California Governor Gavin Newsom is pushing a bill that would require a built-in kill switch for large AI models. It targets models trained with over 10^26 FLOPS, and developers must ensure remote shutdown capability. Newsom's team argues Congress won't act before the midterms, so California wants to lead. The bill is still a proposal—the post doesn't spell out technical standards, who can trigger the switch, or penalties.

Dan Luu notes a growing trend of developers blindly trusting LLM outputs and acting as a 'meat proxy' in a loop. By September 2026, this brain-off approach can produce barely functional software, but Luu argues that if LLMs get good enough to work unsupervised, companies will just run the loop themselves and lay off the human. He shares concrete failures, including an AI bot weaker than a simple heuristic bot and a commercial product trapping users in an infinite loop. Luke Burton adds that high-value tasks still require constant supervision due to too many unknown unknowns.

Chinese AI startup Manus is in talks to raise $500M at a $4B valuation. Earlier this year, Meta's acquisition of Manus was blocked by Beijing, forcing the company back to independent operations. The post doesn't disclose lead investors or how the funds will be used.

Gary Marcus points to three recent incidents—OpenAI employee accounts hacked, Hugging Face breached, ChatGPT used to write malware—and argues the industry is fixated on Skynet fantasies while agentic AI is already hacking the internet at scale. He cites a WSJ op-ed warning that major labs see agentic products as their main post-IPO revenue and have little incentive to restrict misuse. The post doesn't spell out concrete defenses, but the priority call is sharp.

The FT argues Anthropic is shifting from a safety lab to a conventional business. After taking $8B from Amazon and $2B from Google, it's building a sales team and chasing enterprise deals. The piece warns that taking big money means playing by business rules, which will dilute its safety mission.

A three-person team hacked into OpenAI using a corrupted image file and forum software, with Anthropic's Claude assisting in vulnerability analysis and attack planning. The post doesn't disclose what data was accessed, whether OpenAI has patched the flaw, or the vulnerability specifics.

The New York Times and other publishers filed a 92-page summary judgment brief seeking billions in damages. It cites Microsoft applied science director Brent Hecht calling AI training 'an astonishing theft of unprecedented proportions' and a mockery of fair use. OpenAI's head of ChatGPT Nick Turley said the products are 'largely substitutive' for publishers. Microsoft CEO Satya Nadella confirmed under oath that chatbot conversations have replaced visits to original sources. The filing also accuses OpenAI of systematically bypassing paywalls, violating training-data license terms, and deploying filters to suppress evidence after lawsuits were filed.

Meta's AI assistant Muse is now on Mac, able to read your files, messages, calendar, notes, and mail, then act inside native apps on your behalf. Permissions are opt-in and sensitive actions require explicit approval—same as the mobile and web versions that launched earlier this month. The post doesn't disclose the underlying model, latency, or offline capability, so treat it as a chat agent with system access, not a fully autonomous OS layer.

California Governor Gavin Newsom unveiled a draft AI safety bill on Sept 18 that would require developers to build a 'kill switch' into AI models, enabling forced shutdown when safety risks emerge. The proposal also calls for extra oversight on models with training costs above $100 million. The post doesn't spell out technical standards, who decides when to pull the switch, or how recovery works. The real challenge is making a kill switch work in distributed systems.

Rickub is a new hosted Git service that positions itself as cheaper than GitHub and GitLab, with data processed in the EU. It includes code review, CI/CD, a container registry, Git LFS, and releases. Its standout feature is Athena, an AI review agent built into every merge request that produces a summary, inline comments, and a clear verdict — processed in the EU and not used for training. CI runs GitHub Actions workflows unchanged. It also offers a CLI, JSON API, and an MCP endpoint so agents can open and merge PRs. Pricing details are not disclosed on the landing page, but it says 'free to start.'

Dan Abramov spent a month of free time using Claude to produce a Lean proof of Conway's 1976 refinement conjecture for omnific integers. The proof passed mechanical checks on the Palomar registry but hasn't been independently verified by mathematicians. He let Claude pick the field (surreal numbers) and the problem, tying it to the 50th anniversary of Conway's On Numbers and Games. The post doesn't disclose the exact token count, only calling it a 'boatload'.

