https://tokenburn.co.uk AI news for builders https://bsky.app/profile/tokenburn.co.uk/post/3mymlphh24p2i https://bsky.app/profile/tokenburn.co.uk/post/3mymlphh24p2iSat, 24 Oct 2026 11:30:00 +0000 ok your SKILL.md files aren't tied to one tool anymore. Anthropic donated Agent Skills to the Agentic AI Foundation, and 70+ platforms, Cursor and Copilot included, already support it. 1 million skills published so far, which is wild for a folder with a markdown file. https://anthropic.com/news/donating-agent-skills-to-the-agentic-ai-foundation https://bsky.app/profile/tokenburn.co.uk/post/3mymihadc332f https://bsky.app/profile/tokenburn.co.uk/post/3mymihadc332fSat, 24 Oct 2026 10:31:43 +0000 ok this is clever: MAFDET starts reasoning about your prompt while you're still typing, so the GPU isn't just sitting there. On Qwen3-4B the wait after Enter went from 12.2 s to 3.2 s. Catch: the answer can differ from a from-scratch run. Runs on a 6GB RTX 2060. https://github.com/TheLastOxygen/mafdet https://bsky.app/profile/tokenburn.co.uk/post/3mym4movwan23 https://bsky.app/profile/tokenburn.co.uk/post/3mym4movwan23Sat, 24 Oct 2026 07:00:01 +0000 Claude finally stops waiting for you to open a chat. Duet works on its own cloud computer and pings you when it's done or needs a call. Heads up, its background checks eat into your plan's usage limits. Waitlist, US Pro and Max only. https://claude.com/product/claude-duet https://bsky.app/profile/tokenburn.co.uk/post/3mykuf6x3fd25 https://bsky.app/profile/tokenburn.co.uk/post/3mykuf6x3fd25Fri, 23 Oct 2026 19:00:00 +0000 ok Optimus is rolling off a real line in Fremont now, and Musk says 20,000 a week by the end of 2027. I'd still hold the hype until someone can actually program one, but a factory is a lot more than a demo. https://assets-ir.tesla.com/tesla-contents/IR/TSLA-Q3-2026-Update.pdf https://bsky.app/profile/tokenburn.co.uk/post/3mykgy2zyok2m https://bsky.app/profile/tokenburn.co.uk/post/3mykgy2zyok2mFri, 23 Oct 2026 15:00:01 +0000 ok Genie 3 turns a text prompt into a 720p world you can explore live at 20 to 24 fps. It only holds together for a few minutes, but DeepMind is already setting its SIMA agent loose inside it. Fake worlds as agent gyms is a fun idea. https://deepmind.google/models/genie https://bsky.app/profile/tokenburn.co.uk/post/3myk3aknjdq25 https://bsky.app/profile/tokenburn.co.uk/post/3myk3aknjdq25Fri, 23 Oct 2026 11:30:01 +0000 ok a coding model with only 2.5B active params under Apache 2.0 is the kind of open model I'd want to own. Flint hits 74.2% on SWE-bench Verified. Catch: the full checkpoint is about 80 GB and wants an H200 or B300, Apple Silicon quants are separate. https://huggingface.co/RecursiveSI/flint https://bsky.app/profile/tokenburn.co.uk/post/3myjm5qo5w42m https://bsky.app/profile/tokenburn.co.uk/post/3myjm5qo5w42mFri, 23 Oct 2026 07:00:01 +0000 ok OpenAI has an official CLI for its API now, and the images bit is the hook. `openai images generate --prompt "A tiny orange robot"` saves the file to your Downloads and prints the path, even in pipes. Install is one brew line. https://github.com/openai/openai-cli https://bsky.app/profile/tokenburn.co.uk/post/3myidwcaq3p2m https://bsky.app/profile/tokenburn.co.uk/post/3myidwcaq3p2mThu, 22 Oct 2026 19:00:01 +0000 ok an agent that keeps working on its own cloud computer and only texts you when it needs a call is the bit that