Meta posed as teens to bait rivals 🔑, one retry loop billed 21x 🧑‍💻, a Pixar short for $1.33 🎬

Hundreds of contractors on a Meta project codenamed Cannes were told to pose as minors and test how ChatGPT, Gemini and Character.AI handled high-risk prompts. They created dummy under-18 accounts and sent prompts about suicide, eating disorders and sex, with one August 2025 round running 45,000 prompts past unaware rivals. If you run a chatbot, adversarial testing may now arrive from a competitor wearing a child's identity.

The Bank for International Settlements used its 2026 annual report to warn the trillion-dollar AI investment race resembles past bubbles, from 1800s railway mania to the dotcom boom. It estimates the five largest hyperscalers will spend over $1tn on AI capex this year, commitments that already outpace their cash flow. Its warning that weak returns could trigger a protracted investment bust is a signal the cheap-inference subsidy may not last.

Cursor now offers a native iOS app in public beta that lets developers drive coding agents from a phone. From the app you start cloud or local agents against any repo, give instructions by voice, then review diffs and merge pull requests. For engineers, it pushes the always-on agent into the on-call workflow, where you open an incident investigation by phone and return to a ready PR.

Governor Gavin Newsom announced a partnership making Anthropic's Claude the first AI productivity tool available to all California state agencies, plus its cities and counties. Agencies get Claude at a 50% discount with free workforce training and technical help from Anthropic developers, framed as augmenting rather than replacing public workers. For vendors, it is a template for how a frontier lab lands a whole-of-government account across an entire US state.

South Korea has earmarked at least $880bn in planned investment to expand its chip manufacturing, AI data centres and robotics capacity over the coming years. President Lee Jae-myung set out the Three Mega Projects alongside Samsung and SK Hynix, calling semiconductors, physical AI and data centres a national triple axis. It marks where new memory and accelerator supply will be built, as chip shortages already lift prices.

Amazon is hunting for alternatives to Claude after a renegotiated contract shifts it to token-based pricing, according to The Information. The pricing does not start until next year, and Amazon disputes its costs will rise, but it already pressures tools like Kiro, Quick and Alexa for Shopping. It is a reminder of single-provider renegotiation risk, even for a backer that put $4bn into Anthropic.

Even San Francisco engineers on $180,000 say they can no longer afford the city as AI wealth floods in. One couple earning $180,000 and $185,000 gave up after three months hunting a one-bedroom under $5,000, and one $5,200 listing drew 30 sign-ups within an hour. For engineers weighing where to work, it signals how OpenAI and Anthropic's IPOs are repricing the city even for the well-paid.

A SaaS shipped in two days re-ran one tenant's LLM batch job 21 times, costing more than a month of servers. Every LLM call returned 200 and was billed, but the final database write hit a missing column, so the job 500'd and each retry re-ran the whole expensive batch. Wrap LLM batch jobs in idempotency keys and re-run guards so a late failure never re-bills calls that already succeeded.

Runsybil rebuilt last year's DARPA AI Cyber Challenge results using only general-access models and a naive harness. Stock Opus 4.8 with Claude Code and GPT-5.5 with Codex each flagged more vulnerabilities than any AIxCC competitor, and Opus proved and patched bugs on par with the third-place team, for about $600. Before building a bespoke security pipeline, test whether an off-the-shelf model plus a thin harness already clears your bar.

The Zenith harness points itself at a hard, long-horizon engineering task and builds a control loop tailored to it. On Frontier SWE's seventeen ultra-long-horizon tasks, that loop took GPT-5.5 from fifth on the default Codex harness to first by mean@5, ahead of the larger Claude Fable. When an agent stalls, improve the loop around the model before swapping in something bigger.

PAR Technology built a multi-tenant text-to-SQL agent where the same question returns each user only their slice of data. Rather than trusting the model to filter, three independent layers enforce isolation, AWS SigV4 request signing, semantic validation on Bedrock, and Split-Plane SQL, so a compromised LLM still cannot cross tenants. Treat the model as a probabilistic generator, not a policy engine, and enforce row-level security in layers it cannot bypass.

After being let down by past local models, Piotr Migdal rates Qwen 3.6 27B the first that feels like general intelligence on his own machine. He recommends the slower dense 27B over the faster 35B mixture-of-experts variant, running an 8-bit quantisation through llama.cpp, noting it runs hot enough to warm his knees. For everyday local coding and writing, start with the dense 27B at Q8, not the MoE.

Vipul Ved Prakash argues AI is repeating the PC era's modular unbundling, splitting the stack into specialising layers. He points to three standardised interfaces, the transformer architecture, OpenAI-compatible inference APIs, and agentic harnesses, while Together's token throughput climbed from 30B to 400T a month, 10,000x. It reframes open-weights models as an economic consequence of modularisation rather than a values choice, useful when you decide which layer to build in.

Marketing AI agents as colleagues rather than tools degrades the humans meant to supervise them. Emma Wiles's study of 1,261 managers found people caught 18% fewer errors and were 44% more likely to escalate questionable output when it came from an 'AI employee' rather than a chatbot. With 23% of firms already listing agents on org charts, the framing you choose shifts accountability away from the human reviewer.

Tom Tunguz models how far the broader software market sits from AI-native firms' compute spend. Anthropic spends about 2.3x payroll on compute, roughly $2M per employee a year, against the top 1% of firms at $89k and the median near $137. His bear, base and bull scenarios project where the market lands by 2029, so the question is which trajectory you budget for in 2027.

The authors argue formal verification is now practical because AI eliminates the labour of writing proofs. They show a permission system whose thorough, passing test suite still misses a role that escalates single-environment access to near-universal, a bug a subset-invariant proof makes impossible by construction. The shift is from the code did not fail at runtime to the code cannot fail, for the invariants you cannot afford to get wrong.

Hamel Husain reframes 'our product is hard to eval' as a product smell rather than an evals problem. If an artifact is hard for you to verify it is usually hard for users too, so designing for verifiability should precede building evals. Surface the assumptions behind an answer, a data agent's query, its metric definitions, its unmatched rows, so outputs are checkable without users redoing the work.

Lark CLI is the official command-line tool for Lark and Feishu, built for humans and AI agents. It exposes 200+ commands and 26 prebuilt Agent Skills across 18 business domains, from Messenger and Docs to Base, Sheets, Calendar and Meetings, with structured output tuned for agents. Point a coding agent at it and the agent can send messages, update Base tables or schedule meetings without custom integration work.

codebase-memory-mcp is a code-intelligence server that turns a repository into a persistent knowledge graph of functions, classes and call chains. It parses 158 languages via tree-sitter, full-indexes the 28M-line Linux kernel in three minutes, answers structural queries in under a millisecond, and comes as a single dependency-free binary. An agent queries the graph instead of grepping file by file, which the project measures at roughly 120 times fewer tokens.

OpenMontage is an open-source, agentic video production system that turns an AI coding assistant into a full video studio. From a plain-language brief the agent researches, scripts, builds a corpus from free stock and generated clips, edits a timeline and renders the result through Remotion. Builders use it to ship a finished short end to end, like a 60-second Pixar-style animation the project produced for $1.33.