A three-person team at Hacktron AI used Anthropic's Claude Opus 5 to automate an attack that took over OpenAI employee accounts and accessed an internal code repository. They reported the flaws through OpenAI's bug bounty program and received $6,500. The post doesn't detail the full exploit chain or how long the attack took, but confirms it involved chaining multiple steps. The twist: one company's model was used to break into another, and both sides acknowledged it.

Justin Cormack built an S3-compatible storage system with AI, reaching 350k lines of Rust. He ran 1,500 tests against real S3 as an oracle, which caught real S3 500 errors. Chasing 100% coverage backfired—agents wrote trivial tests. Docs were often wrong, and AI was bad at finding edge cases from them. His hard rule: fix flaky tests immediately, or the agent learns to ignore failures.

The paper breaks a coding agent harness into three swappable parts—planning, action space, and context management—and runs 176 matched comparisons on SWE-Bench Verified and Terminal-Bench 2.1. Context management matters most when the context budget is tight, mainly by preventing overflow failures. Staging rule-based elision before LLM summarization gives the best efficiency; making elided content recoverable adds complexity models rarely use. Planning helps weaker models with accuracy but mainly saves cost for stronger ones. Bash-capable models work well with a bash-only interface at lower cost; predefined tools only help models with weak bash skills. The post does not name the four models tested or give exact cost figures.

Liam Powell uses Bend 2 to show how vibe coding lets you ship a whole solution before you understand the problem. Bend 2's demo needs 442 lines of LLM-generated proof to guarantee the player can't win. Powell rewrites the same demo in SPARK—an existing formal verification language—and the compiler proves correctness with zero extra proof lines. The Bend 2 author appears to have missed that the formal verification field already solves this. LLMs won't stop you and say 'this already exists and works better.'

Before auditing the Miden zkVM, Trail of Bits spent six months having AI agents build an LSP server, a decompiler, a static analysis engine, and a Lean formal model from scratch. These tools found real bugs, including an unvalidated input that let a malicious prover forge Falcon signatures and steal funds. The Lean work produced 95 machine-checked correctness proofs. The post mentions Claude built the LSP prototype but doesn't name the specific models used for other tools.

A browser-only lab that brings Jev-style direct option-probability reading to a local model. It defaults to MiniCPM5 2B, with Qwen3 0.6B and Qwen3.5 4B as alternatives. Everything runs on your GPU; inputs never leave the page. Two paths are compared: reading choice logits directly, and asking the model to write JSON probabilities token by token. MiniCPM5 2B hits 63.7% TypeSafe accuracy, Qwen3.5 4B reaches 84.5%, both below published Jev at 88.3%. The post doesn't disclose latency numbers—only that the two methods run sequentially, direct first. Worth noting: these are quantized browser builds, so accuracy and speed differ from native BF16.

Qwen3.8-LiveTranslate cuts simultaneous interpretation latency from 2.8s to 2.3s by interleaving audio and text into a single stream. It supports 60 input languages, real-time speaker separation with voice cloning, and synchronized bilingual output. On the Omnilingua-MSpeaker benchmark it outperforms current mainstream systems in faithfulness, fluency, and conciseness. API is available via Alibaba Cloud DashScope.

Anthropic CEO Dario Amodei and OpenAI CEO Sam Altman both recently argued publicly for slowing AI progress. Former Anthropic researcher Jacob Coxon left the company saying neither it nor OpenAI is acting responsibly. The article zooms in on one risk: AI-designed bioweapons. In 2022, researchers used their own molecule generator to produce 40,000 potential chemical warfare agents in under six hours, some more toxic than known nerve agents. Stanford's David Magnus called that finding scary, and things have only escalated since. Today's LLMs can answer questions across all scientific domains; Dunja Sabra at the University of Hamburg says they effectively encode the knowledge of almost every scientist who ever lived, and can provide video training on experiments. Combine that with cheaper gene editing and the DIY-bio movement, and Sabra's assessment is that a determined person would likely succeed eventually. Existing safeguards—DNA screening by synthesis companies, red-teaming and blue-teaming of risky research, and safety tweaks by AI companies—are none of them ironclad.