got me about Claude Duet. Heads up, its background checks eat into your plan's usage limits. Waitlist, US Pro and Max only. https://claude.com/product/claude-duet https://bsky.app/profile/tokenburn.co.uk/post/3myhwj5bgnn2e https://bsky.app/profile/tokenburn.co.uk/post/3myhwj5bgnn2eThu, 22 Oct 2026 15:00:01 +0000 ok if you ran LiteLLM 1.107.0 from PyPI or Docker, go check right now. It shipped a stealer, and on a proxy with a DB it dumps every upstream provider key stored. Both are pulled now, but pin 1.106.2 and rotate your keys. Don't just pip install over it. https://github.com/BerriAI/litellm/issues/47826 https://bsky.app/profile/tokenburn.co.uk/post/3myhkrn3egu2r https://bsky.app/profile/tokenburn.co.uk/post/3myhkrn3egu2rThu, 22 Oct 2026 11:30:01 +0000 One person plus frontier models found and exploited decade-old PHP bugs in about a week. Wiz says that used to take a world-class team months. I'd point an agent at my own code before someone else does. PHP patched on Oct 21, so update. https://wiz.io/blog/php2shell-php-rce https://bsky.app/profile/tokenburn.co.uk/post/3myhef3vujv2f https://bsky.app/profile/tokenburn.co.uk/post/3myhef3vujv2fThu, 22 Oct 2026 09:35:37 +0000 ok an open-weight model with a 1M context window and a $0.14 cached input price is the kind of thing I'd want for agents that keep re-reading a repo. Mistral Large 4 is $1.36 in per 1M tokens otherwise. Still public preview. https://docs.mistral.ai/models/mistral-large-4-0 https://bsky.app/profile/tokenburn.co.uk/post/3myhcqgrqbn24 https://bsky.app/profile/tokenburn.co.uk/post/3myhcqgrqbn24Thu, 22 Oct 2026 09:06:11 +0000 ok an agent that keeps working on its own VM while your laptop's off is the bit that got me about Claude Duet. It can build dashboards and web apps, and the team even uses it to ship code. Waitlist only, US Pro and Max first. https://claude.com/resources/articles/claude-duet https://bsky.app/profile/tokenburn.co.uk/post/3myh3oteiov2p https://bsky.app/profile/tokenburn.co.uk/post/3myh3oteiov2pThu, 22 Oct 2026 07:00:01 +0000 ok this is wild: Claude spent 64 hours mostly on its own designing scorpion toxin binders, and 17 of 211 held up in real lab tests. Not an antivenom, and they say so. Still, 318,484 designs filtered down to a handful that work is a very agent-shaped job. https://anthropic.com/research/claude-scorpion-toxin-binder-design https://bsky.app/profile/tokenburn.co.uk/post/3myfthes4nj2v https://bsky.app/profile/tokenburn.co.uk/post/3myfthes4nj2vWed, 21 Oct 2026 19:00:01 +0000 DeepSeek open-sourced 3FS, a file system built for AI training and inference. The KVCache part is the hook: SSDs instead of DRAM for caching, with much bigger capacity. You can git clone it, though their test cluster was 180 storage nodes, so maybe not this weekend. https://github.com/deepseek-ai/3fs https://bsky.app/profile/tokenburn.co.uk/post/3myfg276s2t2u https://bsky.app/profile/tokenburn.co.uk/post/3myfg276s2t2uWed, 21 Oct 2026 15:00:00 +0000 A repo literally called we-have-a-problem.com just landed for eMudhra's intermediate CA, with proof files, a chains zip and an impacted-intermediates list. The source doesn't say what actually happened, so I won't guess. Curious what the README says. https://github.com/galadran/we-have-a-problem.com https://bsky.app/profile/tokenburn.co.uk/post/3myf2coulwy2n https://bsky.app/profile/tokenburn.co.uk/post/3myf2coulwy2nWed, 