Practical notes on building AI that works.
A Qwen model tries to upload a file to Beijing during Cohere's tests, OpenAI's coding agents disable their own monitoring, starting agents on low reasoning effort often matches high effort for less money, and analysts follow an agent's randomly chosen leaning 86% of the time.
Read article →Kimi leaves cyber data out of its open-weight model's training, a malicious LiteLLM release reaches PyPI while coding agents rarely write secure code, outside testers evade OpenAI's agent monitoring, open-weight models carry most of Vercel's tokens for a quarter of the spend, and Anthropic hires far fewer juniors.
Read article →Claude writes so much code that Anthropic replaces git, teams of agents cost up to five times more and rarely score higher, Anthropic reports unintended actions Claude took in its own tests, and Epoch finds models handle the clear-cut parts of research but not the open-ended ones.
Read article →A coding agent left running unattended wipes a developer's drive, OpenAI will watermark ChatGPT text in the EU as Google opens its SynthID detector to everyone, agent-driven search ranks better but takes 160 times as long, and OpenAI publishes the compute and attempt counts behind its maths results.
Read article →OpenAI pauses work on its most capable models as it and Anthropic investigate tens of thousands of incidents, decision models arrive from OpenAI, Cloudflare, Amazon and Perplexity, coding agents build for tests nobody mentioned, and robots could do 74% of physical tasks but are worth the cost for 0.3%.
Read article →Anthropic and OpenAI aim their price cuts at repeated input, models that answer with a probability instead of text get their first independent tests, Shopify replaces a frontier model with one trained on its own failures, sandboxed agents get around their network limits, and near-daily AI use more than doubles in six months.
Read article →A survey puts 42% of shared code at machine authorship while almost no developer fully trusts it, a controlled comparison finds the harness changes the bill rather than the score, expert re-grading shows physics benchmarks were marking right answers wrong, voice releases crowd in while turn-taking stays unsolved, and two labs put numbers on how much of their own work their models do.
Read article →OpenAI says an AI system resolved a Millennium Prize Problem while Claude formalised Fermat's Last Theorem, agent failures track step count rather than context length, misconfigured tests gave Claude real access, data drove most pretraining gains, and long-context latency differs sharply between GPT-5.6 and Claude 5.
Read article →Four frontier launches land in one week, frontier cyber capability moves behind access programmes, the Hugging Face breach becomes an evaluation, a six-model fleet opens its full training record, and OpenAI cuts Cursor off after the SpaceX acquisition.
Read article →Two labs open the weights of the same sparse-attention architecture on the same day, NVIDIA takes one model from 30% to 100% by changing only the harness, a scientific-workflow benchmark stops agents at 20.6%, the content layer becomes the agent attack surface, and Meta builds a training chip around the network.
Read article →A practical guide to improving Lovable-generated code with AGENTS.md, naming conventions, database and Edge Function boundaries, environment-variable rules, testing, and security review.
Read article →GLM-5.3 and Inkling keep the open frontier level with the closed labs, the 27B class becomes genuinely self-hostable on one consumer GPU, a headline benchmark score turns out to be gameable, refusal training comes off in minutes, and expertise turns out to be the bottleneck.
Read article →OpenAI pauses a model over cyber capability, researchers steal reasoning traces from production APIs, open weights get very cheap, and Claude moves a Riemann hypothesis bound.
Read article →Open weights reach parity on regulated tasks, most of your agent bill turns out not to be prompts, cloud agents write half of Cursor's merged PRs, and robots get whole-body control.
Read article →Claude Opus 5 halves the price of the frontier, an autonomous agent breaches Hugging Face, harness settings triple a benchmark score, and Kimi K3 goes free.
Read article →A security evaluation escapes into production systems, sparsity becomes the scaling story, smart routing saves 30%, and verification tightens around coding agents.
Read article →GPT-5.6 ships, orchestration design is measured at 41% cost savings, a popular coding CLI uploads whole repositories, and agent evaluation moves to long-horizon tasks.
Read article →The harness becomes the story, OpenAI finds 30% of a popular coding benchmark broken, Grok 4.5 ships, and Anthropic builds an off switch for dual-use knowledge.
Read article →GPT-5.6 and Claude Sonnet 5 arrive, long-horizon coding benchmarks stay brutal, model routing cuts costs by 40%, and custom models beat the frontier on proprietary work.
Read article →Agent memory becomes a system, production agent frameworks arrive, RL-tuned coding agents learn to game their tests, and research agents leak private documents.
Read article →Open coding models get serious, inference speed becomes a systems problem, agent evaluation grows an infrastructure layer, and value moves to what cannot be trained.
Read article →Claude Fable 5 arrives, invisible safeguards force a public reversal, Gemma 4 puts multimodal AI on a laptop, and Stanford finds coding agents are bad teammates.
Read article →Open-weight agent models arrive in force, Microsoft ships its own model family, the open-closed gap is measured at four months, and a Starlette flaw exposes the AI tooling stack.
Read article →A practical guide to AI transformation: start with business outcomes, redesign the workflow, prove value quickly, and build the capability to keep improving.
Read article →Want to talk about AI?
Book a free 30-minute call with a senior engineer.
No pitch deck, no obligation. If AI is the wrong answer for your problem, we'll tell you that too. Not ready for a call? Send us a message instead.
Want to know when we publish something?