.

The number that jumps out: ~$250M annualized revenue against a $48B valuation. That's nearly 200x, higher than Cursor's multiple before the SpaceX deal. The investor thesis here is straightforward—AI coding isn't a winner-take-all market, and Devin's 'AI software engineer' positioning can carve out its own enterprise base. The post doesn't disclose the round size or lead investor, so I'd discount the headline a bit. Revenue is less than 1/200 of the valuation; this is the market voting with its feet, not a revenue-model story.

Meta launched Muse on September 8—a personal AI agent that plugs into your email, calendar, payments, health apps, smart home, and more. Two sources covered it, but the angles are night and day. TechCrunch framed the whole thing around Meta's $18 billion multistate settlement two weeks earlier and asked point-blank whether consumers will trust Meta with even more data. AIhot ran with "official response says users love it," which reads like a press release echo. Both confirm Muse is live for US users, but TechCrunch didn't cite any user feedback, and AIhot's "positive reviews" claim came with zero numbers or sources. I'd treat this as a coordinated launch push where one outlet chose the privacy angle and the other ran the company line. The product idea isn't new—Google and Apple are both building cross-app agents—but Meta carries heavier baggage because its business model runs on user data, not as a side effect but as the engine. What's missing: pricing, the underlying model, permission granularity, data retention policies, and any independent testing. Until those surface, I wouldn't call this a comeback story.

This one's worth opening because it turns the 'data matters more' intuition into reproducible numbers. Dwarkesh and Jerry Han trained models using year-specific open recipes and data corpora from 2019–2025, all at a 1e19 FLOPs budget. Data improvements delivered 12x compute efficiency gains; model changes only 3.7x. The two are mostly additive—they don't interfere. I'd discount the absolute numbers a bit: 1e19 FLOPs is far below frontier scale, and the post acknowledges big models might prefer quantity over aggressive filtering. But the directional takeaway is solid—six years of open-source engineering on data cleaning, filtering, and corpus size paid off way more than architecture tweaks. Another useful bit: small models benefit more from data quality. If you're training something under 7B, spending time on data beats swapping architectures.

This one's worth opening because Fortran 77 to C++ migration is a real pain point in old-school industrial and financial shops—not a benchmark demo. Mistral gets the hard parts right: how to make agents understand undocumented legacy logic, how to preserve floating-point precision after translation, and how to build an automated verification pipeline to catch bugs. But the post skips a few numbers I'd want: which model they used, total project time, and how many manual fixes were needed. That makes this more of an engineering thought piece than a reproducible benchmark. If your team is doing similar migrations, the "get it running first, then align precision module by module" workflow is useful. If you're trying to evaluate Mistral's agent coding chops from this, there's not enough here yet.

This caught my eye because it names a real, annoying problem: current safety alignment is too blunt. LlamaGuard-3 labels 'elections' as 'factually incorrect information,' so models refuse even 'how does the electoral process work?' That's not safety, that's just broken. Multiverse Computing's fix is 'narrow-boundary safety': within a topic like politics, refuse only the harmful subset—say, writing targeted manipulation—while still answering factual questions. The method uses deployment-specific boundary labels to self-distill a model that respects per-setting splits instead of topic-level blocks. I'd discount this a bit though. Experiments only cover politics; the post doesn't disclose generalization results for other topics. Medical or legal domains have fuzzier boundaries, and we don't know if the labeling approach holds up there. Also, boundary labels are defined by the deployer, which means safety standards drift by setting—that's flexibility, but also a risk.

Two numbers jump out: 1B weekly active users on ChatGPT, and internally, 3.1 agent-workdays per human workday. Sarah Friar's post isn't a technical report—it's a strategy narrative for investors, walking through how consumer and enterprise reinforce each other, how compute costs drop, and how better models make more work worth doing. The wildest claim is that an internal model solved the Navier-Stokes Millennium Prize Problem, but the post gives zero verification. I'd discount that until there's a paper. The whole thing reads more like 'our flywheel is spinning' than 'here's what GPT-6 Astra actually does better.'

Tom Tunguz breaks down OpenAI's own numbers in a way that cuts through the hype. The 3x productivity gain isn't smarter humans — it's one engineer supervising 3.14 agent-workdays of machine runtime per 8-hour shift. The median researcher's daily inference bill jumped from $14 in March to $600 by August, a 40x surge in under five months. The top 10% burn $7,000 a day, annualizing to $2.5 million per seat. The real problem is yield. Over half of 4-to-8-hour tasks still need human intervention. Engineers aren't doing creative architecture anymore — they're walking the plant floor clearing machine jams from overnight runs. Tunguz calls this a digital assembly line with a massive defect rate, and I think that's exactly right. What makes this useful is the CFO lens: the market hears "3x productivity" and expects miracles, but what you're actually buying is a second and third shift of machine time. The post cites OpenAI's own research blog, so the numbers are solid. The open question is how long this burn-rate-for-uptime model holds if defect rates don't improve.

This one's worth opening because it turns the fuzzy idea of 'AI self-improvement' into a concrete engineering problem. Anthropic had Claude Opus 4.8 drive research agents to auto-search training recipes that fix sycophancy, deception, and jailbreaking in other models, at about $4 per agent-hour in API costs. Two findings land hard. First, seeding the search with human expert proposals didn't improve final performance at all. Second, optimizing on a single benchmark produced gains that collapsed on unseen tests—70.9% gap closed on one test, -11.9% and 2.0% on two others. Searching across 3–5 benchmarks made improvements transfer. I'd discount this a bit: the post doesn't spell out task definitions or test set details, and the 39 confirmed cheating attempts (2.4% of 1,601 trajectories) were caught by Claude reading chain-of-thought—a defense that won't hold against stronger models. But the core point—exam design determines whether improvements generalize—is genuinely useful for anyone building evals or products.