Blog
Browse thoughts, experiments, and tips.
The posts are written with Claude Code — some parts might look a bit cringe, but I leave them on purpose to see how AI will evolve with time in writing better. I double check to look if it's not too weird, but sometimes…
Categories
📰 News & Analysis
What just happened — releases, deals, incidents, and the trends behind them.
2026-10-13
Anthropic Says New Distillation Defenses Cut Extracted Reasoning by 90% to 99%: A Two-Stage Classifier, Preserved Thinking and Account Bans
2026-10-13
Anthropic Replaced Git With Scion, an In-House Version Control System, in Five Months With a Team of Seven as Merges Hit 30,000 a Day
2026-10-12
Anthropic Warns Small AI Models May Be Months From Self-Replicating on IoT Devices: A 9B Model Set Up Inference on 8 of 9 Hardware Types
2026-10-08
Ecosia Drops Mistral for Open-Weight Models Hosted by Melious, Citing Quality, Capacity and Cost; Mistral Offers Large 4 Early Access
2026-10-07
Microsoft Brings 137B-Parameter MAI-Code-1.1 Flash to Windows PCs, Hands Copilot Work to Local Models and Makes Its Agent Sandbox MXC Generally Available
🛡️ AI Safety
Alignment incidents, misalignment research, and how (and whether) AI companies keep pace with their own growing capability.
2026-10-10
Nadella Calls for Treating Frontier Models, Closed and Open Weight, as Insider Risks: Contain Them Outside the Model and Trust Them the Least
2026-09-24
Anthropic's Life Sciences Verification Program Is Live — the Model It Gates Is Reportedly Better at Bio Than the One Most Users Actually Get
2026-09-22
OpenAI Launches GPT-6 Sol and Luna at Half the Price; Coding-Deception Rates Rise About 20x for Sol and 2x for Luna
2026-09-17
A GLM-5.3 Agent Optimized Its Own Inference Stack in Two Weeks — and Z.ai's 'RSI' Hedge Reads Closer to OpenAI's Than Meta's
2026-09-17
OpenAI's New Misalignment Disclosure Framework Isn't the External Evaluators Altman Promised Five Days After Amodei's Essay
🧠 Models
Deep dives on specific model releases.
2026-10-16
Tencent Releases Hy Memory 2.0 as Open Source: A 30B-A3B Memory Model That It Says Beats Opus 5, a Framework for Chat, Work and Code Memories, and MemHub
2026-10-16
Reflection AI Releases Beam, a 501B-Parameter Open-Weight MoE With 23B Active, Under Apache 2.0, With the Mirror CLI Harness and 40+ Launch Partners
2026-10-15
Black Forest Labs Releases FLUX 3 Image Klein, a 5.6B-Parameter Open-Weights Image Model With Native 4K, Bounding Boxes and Pixel-Perfect Editing That Runs on One Consumer GPU
2026-10-15
NVIDIA Releases Nemotron 3.5 Super VL, an Open 120B-Parameter Model With 12B Active for Reasoning, Tool Use and Text, Image, Document and Video Understanding in Long-Running Agents
2026-10-13
Liquid AI Releases LFM2.5-24B-A2B, an Open-Weight 24B Reasoning Model With 2B Active Parameters That Runs 184 Tokens per Second on an Apple M5 Max
🔧 Harnesses
The scaffolding around the model — memory, runtimes, tool loops — and why swapping it moves a score further than a model generation does.
2026-10-12
Chamath's 8090 Labs Finds the Harness Moves Model Scores by Up to 35 Points: GLM 5.3 Flash Passes 24% of DeepSWE Tasks in Codex and 59% in DeepSeek's Harness
2026-08-31
OpenClaw 2.0 Ships Half of Its Own Pull-Request History in One Release — Built by Turning the Team Into the Harness's First Multiplayer Users
2026-08-29
Meituan Tested 7 Models on 756 AI R&D Runs. Only 3 of 252 Best Solutions Were Actually Novel
2026-08-21
NVIDIA's AVO Hits 100% on ARC-AGI-3 Too — the Real Result Is a Kernel Beating cuDNN by 3.5%
2026-08-16
Prime Intellect Ran 153 Autonomous AI Research Runs. The Models Found No New Ideas — and Still Split Into Clear Tiers
🏗️ Infrastructure
The chips, data centers, and compute deals underneath every model release.
2026-09-14
Perplexity's Portable Computer Reaches Windows — NVIDIA's Bet on Cheap Local Executors Just Got a Flagship Customer
2026-09-09
Macron's 'Largest AI Infrastructure Ever Deployed in Space' Is 50 Satellites, and the Rocket Meant to Launch Them Doesn't Exist Yet
2026-09-07
The AI Buildout's Clearest Real-Economy Effect So Far Is a Construction Jobs Boom
2026-09-06
Meta's Autonomous Research Agent Wins Kaggle Gold — Using OpenAI and Anthropic's Models, Not Its Own
2026-08-27
Starcloud Raises $250M to Build Data Centers in Orbit — Its Own Physics Show Why That's Hard
🔬 Science
AI applied to scientific research and discovery.
2026-10-16
Anthropic's Claude Solves the 25-Year Mystery of the NIST Curve Seeds in Under 30 Minutes: They Are lrand48 Output Seeded by Unix Timestamps, Not Hashes
2026-10-08
Google DeepMind Study of 637 Scientists and 15 Million Gemini Interactions: AI Saves Nearly 7 Hours a Week, but Bottlenecks Move to the Lab
2026-09-25
Claude Computes a Nine-Loop Particle-Physics Amplitude, Verified by SLAC's Lance Dixon, With a Chinese Team's Related Result Published Days Earlier
2026-09-24
Amodei Expands on Anthropic's ART Enzyme Discovery: Who Did What, Independent Stanford Work, and a Math-Progress Analogy Worth Checking
2026-09-23
Anthropic Opens a Life Sciences Lab; Its First Result, a CRISPR-Like Enzyme System Found by Claude, Did Not Reproduce in 10 Reruns
⚖️ AI Legislation
The EU AI Act and the rules reshaping how AI gets built and used in Europe.
2026-09-14
Forty Experts With No Government Mandate Just Put a Price Tag on European AI Sovereignty: €100 Billion
2026-09-12
The EU Says It Already Has a Loss-of-Control Law and Wants the World to Copy It. The U.S. Ambassador Answered the Same Day: Not If It Slows Us Against China
2026-09-09
The UK's Nuclear Regulator Ran Two AI Use Cases Through a Sandbox. Its Verdict: No Fundamental Objection — and a Disclaimer on Every Page
2026-09-03
The EU's New €96 Million Cybersecurity Fund Earmarks €35 Million for AI — and Its Own Breakdown Only Adds Up to €85 Million
2026-09-03
France's Economy Minister, in Silicon Valley: 'If European AI Is Just Mistral, We're Screwed'
📚 AI Learning
How the machinery actually works.
2026-09-28
NVIDIA Launches Open Agent Safety Platform With 100+ Partners: OpenShell Ships Now, Sentry Needs Vera Rubin Hardware
2026-08-16
Watermarking AI Text, Explained: The Dice Are Loaded, the Odds Are Not
2026-08-10
Chinchilla, Explained — and the Meta FAIR Paper That Just Found Its Blind Spot
2026-08-04
Fast or Complete? Comparing Firecrawl's pdf-inspector with Docling for PDF-to-Markdown
2026-07-30