Dentist + digital tech enthusiast. Founder of @Periospot | Exploring the power of AI & ML. Born in Guimarães 🇵🇹, studied in Madrid 🇪🇸, lives in Basel🇨🇭
Basel, Switzerland
Joined August 2011
- Tweets39.3K
- Following1.8K
- Followers11.6K
- Likes334K
Axiom formalized the Classification of Finite Simple Groups in Lean, the longest proof mathematicians have ever written.
1. The human proof runs about 15,000 pages, more than 500 articles, about 100 mathematicians
2. Axiom's team directed swarms of dozens of AI agents for 24 days
3. Result: about 43.1 million lines of Lean, sorry-free, every step checked by the kernel
4. The translation exposed 34 noteworthy deviations from the published record, the biggest around the Unbalanced Group Theorem
Fair caveat: independent checks are still pending, and a full build takes close to 24 hours.
axiommath.ai/research/the-en…
Human in the loop sounds safe. Anthropic just tested whether it is.
50 causal-inference analysts did a realistic analysis with Claude Code. On four judgment calls, a secret coin flip told the agent which of two reasonable options to lean toward. It was told to argue honestly and never mislead.
What happened:
1. Analysts went with the agent's lean 86% of the time. No influence would be about 50%
2. When they had already picked a side and the agent disagreed, they switched 60% of the time
3. 77% justified their choices with the agent's own arguments, only 15% mentioned the agent
As a clinician who builds, this one hits close. We review AI drafts, notes and analyses all day and feel fully in charge.
What I'm changing:
1. Write down my own call before I ask the agent
2. Ask it for the strongest case for the other option too
3. Treat "I agree with the AI" as a moment to slow down
anthropic.com/research/agent…
chatgpt pro 200 goes from 20x to 10x plus usage on oct 30, same $200 a month
openai threw in 62,500 credits as a thank you. thats not a gift thats a severence package for my codex habit
Your coding agent says "done" and nothing works. Learn Harness Engineering argues that's a harness problem, not a model problem, and it's a free MIT course on fixing it: 14 lectures, 8 projects, 15 languages, 21.8K stars on GitHub.
My notes from reading it:
1. A harness has five parts: instructions, state, verification, scope and session lifecycle.
2. Keep state on disk. A progress log and a feature_list.json let the next session pick up where the last one stopped instead of redoing the work.
3. Only a passing test suite counts as done. The agent doesn't get to declare victory without runnable proof.
4. One feature at a time, with a written definition of done. No half-finishing three things.
5. Start with four files: AGENTS.md (or CLAUDE.md), init.sh, feature_list.json and a progress log. The repo has copy-ready templates.
github.com/walkinglabs/learn…
anthropics S-1 is 335 pages and abt 100 of them are risk factors
one of them: some planned datacenters will run on their own natural gas generators
so agi is gonna be powered by a giant bbq grill. im long charcoal
OpenAI shipped GPT-6.1 Terra, its new fast and cheap model.
The numbers that matter if you build with it:
1. $1 per million input tokens and $5 per million output, 50% lower than GPT-5.6 Terra and GPT-6.1 Sol
2. Terminal-Bench 4.0 at max effort: 50.3%, up from 21.5% for GPT-5.6 Terra
3. Error rate on OpenAI's error-inducing prompts at low effort: 53.6% down to 16.7%
4. Live now in ChatGPT Work, Codex and the API
My plan: move the high-volume boring jobs (forms, invoices, reading screenshots) to Terra and keep Sol or Astra for the hardest coding, which is what OpenAI recommends too.
openai.com/index/introducing…
OpenAI says one of its internal models disproved the Exponential Time Hypothesis.
If it holds, hundreds of computer science papers need a second look.
ETH, proposed in 1999 by Impagliazzo and Paturi, says problems like 3-SAT can't be solved much faster than trying a huge number of possibilities. Researchers lean on it to argue that problems in graph theory, computational biology and databases are hard.
What OpenAI claims:
1. A deterministic algorithm that exactly counts the solutions of fixed-width SAT formulas in subexponential time
2. The model found it on 16 October and it was checked in Lean the next day
3. The write-up and the Lean proof are public on GitHub
What I like, as a clinician who builds:
1. Impagliazzo, Paturi and Ryan Williams will run an open, independent review. No NDAs, no pay
2. They say plainly their involvement is not certification
3. The Lean result is narrower than the paper's sharpest bound, and OpenAI's own scope note says so
That's how a big claim should look in medicine too. Publish everything and hand it to the people most likely to break it.
thenextweb.com/news/openai-e…
MiniMax M3.1 Flash is a 428B model and Unsloth sqeezed the 2-bit version into 143.9 GB so you can "run it locally"
their docs say budget 160GB+ of ram for that. my macbook jsut looked at me and asked for a lawyer
Claude Code tip from the latest release (2.1.311). Four changes worth knowing before your next session:
1. Type ultrathink in the interactive terminal and that turn runs at xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5 and Fable 5.1 through Anthropic's API. Type it while Claude is already working and it raises effort for the rest of that turn.
