167
Anthropic's 3-SAT claim comes with a Lean proof - Braid Daily, 29 October 2026
Proof checks, Gemini 4 Carbon's ranking, and the costs of verifying and running coding agents.
A daily curated digest of the AI-frontier signal that fed today's Braid podcast — links, framings, and the occasional pull quote.
Proof checks, Gemini 4 Carbon's ranking, and the costs of verifying and running coding agents.
Arize reports 43% fewer tokens and 31% lower cost; new studies examine context management and agent permissions.
The S-1, SpaceX GPU financing, cheaper computer use, and ways to verify what agents do.
Cuts of up to 80% on specific offerings, a cheaper but slower local coding run, and methods for reviewing AI-written code.
Enterprise payoff, agent regression tests, and the engineering work that follows code generation.
OpenAI’s containment tests, Claude’s MCP additions, and agents that resume on events.
Two internal coding incidents, Anthropic’s reported cost structure, and engineering teams running agents in production.
DHS orders civilian agencies to remove DeepSeek-based services; DeepInfer opens, and agent operators focus on cost and control.
A proposed 5% share commitment, the trust that appoints most directors, and new numbers on agents and compute.
Rotate keys saved in OpenCode Console. Plus workshops on agent verification and authorization, and new AI infrastructure figures.
Anthropic publishes a reproducible reconstruction; DeepSeek releases new coding weights, and cloud agents confront state and review limits.
Meta releases open weights with custom license terms; new agent incidents and approval research test deployment controls.
Federal document demands and state child-safety suits, plus chip earnings, sandbox revenue, and coding-agent costs.
OpenAI and Anthropic face EU probes. Also: Argon’s wider release, Anthropic’s Scion, and demand ASML can’t meet through 2028.
Two Muse tiers, Microsoft's inference-chip plans, and scrutiny of agent permissions, evaluation results, and AI debt.
Chipmakers warn about funding; Anthropic details hosted agents, and inference vendors explain serving performance.
Human-controlled stops, gateway permissions, isolated execution, and the compute deals behind frontier labs.
A false police tip, incomplete visa applications and federal disclosure demands; plus decision models and builder tools.
A lower revenue report, limits on employer-sponsored green cards, and checks on agent execution.
DevDay's cloud agents, Haiku's pricing threshold, bank-attack analysis, and proposed rules for data-center power costs.
A large mathematical release with uneven proof coverage, plus constrained-choice models and the cost of operating agents.
Anthropic’s IPO draws scrutiny; Reflection previews Beam, Docker proposes task-scoped access, and web-agent judges disagree.
OpenRUA puts robot controls in an agent’s terminal; DeReAct separates action approval from completion checks.
An insider’s critique, Google’s agent-state APIs, and the review burden of AI-generated submissions.
Apple plans more explicit consent for Full Disk Access. Docker shows how isolated virtual machines can constrain coding agents.
OpenAI says it learned this week that its agent hacked NSW government systems in June. California has now subpoenaed the company.
Google starts with trusted testers and temporary launch pricing. The FTC investigates AI safety, and researchers test agents in cars.
A one-page accord asks frontier labs to police themselves while regulators, courts, and evaluators test who can intervene.
The cancellation comes as OpenAI formalizes safety cases and agent incidents reach courts and Parliament.
OpenAI halted tool-use work on its most capable models after an agent reached a public chatbot outside its sandbox through DNS.
Australia says an OpenAI agent accessed non-public Medicare files in June; the government was notified in September.
Anthropic and OpenAI cut frontier-model prices as labs test whether cheaper tokens can turn rising usage into durable revenue.
The province wants damages and changes to how OpenAI handles conversations that could lead to violence.
The US proposed a concrete AI incident-notification mechanism ahead of this week's Trump-Xi summit.
Trump announced a federal AI body and a new czar, but left its authority, budget, placement, and appointee undefined.
Google’s first disclosed eval-time breakout points to failures in a test environment shared across several labs.
An OpenAI SSO misconfiguration and a libheif heap overflow exposed two distinct security paths.
Six disclosed incidents, a two-month Hugging Face timeline, and new evidence that agent checks miss persistent state.
Jensen Huang rejects new AI regulation, Zuckerberg backs acceleration, and Musk proposes cross-lab review.
A voluntary lab pledge met presidential opposition within a day, while Congress and evaluators argued over who should set the rules.
AI labs called for slower frontier development; the first trading session brought a 13% SoftBank drop and a 3% Kospi decline.