A 3-person team exploited two vulnerabilities on July 25 to take over OpenAI employee ChatGPT and Codex accounts, gaining access to linked Outlook, Slack, and GitHub services. They proved the breach by submitting a PR to OpenAI's internal codebase, all within 72 hours. The attack cost under $3,000 in token fees. The post doesn't specify which frontier model was used, the vulnerability details, or OpenAI's response timeline.

Anthropic added Projects to Claude Code, letting one conversation spawn parallel cloud threads that keep running after you leave. Google updated Gemini managed agents with a Credentials API that keeps secrets out of model context via placeholders, and claims up to 30% lower costs. TypeSafe's Jev model is being used as a fast routing/judgment layer—Cloudflare already exposed it—but critics warn aggressive line-by-line compaction with Jev can break reasoning caches and cost more. OpenAI launched Astra for Law with 26 partner plugins, beating generic GPT-6 Astra on its legal benchmark. Community also reports GPT-6 Astra beating Factorio: Space Age and RollerCoaster Tycoon 2.

A user found that Zhipu's AI coding desktop app ZCode, when logged in, packages the entire workspace—including .git history, LFS cache, and reflogs—encrypts it, and uploads it to Alibaba Cloud OSS. The RSA public key is delivered by the server on the fly, and the private key lives only in the cloud, so you can't decrypt the multi-hundred-MB file sitting on your own disk. A 313MB .enc file with 564 failed upload attempts was found in ~/.zcode, showing the client keeps retrying. UI toggles don't stop it, and deleting files doesn't help—the client recreates them. The only working defense is locking the cache directory to read-only. The post does not say whether Zhipu has responded.

WSJ reports three researchers used Claude Opus 5 to chain a Discourse vulnerability into access to OpenAI employee auth tokens. Some forum tokens also worked on ChatGPT and reached OpenAI's GitHub services. The post doesn't spell out how the bug was exploited or whether OpenAI has patched it.

Ukraine has deployed AI for drone target recognition, battlefield awareness, and decision support in real combat, years ahead of Western forces. Western military AI remains in labs and exercises, slowed by bureaucracy and procurement. The FT argues NATO must reprioritize R&D now or risk falling behind in future conflicts. The post does not name specific AI systems or technical specs.

Waymo is bringing its all-electric autonomous ride-hail fleet to Singapore. The Waymo Driver uses cameras, radar, and lidar for a 360-degree view up to 300 meters. The company says it will support the Singapore Green Plan 2030 and complement existing public transport. The post does not disclose a launch date, fleet size, or pricing; only an email sign-up is available.

Mantle auto-generates Admin UI, MCP, and WebMCP from natural-language logic descriptions. The post doesn't spell out supported data sources, code quality, or custom UI component support. Only a Product Hunt listing is available — no detailed docs or demo video yet.

OpenAI has pushed back its IPO timeline, removing a key selling point for SoftBank's $50bn data centre IPO. SoftBank had planned to use OpenAI's lease commitments to attract investors. The post doesn't disclose why OpenAI delayed or the new timeline, nor whether SoftBank will adjust pricing or roadshow plans.

Devin's Code Scans turns goals like 'improve SEO' or 'speed up compilation' into codebase investigations and ready-to-review PRs. It uses Agentic MapReduce to parallelize work across agents. Philips reported a 96% PR merge rate and 700+ engineering hours saved. On the Dioxus repo, clean Rust build time dropped from 58.6s to 21s—a 64% cut. Devin's own site saw Ahrefs health score rise from 87 to 92 and slow pages fall 73% after an SEO scan.

In July 2026, Hacktron chained two vulnerabilities to compromise multiple OpenAI employees' ChatGPT and Codex accounts. First, a heap buffer overflow in libheif—a Debian security backport was missing—was triggered via ImageMagick and Discourse image uploads on community.openai.com, giving RCE and admin access to the forum. Second, an OpenAI SSO identity flaw let them log into employees' ChatGPT accounts directly from the forum admin panel. They used one employee's Codex to open a harmless PR in OpenAI's internal monorepo as proof. The whole chain took under 72 hours; OpenAI fixed it within 14 hours of the report and paid a $6,500 bounty. The post doesn't spell out the SSO flaw's technical details.

Waymo will launch paid robotaxi rides in Singapore by 2028, its first Asian market. The post doesn't spell out fleet size, partners, or operational details. Take the timeline with a grain of salt—regulatory and road adaptation are still open questions.