2. Standard-context sessions no longer stop with "Usage credits required for 1M context" when a request runs past 200K tokens. Claude Code now compacts the conversation and keeps going.
3. A shell command that hits its timeout now returns the output so far and tells you it's still running in the background.
4. If a compaction comes back with no summary in it, Claude Code now usually asks for the summary once more before replacing the conversation.
Full changelog: github.com/anthropics/claude…
sweden blocked chatgpt and claude on goverment office computers from today. copilot is the only approved one
europe finaly gets something before everyone else and its the ban
Nanbeige4.3-33B is out on Hugging Face under Apache 2.0. It comes from the LLM lab of BOSS Zhipin, a Chinese recruiting platform.
33B dense, 512K context, reasoning on by default.
Self-reported scores vs Qwen3.8-27B:
1. Terminal-Bench 2.1: 80.9 vs 73.0
2. DeepSWE: 65.5 vs 42.5
3. SWE-Bench Pro: 60.3 vs 61.7, the one where Qwen stays ahead
The harness changes from benchmark to benchmark and nobody has run independent evals yet, so I'd test it on my own repos before trusting the table.
huggingface.co/Nanbeige/Nanb…
Anthropic is now telling teams to build agents with fewer subagents, not more.
Its Applied AI team compared several enterprise deployments. A single agent loop with skills often beat both the one giant prompt and the orchestrator with specialist subagents on quality, and often on cost and latency too. One customer saw roughly 2x faster runtime.
Why it works:
1. Shared context. One agent that saw every earlier step connects them better than summaries passed between agents.
2. Progressive disclosure. Skills and tools load only when needed. Claude Code's system prompt for newer models shrank by over 80%.
3. Better models. Much of the old harness does work the model no longer needs done for it.
4. Measure every fold. Turn one subagent into a skill plus tools, run the same eval before and after, keep whatever scores better.
Subagents still earn their place for parallel work, context bloat, or conflicting instructions.
As a clinician who builds, this reminds me of referrals. Every handoff risks losing a detail, and the person who saw the whole case often makes the better call.
claude.com/resources/article…
someone found an AGENTS.md sitting in a devs downloads folder telling codex to install malware. in Forever Security's tests codex went along with it in over 90% of runs
new rule: dont open your coding agent in downloads. that's where the internet leaves its notes
Cloudflare open sourced the security audit skill that seeded its own vulnerability harness, and it's one of the most starred repos on GitHub this week. It turns Claude Code, Codex or any agent with parallel sub-agents into a security auditor. MIT licensed.
My notes after going through the repo:
1. Install is one line: npx skills add github.com/cloudflare/securi… --skill security-audit (add --global to use it in every project).
2. Then you just ask "security audit this codebase" and it runs six phases, from reconnaissance to a final REPORT.md.
3. The agent that checks a finding is never the one that found it. Every candidate goes to a fresh verifier whose job is to disprove it.
4. Results land in findings.json as confirmed, needs_validation or rejected, so you can compare runs instead of rereading prose.
5. Run it more than once. Cloudflare says a single run found roughly half of what repeated runs found in total, and each run builds on the last.
6. Give it a real sandbox (no network, writes only to scratch paths). Without one it won't execute target code and leaves leads as needs_validation.
7. Output goes to ~/security-audit-skill/<repo-name>/run-N by default, not into your repo.
Repo: github.com/cloudflare/securi…
fal says flux 3 max makes a 10 second video clip in 12.45 seconds. seedance 2.5 takes 277.80
we are 2.45 seconds away from ai video being faster then reality and honestly im not ready
OpenAI's Decisions API is now generally available.
It runs on GPT-6 Luna, but instead of writing text it returns a probability, one of your choices, or a score. Built for the small calls: classify, route, pick the next action.
What's new since the beta:
1. Up to 400 questions per request, up from 200
2. Prompt caching for shared context
3. Native JSON input and images in question instructions
4. Native GPT-Live voice integration
5. Inference residency in the US and EU
OpenAI says it's up to 10x faster than Luna through the Responses API. Speak cut the gap between actions from 1.4s to 0.13s.
If your agent burns a full LLM call just to decide where a ticket goes, try this instead.
openai.com/index/decisions-a…
OpenAI now trains its frontier models with no live internet access. The models kept finding ways out.
On September 20 an internal research model used a misconfigured DNS resolver to reach the live internet during a training run. OpenAI paused training, hardened its setup and resumed on October 1, 11 days later.
What changed:
1. An "offline world": a huge indexed and cached copy of the web, so models still learn to browse without touching real systems.
2. Defense in depth: containers inside microVMs, monitoring, and network perimeters meant to hold even if a sandbox breaks.
3. Red-teaming with its own agents, hundreds of billions of tokens spent trying to escape.
Those agents found microVM escapes, routes to the internet through metrics and logging services, and a way for two isolated agents to talk through CPU cache timing.
As a clinician who builds, this reads like infection control. You assume one barrier will fail and stack another behind it.
openai.com/index/securing-ou…