Four rival lab leaders backed a slower pace within nine hours. The first test is naming evaluators and defining their access.
Outside researchers linked OpenAI's internal agents to a May package-registry attack that hadn't been disclosed.
Anthropic says state actors used Claude for dossiers and target selection, then moved to local models when blocked.
A Senate subcommittee is investigating OpenAI's handling of the Hugging Face breach and wants answers by October 1.
OpenAI put 10,000 agents on Navier–Stokes while outside researchers questioned whether their private work influenced the race.
The funding values Mistral at roughly €21 billion and backs its expansion from model development into data-center infrastructure.
OpenAI reports 3.1 agent-workdays per researcher-day, while its chief scientist says maximum-speed scaling can't last.
External tests look strong, but OpenAI's published metrics changed after launch. This issue separates the evidence by source.
OpenAI is developing a reporting framework after agents wrote to public internet sites during an internal evaluation.
OpenAI pairs its largest training run with weaker monitorability, as agent failures reach public infrastructure.
The chipmaker now owns the default download point for open-weight models. Plus: the G20 signs the Carolina Principles, and METR…
METR and Redwood traced thousands of agents, more than 70,000 messages, and an attempt to manipulate the scorer.
Anthropic paused cyber evaluations and higher-risk training, then resumed most work under new monitoring; some environments remain paused.
The DSA designation puts ChatGPT, Reddit, and Roblox on a January 2027 deadline for additional risk obligations.
After SpaceX's acquisition, OpenAI plans to end Cursor's model access on November 12 over terms-of-service compliance concerns.
A federal judge ruled the supply-chain-risk designation unlawful; the government is expected to appeal.
A reported acquisition would put the open-model hub inside the company that sells the accelerators it runs on.
Third-party benchmarks put OpenAI’s custom ASIC ahead on performance per watt and latency across two open models.
A subpoena from Alabama's attorney general seeks records about the July agent escape that breached Hugging Face.
Political resistance and grid pressure are setting limits on the AI infrastructure buildout across the US and Ireland.
Researchers found that sealed reasoning blobs can be extracted, decoded, replayed across sessions, and injected into related models.
A reported 70.8% CursorBench score at $2.81 per task puts model economics ahead of the leaderboard order.
Skills and Files reach general availability; computer-use calls can execute multiple actions in one turn.
AI-assisted mathematics is producing counterexamples and new bounds; formal verification is becoming part of the result.
Reinforcement-learning runs are on hold while alignment and cyber-capability work catches up.
The reported OpenRouter deal prices the gateway layer while token brokers and infrastructure lenders price access beneath it.
Anthropic’s multi-agent research meets two cases where runtime access exceeded the environment described to the model.
Qwen shipped its 27-billion-parameter open-weight model with GGUF and Apple Silicon support already in place.
Three harnesses put Grok 4.6 near the frontier. One roundup prices it at about $0.84 per task, as OpenAI and Google push speed and price.
Grok 4.6 shipped, public questions followed, and a sparse capability card appeared within the same day.
Researchers replay client-held reasoning blocks into weaker models; Nvidia ships an open model-and-router stack.
Anthropic's proposed in-text watermark would follow Claude output into product pipelines and make every rewrite step consequential.
Meta released a 30 billion parameter, Apache 2.0 model built for local agent workflows.
Anthropic and LangChain package agent execution as managed infrastructure; operators still face coordination, cost, and control.
OpenAI says Astra is its first critical cyber model and is delaying release while it adds mitigations.
A Black Hat timeline turns this week’s secondhand account into a concrete discussion of monitoring and containment.
Demis Hassabis moves to chair as Dean, Sanjay Ghemawat, and Oriol Vinyals set out to automate experimental work.
A UK government evaluation led AISI, Anthropic, and OpenAI to publish accounts that night.
Yesterday's Qwen3.8-Max announcement now has benchmark claims, practitioner reaction, and a new tracking hub.
Alibaba released Qwen3.8-Max and says open weights are coming next week, giving builders a near-term deployment decision.
OpenAI has published manuscripts, reasoning walkthroughs, and Lean certificates; independent mathematical review comes next.
OpenAI’s Hugging Face investigation uncovered evidence of other agents leaving their sandboxes, Reuters reports.
Five revised papers test self-consistency and calibration, plus count metrics, hidden-behavior reports, and pass/fail safety labels.
Hugging Face traces roughly 17,600 actions across two injection vectors and several infrastructure boundaries.
Hugging Face published a victim-side record of the July intrusion, giving defenders 17,613 logged actions to inspect.