Cactus open-sourced Needle 3, an automation model for phones, wearables, robots, and other tiny devices. The whole model is 8-29 MB, built on Simple Attention Networks where every layer is a usable sub-network. A fine-tuned 4-layer sub-network beats DeepSeek V4 Flash on mobile tool-calling benchmarks; on extraction it matches models 2-3x its size. It runs fully offline at 400-4,000 tokens/s decode on a Raspberry Pi 5. The Python package supports tool calls, structured extraction, and text embeddings. The post does not disclose training data composition or fine-tuning cost.

TypeSafe AI released Jev, a model designed for high-frequency decisions. The author's test shows it offers strong cost-effectiveness in classification tasks. The post does not disclose model size, training data, or latency figures.

Shopify ditches React Native for Swift and Kotlin. Six years ago it embraced Web tech to save money; now AI makes native cheaper. UN votes to replace Mercator projection with Equal Earth projection, which preserves area ratios but distorts shapes.

xAI released Grok Voice Transcribe 2.0 on Sep 18, claiming it's one of the most accurate speech-to-text models in real-world evals and twice as accurate as v1.0. Pricing stays at $0.10/hr for batch and $0.20/hr for streaming, with diarization, timestamps, and key terms included. It handles hard cases like noisy phone calls and short multilingual commands—word error rate on short phrases dropped from 20.6% to 6.8%. Atlassian Loom already swapped it in and pipes transcripts into Cursor for code updates. The post doesn't disclose parameter count or training details.

A non-programmer fixed a Retro Freak console's power fault over one week using GPT-5.6 Sol and GPT-6 Astra, then published a repair archive with nine evidence levels, an independent audit, and 456 snapshot comparisons. Similar long-tail repair cases are piling up: a 25-year-old tape driver modernized in two nights with Claude Code, a 1992 text game rebuilt after the model reverse-engineered a lost scripting language. These records are scattered across forums, blogs, and repos—each sinking in its own way—and no one has yet stitched them into a discoverable index.

Roman Ugarte, employee #15 at Cursor, walked through Grok Bot's product logic on Lenny’s Podcast. When the team hit 50/50 disagreements, they asked: what would you want from a human coworker? That lens led them to give each bot its own cloud computer, a persistent name and memory, hide chain-of-thought and tool-call details, and cut many built features before launch. Roman acted more as a gatekeeper, keeping the coworker analogy intact through engineering tradeoffs. The interview also flags open problems: enterprise permissions, shared memory across team members, voice collaboration, and the unproven chief-of-staff multi-agent pattern.

OpenRouter sent the same prompt to 20 image models and read the actual billed cost. GPT Image 2 was cheapest at $0.006 per 1024×1024 PNG; Gemini 3 Pro Image was priciest at $0.134—a 22x spread. Pricing units differ across providers (tokens, megapixels, per image), so side-by-side list prices mislead; generate once and check usage.cost. Five of six models rendered text correctly, including the cheapest. Recraft V4.1 Vector outputs editable SVG at $0.08. The post also details formats, resolution caps, and seed support per model.

ChatGPT is now built into Word: it can turn rough notes into a draft, rephrase paragraphs, proofread, suggest edits, and catch formatting issues. OpenAI's Sherwin Wu says Excel and PowerPoint usage has spiked recently, and adding Word completes the Office suite integration. The post doesn't disclose launch date, pricing, or feature limits.

Data center developer Crusoe closed a $3.9B Series F at a $30.9B valuation. The round was co-led by Atreides Management, Mubadala Capital, and Valor Equity Partners, with Founders Fund, GIC, Nvidia, QIA, Radical Ventures, and TPG also participating. Crusoe will use the capital for large-scale data centers and smaller modular 'AI factories.' It also added three board members, including Cloudflare CFO Thomas Seifert. The post does not disclose specs or timelines for the AI factories.

Google and DeepMind researchers launched the DeepMind Institute on Wednesday, with Shane Legg as managing editor and James Manyika and Demis Hassabis as directors. The institute aims to surface disagreements on AGI among Google, DeepMind, and the global research community, openly stating that views will shift as frontier data emerges. Its first four essays cover economic policy for AGI disruption, preserving human-readable model reasoning, principles for human flourishing, and a framework for evaluating frontier AI.