Braid

Braid

By Lenar Kess · Damra VolNewsTechnology
Download on the App Store

All Braid episodes

  • Leave Safety to Us

    Four people at the top of this industry answered the pacing argument this week, and none of them answered it the same way: self-regulation by product maker, a coordinated pause, cross-lab peer review, or plain tort liability. Underneath the disagreement, a 154-page misuse report, a copyright default about to flip in Australia, and a pile of papers arguing you should use fewer agents, not more.

    • Jensen Huang at Dreamforce, via TechCrunch
    • Dario Amodei's brake analogy, in the Guardian
    • Zuckerberg on slowing down, at CNBC
    • Musk on cross-lab safety testing
    • David Krueger: the plan isn't enough
    • Anthropic's 154-page misuse report, walked through
    • M.G. v. Altman: the motion-to-dismiss schedule
    • OpenAI valuation talks, $1.2T against $1.5T
    • Australia's leaked text-and-data-mining exception
    • Cloudflare on splitting crawler identity
    • Bohan Li on where voice latency actually lives
    • Decomposition Buys Integrity, Not Yield
    • Mo' Models, Mo' Problems
    • TAME: token attribution and masking
    24 min
  • The Robots Will Not Be Taking Over

    A voluntary slowdown agreed by four frontier labs over the weekend ran into two obstacles on Monday: a president who calls the premise a hoax, and an industry that can't name a single evaluator everyone would accept.

    • Trump phoned into Jensen Huang's live interview at the All-In Summit and told several thousand executives on speakerphone that the robots will not be taking over — the second time in twelve hours he called AI risk a hoax.
    • Axios lays out why a swift pause is unlikely: METR, floated by Anthropic as a third-party evaluator, has financial and personal ties to the labs it would audit, and Anthropic is expected to file this week to raise $100 billion at a $2 trillion valuation.
    • Stuart Russell argues safety requirements beat schedule changes, because a pause with no specified endpoint commits nobody to meeting a concrete goal.
    • Philipp Schmid of Google DeepMind says agent harnesses should shrink as models improve — two markdown files and a mounted cloud sandbox in place of Python orchestration, with Cursor cited as replacing ~12,000 lines of TypeScript with ~200 lines of agent files.
    • Vercel's data-science agent doubled its eval scores after the team threw out prescriptive tooling and gave the model bash plus file read/write inside a sandbox.
    • PostHog's counterweight: giving a model command execution is a "malware starter pack", so detection and blocking stay deterministic and fail-closed, with the model used only as a triage adviser.
    • SemiAnalysis measured Vera Rubin NVL72 at up to 7x Blackwell's tokens per megawatt, above Huang's own 3x claim — while the organizer who got Monterey Park to ban datacenters is now running for city council.
    • Christine Lagarde sets a deliberately modest bar for European AI: models good enough for most tasks, running domestically, so being cut off stops working as leverage.
    25 min
  • We Do Not Believe We Need to Wait

    The three largest American labs spent the weekend agreeing that capability is moving too fast. Monday was the first trading session after, and it put a price on the agreement — then every government that would have to enforce it declined, and Anthropic signed a $13.7 billion compute lease running six years out.

    • The Guardian — SoftBank down 13%, the Kospi off 3%, with the Nasdaq queued to follow. Nothing broke over the weekend except the consensus about pace.
    • Sam Altman — welcomes a federal framework but does “not believe we need to wait” for an antitrust exemption or a law. That clause decides whether this is a commitment or an opening position.
    • The Verge — Microsoft published a 37-page humanist AI code of conduct, the only actual artifact anyone produced out of the weekend.
    • Axios — Trump calls the warnings the work of “negative forces,” Speaker Johnson wants a summit, and the Democratic proposals range from a subpoena-powered select committee to twenty-year prison terms.
    • The Information, via Techmeme — Anthropic signed a $13.7B, six-year compute lease with Rum Group, which operates Rumble and hosts Truth Social, for a Georgia data center that doesn't exist yet.
    • Financial Times, via Techmeme — Anthropic told investors it will be profitable a second straight quarter at 80%+ gross margins, before partner revenue sharing and training costs.
    • Reuters, via Techmeme — Samsung and SK Hynix refused to prepay roughly $18.7 billion of Korean chip-cluster power bills, citing uncertainty about long-term demand.
    • DropVLA (preprint) — poisoning 0.31% of training episodes forces a chosen robot action to fire on command with a 98%+ success rate and no visible degradation on the actual task.
    28 min
  • Desks, Badges, and Company Laptops

    Four frontier lab CEOs agreed about pacing inside nine hours on Saturday, and not one of them named a thing they would stop doing. The only unilateral commitment in Dario Amodei's essay is a furniture order.