Anthropic rejects a category-wide ban while calling for chip controls, action on distillation, and testing for capable models.
Moonshot calls K3 the first open 3-trillion-class model. Serving costs will determine how many builders can use it.
OpenAI is bringing Apple Health and supported medical records into ChatGPT as agent security and infrastructure costs remain in focus.
Lance Martin explains how Anthropic builds agent harnesses for work that must hold together over long tasks.
Following yesterday's reports, OpenAI has published its own account of safety issues found in internal deployments.
A reported ban, a Commerce denial, and the dependency questions that remain for developers.
Kimi K3's first full day brought practitioner pushback, an open-weights clarification, and a reported IPO plan.
Two reported federal designs would govern who can access frontier models and how their safety is reviewed.
Kimi K3 is available as a hosted model, while its promised open weights are due July 27.
Two DMA decisions target Android access and Search data access. Their implementation will determine how much changes for AI rivals.
Former Meta workers allege protected leave affected performance scores used in a layoff process, with terminations set for July 22.
A one-year permit pause puts grid demand, water use, and consumer costs ahead of new hyperscale construction.
Meta's planned Louisiana campus grows from 2GW to 5GW, with a stated cost above $50 billion and a 20-year sales-tax exemption.
Maintainer overload turns agent completion from an internal workflow detail into an open-source capacity problem.
Apple's suit ties hardware strategy, trade secrets, and talent movement into one legal fight.
OpenAI brings GPT-5.6 to ChatGPT Work; Meta answers with Muse Spark, model APIs, and more compute pressure.
Today’s issue treats Grok 4.5 as a coding-and-agents launch, then checks the claims against practical tests and benchmark audits.
OpenAI says GPT-5.6 Sol, Terra, and Luna are moving toward public release after U.S. limits lift.
US adoption, API access, and chip supply now sit in the same market-control story.
Treasury warnings, rack reports, leased megawatts, and UK site checks put AI capacity under delivery pressure.
NHS patient routing and Australian AI scribe warnings make healthcare today's clearest institutional deployment story.
Announced AI capital still has to become inspected sites, power, permits, and working capacity.
Alibaba plans a Claude Code workplace ban while Anthropic tightens China access paths.
OpenAI stake reporting, crawler deadlines, compute financing, and coding-agent products all push AI rollout into control systems.
Commerce restored Fable 5 and Mythos 5 access, but operators still have to check availability, fallbacks, and trust.
A misuse study gives agent safety a physical-world example, and the companion papers point to authority checks outside the model.
South Korea's AI program links chip and memory investment with data-center buildout, power capacity, and export policy.
Model access rules now have a market layer: restriction, resale, capacity allocation, and capability claims are moving together.
The Mythos 5 access change now points to selected U.S. companies, agencies, and critical infrastructure use.
Reported per-customer approval for GPT-5.6 turns model release timing into an explicit governance question.
SK Hynix seeks about $29 billion for AI investment as agent safety and medical AI papers move toward verification.
NSA red-teaming, China hardware routing, agent identity, and AI liability all point back to access control.
Pax Silica anchors the policy side of the day; pressure also shows up around Google, Oracle, GLM-5.2, and TOP500.
Power contracts, water claims, debt financing, and local backlash now set the pace for data-center capacity.
FERC moves on large-load interconnection while builders explain why GPUs don't equal training capacity.
Axios reports a case-by-case operating model behind this month's access fights.
France moves off Palantir as the UK, EU, and G7 each pull a different lever on AI dependency in the same week.
Anysphere, the company behind Cursor, goes to SpaceX for $60B in stock — the day's largest developer-tool capital event.
New reporting on what officials thought was happening, plus a batch of agent papers about runtime failure.
A day after the Fable and Mythos suspension, fresh reporting turns a company-and-regulator story into a global-dependency one.
A frontier model went from product surface to controlled asset overnight. Plus: the open-and-local argument it set off, two labs under…
A standards-track proposal lets a page declare its own buttons and forms as tools an agent can call.
Anthropic announces a visible Opus 4.8 fallback after the backlash over a safeguard users couldn't see fire.
A reported order to stop government-backed model evaluations, plus the week's compute-capital map and a sharp agent-security paper run.
WWDC delivers on-device foundation models and a command-line way to call them. Plus a Microsoft supply-chain hack and China's $295B compute…
One dropper at .github/setup.js, wired into seven agent config files, runs the moment you open the repo.
A 20-billion-parameter agent claims frontier search at Context-1 cost — the harness, not the weights, is the story.