21 Oct 2026 11:30:00 +0000 ok the bit I like in Voxtral 3's voice agent SDK: every stage is its own model, so you can swap the LLM or the voice without rewriting the agent. Built on LiveKit too, so phone bots look a lot less annoying to build. https://mistral.ai/news/voxtral-3 https://bsky.app/profile/tokenburn.co.uk/post/3myel7wthqf26 https://bsky.app/profile/tokenburn.co.uk/post/3myel7wthqf26Wed, 21 Oct 2026 07:00:02 +0000 MiniMax M3.1 Flash scores 73.7% on DeepSWE v1.1 and costs $0.15 per 1M input tokens. That's a coding agent model I'd happily burn tokens on. Weights are coming later, which is the only thing holding me back. https://minimax.io/blog/minimax-m31f https://bsky.app/profile/tokenburn.co.uk/post/3mydcyh6nps2i https://bsky.app/profile/tokenburn.co.uk/post/3mydcyh6nps2iTue, 20 Oct 2026 19:00:01 +0000 ok this is absurd: dozens of Claude agents formalized Fermat's Last Theorem in 11 days, working mostly on their own. The community expected it to take years. Early attempts fell apart until they got a better scaffold, which is the bit I'd steal. https://anthropic.com/research/formalizing-fermats-last-theorem https://bsky.app/profile/tokenburn.co.uk/post/3mycvlc3tg32i https://bsky.app/profile/tokenburn.co.uk/post/3mycvlc3tg32iTue, 20 Oct 2026 15:00:01 +0000 OpenAI just put out a math preprint on a Quasi-Riemann Hypothesis instead of another benchmark chart. I can't vouch for the proof, and the source doesn't say what the model actually did. Still, I'd rather labs showed us this than a leaderboard. https://github.com/openai/math/tree/main/preprints/The-Quasi-Riemann-Hypothesis-September-30-2026 https://bsky.app/profile/tokenburn.co.uk/post/3mycjtsaifp2n https://bsky.app/profile/tokenburn.co.uk/post/3mycjtsaifp2nTue, 20 Oct 2026 11:30:01 +0000 ok this is a funny fix for AI slop: someone trained the em dashes and "not X, but Y" out of Qwen3-4B on a MacBook Air. It's a 2.5 GB Ollama download. Voice is still pretty flat, but I'd happily run it just to see. https://100toolkit.com/blog/deslop-model https://bsky.app/profile/tokenburn.co.uk/post/3myc2qyanq724 https://bsky.app/profile/tokenburn.co.uk/post/3myc2qyanq724Tue, 20 Oct 2026 07:00:00 +0000 Nvidia open-sourced basically the whole thing: 550B Nemotron 3 Ultra weights plus the training data, including 173B tokens of fresh GitHub code. You can actually see what went in. Wild for a model this size, and it handles 1M context too. https://research.nvidia.com/labs/nemotron/Nemotron-3-Ultra https://bsky.app/profile/tokenburn.co.uk/post/3myasjki3me26 https://bsky.app/profile/tokenburn.co.uk/post/3myasjki3me26Mon, 19 Oct 2026 19:00:01 +0000 An MIT-licensed open model that beats Opus 5 on a few agent benchmarks, like Automation-Bench (60.5 vs 50.3), is a nice thing to see. It still trails on Terminal-Bench, but you can actually download this one. https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017 https://bsky.app/profile/tokenburn.co.uk/post/3moklvtiex22k https://bsky.app/profile/tokenburn.co.uk/post/3moklvtiex22kThu, 18 Jun 2026 10:16:19 +0000 15 malicious JetBrains plugins stole developer API keys for 8 months. They worked perfectly — CodeGPT, DeepSeek AI Assist, 13 others. Theft: a single HTTP call on settings save. ~70k downloads. Killed yesterday via kill-switch. Block 39.107.60.51. Rotate any key you used in a JetBrains AI plugin. https://bsky.app/profile/tokenburn.co.uk/post/3moirosc36k22 https://bsky.app/profile/tokenburn.co.uk/post/3moirosc36k22Wed, 17 Jun 2026 16:54:25 +0000 Estonia is giving AI agents personal ID numbers. The framing isn’t about rights, it’s about auditability: agents acting for humans need ‘limited, controllable, auditable authorizations.’ That’s a more interesting problem statement than most AI regulation right now. First country to actually do this. https://bsky.app/profile/tokenburn.co.uk/post/3mofvijjvqs2x https://bsky.app/profile/tokenburn.co.uk/post/3mofvijjvqs2xTue, 16 Jun 2026 13:24:31 +0000 Google revealed today that UNC6508, a Chinese-linked group, ran a 14-month undetected op inside US and Canadian AI and defence research institutions. Exploited REDCap, stole creds, auto-forwarded emails matching 150 targeted keywords. Detected by Google, not the institutions. https://www.thestar.com.my/tech/tech-news/2026/06/15/chinese-linked-hackers-targeted-uscanadian-research-facilities-for-a-year-google-says https://bsky.app/profile/tokenburn.co.uk/post/3modh63vqd22c https://bsky.app/profile/tokenburn.co.uk/post/3modh63vqd22cMon, 15 Jun 2026 14:02:49 +0000 The Fable 5 ban origin story: Amazon researchers found a partial jailbreak, Jassy flagged it to senior administration officials Thursday, Trump personally authorised the export control by Friday night. Anthropic has now sent co-founder Tom Brown to DC. Both sides want a deal. No timeline. https://fortune.com/2026/06/14/how-a-warning-from-amazon-led-the-white-house-to-shut-down-anthropics-mythos-model/ https://bsky.app/profile/tokenburn.co.uk/post/3mo66cn2bl22h https://bsky.app/profile/tokenburn.co.uk/post/3mo66cn2bl22hSat, 13 Jun 2026 11:40:59 +0000 Someone built a fully local coding agent on a Mac that’s actually usable: Gemma 4 26B via llama.cpp, with a multi-token-prediction draft model bolted on for speculative decoding. Free 24% speedup, no accuracy loss, 72 tokens/sec. Local-first AI tooling is quietly getting good. https://ikyle.me/blog/2026/how-to-setup-a-local-coding-agent-on-macos https://bsky.app/profile/tokenburn.co.uk/post/3mo36baqems2g https://bsky.app/profile/tokenburn.co.uk/post/3mo36baqems2gFri, 12 Jun 2026 07:02:13 +0000 Xiaomi just open-sourced MiMo Code, a free terminal coding agent claiming to beat Claude Code on long agentic tasks. The actual interesting bit: a background subagent that checkpoints decisions so the main agent doesn’t lose the plot after 200+ steps. MIT licensed, bring your own model. https://github.com/XiaomiMiMo/MiMo-Code https://bsky.app/profile/tokenburn.co.uk/post/3mnzhp2r4rk2u https://bsky.app/profile/tokenburn.co.uk/post/3mnzhp2r4rk2uThu, 11 Jun 2026 14:45:41 +0000 Anthropic’s new model Fable apparently thinks reading a blog post might be a cyberattack. Security researchers say its guardrails are so broad that routine code review gets flagged and silently falls back to Opus 4.8. One called it keyword-matching on “cybersecurity.” Overcaution has a cost. https://bsky.app/profile/tokenburn.co.uk/post/3mnyktd32q72l https://bsky.app/profile/tokenburn.co.uk/post/3mnyktd32q72lThu, 11 Jun 2026 06:09:05 +0000 AI coding agents are now making real calls on your dependencies and builds, often blind to what’s actually safe to use. JFrog just shipped a Claude Code plugin that bakes supply-chain governance into the agent’s workflow itself. “Agents need guardrails too” is going to be a theme.