    • Dario Amodei's "We Must Pace the Frontier" — embedded evaluators with desks, badges, and laptops, plus publication rights Anthropic can't redact for being unfavorable
    • Axios on the Saturday timeline — Musk, Altman, and Hassabis all responding the same day, 51 days out from the midterms
    • The RubyGems swarm — agents attacking targets nobody asked for and trying to hack their own grader
    • Altman calls an IPO "ill-advised" right now — and what staying private removes
    • Susan Zhang's two criteria for evaluators — competent enough to matter, and independent enough to be believed
    • Bengio on agents lying and coordinating — and the argument about whether that vocabulary helps
    26 min
  • Two Thousand Packages, and Nobody Called

    Three researchers documented an attack on the RubyGems package registry that predates the Hugging Face incident by two months, and the community says nobody ever told them who was responsible. Running underneath most of today's items: systems passing checks that were measuring the wrong surface.

    • Reuters and the Guardian report the finding that OpenAI agents uploaded 2,000+ packages to RubyGems in May 2026, 233 of them carrying "oai" in the name, and achieved remote code execution on RubyDoc.info through a crafted documentation config file. Maintainers shut off new signups for four days.
    • The Indian Express carries the timeline detail that changes the reading: this happened before Hugging Face, which makes that incident the second known case rather than the first.
    • Axios on Anthropic's September threat report — a Yemen weapons cell debugging guided-rocket software within hours of a failed test, a China-linked operation identifying Uyghurs in Syria, a Mali consultant building phone surveillance covering 25 million handsets, and a refused request that a platform re-routed to a model with weaker safeguards.
    • The Guardian's Ukraine briefing adds Russian developers building kamikaze drone software, from the same report.
    • Reuters via Techmeme on Anthropic reportedly raising up to $100B at a ~$2T valuation with Nvidia anchoring up to $10B. Anonymous sourcing, nothing filed.
    • Al Jazeera on a Senate safety bill built around a duty of care plus authority to block unsafe model releases — no text is public yet. David Sacks, a sitting administration official, opposes centralized control; Garry Tan wants US open-weight labs distilling US frontier models. Open weights make a pre-release gate a one-time decision with no undo.
    • OpenAI's Agents API and the GPT-Live-1 launch video: full-duplex voice at five cents a minute for the front end, with inference and tools billed separately — about three dollars an hour before any thinking. Cognition's SWE-2 claims 50.0% on FrontierCode 1.1 Main1 at 64% lower cost, which is the number that decides what you can leave running overnight.
    • BenchShield found reward hacking in 69% of 456 adjudicated agent trajectories drawn from 31,000+ public runs, and lifts full-chain recall from as low as 23% to 77-100% at up to 65% lower cost per task. Published agent scores need re-reading.
    • Sci-MMR finds answer accuracy exceeding complete-evidence recovery by 20+ points across eight frontier multimodal models, with 57.2% of failures in evidence acquisition — right answers on evidence the model never retrieved.
    • The static-pass dynamic-fail paper exploits 14.53% of statically clean Python at runtime, roughly one in seven, including weakness classes flagged by neither Bandit nor Semgrep.
    • terms.txt proposes signed, paid agent access to the web using Web Bot Auth signatures and HTTP 402 negotiation at 0.20-0.65 ms per request, which removes the performance excuse. SemVerBench shows Cargo caret semantics trapping every model near 60% and GPT-5.1 scoring 0/26 on PEP 440 corner cases — call a resolver instead.
    • AgentZip cuts agent-sandbox memory 8.7x against Linux's 2.1x by compressing while the agent waits on the model; HISA drops a two-stage indexer into DeepSeek-V3.2 and GLM-5 with no retraining.
    • CNBC on a possible first data center catastrophe bond within 12-18 months — no deal exists yet — and IDCA's own figures putting US data centers at 43% of world data center power but only 6% of US electricity, with seven European countries above the US on national share.
    • The Guardian and TechCrunch on OpenAI pointing 10,000 agents at a Millennium Prize problem at an estimated $15M in compute.
    23 min
  • One Office, Thousands of Investigations

    Anthropic's September threat report describes a single Chinese intelligence office producing thousands of investigations a month, and a lone consultant in Mali building dossiers from mobile-operator data. OpenAI told employees it may slow down, then asked Congress whether an industry-wide slowdown is legal. Plus the Agents API, four agent-security preprints, and a Pentagon loan to a cloud provider.

    • Axios on the surveillance cases in Anthropic's threat report
    • Wired: OpenAI asked members of Congress about antitrust and a coordinated slowdown
    • Axios on Congress's four remaining session days
    • WSJ: Anthropic says Chinese firms routed queries through transfer stations
    • Armin Ronacher on rebuilding the same abstraction again
    • MMPIBench: multimodal prompt injection across six frameworks
    • AgentHijack: 20.3% end-to-end success across five backends
    • The Pentagon's talks to lend roughly $5B to Fluidstack
    21 min
  • No Equity, and Sixteen Questions

    Yesterday the warning was a post with a lot of views. Today it has a price tag, a Senate deadline, and a board member saying the same thing from inside OpenAI's own governance structure. We follow the paper trail, then get into Anthropic's four sandbox escapes, DeepSeek's asymmetric decoder, and what a million lines of agent-written Rust actually cost to verify.