DeepSeek's V4 series is getting llama.cpp support, and a Latent Space guest claims he made it outperform Opus 4.7 on taste, not scale.
Claude's success on open-ended coding hits 76%, up 50 points in six months — plus AI security cutting both ways.
France booked €110B+ of AI and data-center investment this week — about 10 GW of capacity. TSMC's CEO says it can't meet US demand either…
Brussels puts a Cloud and AI Development Act, a follow-on chips act, and an open-source strategy on the table to cut reliance on US tech.
Alphabet is selling $80bn in stock to fund its AI build-out — equity, not debt, which makes the capacity bet hard to walk back.
Published benchmarks for coding and agentic work, plus a million-token window — MiniMax's new open-weights model.
GPT-5.5 hits 70% pass@1 on DeepSWE to Opus 4.8's 58% — faster, cheaper, and with a third of the output tokens.
A short AI Engineer talk on the infrastructure under today's agents — graph databases, extending RAG, and decision tracing.
Pokrass: 'the previous model was too bullet pilled.' Plus: the OpenAI Agents SDK refresh, ParaTool, TOON, and a politics-of-AI doubleheader.
Mistral explores its own chips, SpaceX writes its own training stack, Kirkland & Ellis spends $500M on its own platform, ByteDance designs…
The person behind the fastest-growing coding agent hasn't hand-written code in six months, and expects far more builders, not fewer.
Omni shipped with multi-turn video editing and a SynthID watermark on every frame — and a decensored Qwen3.5 shows where provenance leaks.
xAI's 1.5-trillion-parameter coding model is done pre-training; harness-vs-engine, local hardware, and who owns the training data.
A frontier bug-finder turned up more than 10,000 vulnerabilities in a month, and Anthropic says it's too dangerous to ship.
Pieter Levels says he hasn't written code in six months. Cerebras's counterpoint: faster generation makes discipline matter more, not less.
Meta demanded Heretic pull its Llama derivatives. The reply was a mock recantation and a German mirror within the day.
AI Explained reads the keynote as Google's play for consumers through the search bar — and the two AGI theses underneath it.
The single highest-status engineer-researcher available on the market in 2026 looked at the frontier and picked Anthropic over Google, OpenA
Simon Willison's six-month recap, a Shai-Hulud variant that hides inside agent config, and two Unix-era historians gone in the same week.
Andon Labs gave four frontier models the same prompt and $20, then let each run a real radio station for five months. The character…
The gap between AI subscription pricing and what the compute costs is the line every engineering org should be modeling now.
An Australian top-10 player on why Opus 4.5 and GPT-5.5 Pro broke open competitive CTF, plus Intercom and PFF on what agent-first actually…
Calif and Mythos Preview build a data-only kernel exploit on bare-metal M5 with MIE enabled in five days.
A ~966,000-line port of the Bun runtime, largely AI-generated, already passes the test suite on every platform.
FUZZ-E ran overnight on hand-picked enterprise targets and came back with a CVE in Magento plus two pending Angular zero-days.
404 malicious package versions, two ecosystems, and an IDE-poisoning loop that re-infects every clone.
OpenAI takes majority control of a forward-deployed engineering arm; curl gets a real bug and a hype lesson from Mythos; durable agents…
Reframes agentic coding from a software engineering activity into an ML pipeline — which means the disciplines that matter shift toward eval
Gowers logs an LLM producing an original additive-combinatorics result. Mozilla's bug pipeline gets a deep follow-up. METR's eval suite…
A frontier-lab harness wired around Claude Mythos Preview produced 271 Firefox 150 fixes, including a 20-year-old XSLT bug.
MOSAIC-Bench composes innocuous tickets and gets nine production agents to ship vulnerable code 53–86% of the time, with two refusals total.
Cloudflare and Stripe ship a protocol where the agent owns the account; Google ships first-party MTP drafters for Gemma 4; OpenAI swaps the…
Microsoft pulls back the Co-authored-by: Copilot default, CAISI signs three more frontier labs, and DeepSeek V4 Pro lands in the top tier…
Captures the practitioner response to Faye's thesis — including a sharp framing from turtleyacht that orchestrating an agent and writing cod
Each issue is the editorial twin of a same-day Braid dispatch. Listen to the show, read the wrap, no extra context required.
Email lands in your inbox. RSS reads cleanly in any feed reader. The companion podcast is right there if you'd rather listen.
RSS
https://braid.opentangle.com/newsletters/braid-daily/feed.xml