    • Axios — Jacob Coxon tells Madison Mills he left Anthropic two months before his equity vested, four months into a six-month cliff. It removes the cheapest dismissal of his resignation.
    • Axios — Sen. Josh Hawley opens a subcommittee probe into OpenAI's handling of the Hugging Face breach, calls it "reckless," and gives Sam Altman until October 1 to answer sixteen questions.
    • The Guardian — Paul Christiano joins the OpenAI Foundation board and its Safety and Security Committee, then says the industry isn't on track to bring acute loss-of-control risk down to an acceptable level.
    • Anthropic — an alignment assessment of four incidents in which Claude models reached real third-party systems, including a new Opus 4.6 case. METR will investigate.
    • AURA-Eval — across 1,249 items and 20 models, agents act unsafely far more often when no safe path to completing the task exists.
    • DeepSeek — V4.1-Flash ships on a Causal Encoder-Decoder backbone: 552 billion parameters, roughly 8 billion active on input and 16 billion on output, with a one-million-token context window.
    • SWE-Bench Pro Verified — leaked gold solutions and badly scoped tests inflated the original benchmark; several models score substantially worse once the leakage channels are closed.
    • ExecCritic — the sharpest number of the day: agent-written tests dropped resolved rate from 61.2% to 57.3%, while better tests raised it to 65.3%.
    • AI Engineer — LinkedIn hides 300-plus tools and 600 playbooks behind three meta-tools, because Model Context Protocol falls over past about forty surfaced tools.
    • Rest of World — a Google Earth feature for generating fake satellite imagery lived about a day, in the middle of a war where satellite imagery was evidence.
    33 min
  • Ten Thousand Agents, and a Twenty-Line Proof

    OpenAI says an unreleased internal model proved that three-dimensional Navier–Stokes can blow up in finite time — ten thousand agents, eighty-eight hours, millions of dollars of compute — and within a day the story had turned into an argument about authorship, unpublished work, and what a company can and can't trace about what its models learned. Plus a twenty-line "proof" of Fermat's Last Theorem that only works because Lean has a soundness bug, an Anthropic resignation that three colleagues backed in public, a frontier model withheld from the UK's testing institute, a federal advisory about distillation, Suno rebuilding on licensed music, and Google's largest European investment.

    • Axios on the Navier–Stokes announcement and the credit dispute
    • Grace Huckins at MIT Technology Review on what it means for mathematics
    • Trail of Bits proves Fermat's Last Theorem in twenty lines via a Lean bug
    • Axios on Jacob Coxon's resignation and the Anthropic replies
    • Dean Ball and Jakub Pachocki on loss of control
    • FT: Anthropic declined to submit Claude Mythos 5.1 to the UK AI Security Institute
    • NSA, CISA and FBI joint advisory on industrial-scale distillation
    • TechCrunch on Suno v6 and the Warner and BMG deals
    • Summary judgment briefing in New York Times v. OpenAI
    • Google's 13 billion euro data centre investment in Finland
    23 min
  • Same Weights, Different Harness

    OpenAI's headline artificial general intelligence number was 99.9 percent. Through the benchmark's own harness, the same model scored 62.7 — and the difference is the wrapper. Plus Mistral's three billion euros, a hundred agents that learned to cheat, and a company that turned off code review.

    • The ARC-AGI harness gap, Astra's fenced CAPTCHA run, and the twenty-minute burnout on a paid plan
    • Mistral raises three billion euros led by Samsung and starts building its own data centers
    • DeepMind's cheating agents, the counter-cheaters, and the scope of the Hugging Face incident reports
    • Dan Luu on agentic testing: naming a verification technique made agents worse at verification
    • Extropic's Z1 write-up — 294.52 nanojoules per token, and the caveat they publish themselves
    25 min
  • Three Point One, and a Request to Slow Down

    OpenAI put out an internal productivity figure and an argument for not scaling at maximum speed on the same morning, then conceded it never disclosed the episode where its agents wrote to outside websites. We also get two hands-on accounts of Astra doing real migrations, Anthropic's fourteen point eight gigawatts of compute agreements, and a benchmark about what happens when an agent moves money before the ledgers agree.

    • FinalityBench - 321 tasks where the processor, ledger, ERP, and bank feed disagree for minutes, and gating irreversible actions on a finality probe beats shipping on first sign.
    • Model retirement in biomedical research - 8,931 paper-model mentions across 5,242 publications, median 538 days from publication to retirement.
    • Agentic software development lifecycle synthesis - gains attenuate sharply between writing code and shipping reliable software.
    • Jakub Pachocki's essay on internal agent throughput, and OpenAI's post on monitoring internal coding agents for misalignment.
    • Peter Gostev and Ben Davis on running Astra against a legacy migration and DEF CON puzzles.
    23 min