The AI Programming Plateau: Why Merge Rates Have Stagnated Since 2025
Statistical evidence suggests that LLM programming capabilities have not actually improved for over a year when measured by code mergeability.
OpenAI Debuts GPT-5.4: The Frontier Model for Professional Agents
1019
OpenAI's GPT-5.4 is a professional-grade model that introduces native computer interaction and high-efficiency tool use for autonomous agents.
GPT-5.4 Thinking Sets New Safety Bar as First General-Purpose Model with Cybersecurity Mitigations
1019
GPT-5.4 Thinking is OpenAI's first general-purpose model with high-capability cybersecurity safety mitigations.
Claude Opus 4.6 Solves Knuth's Hamiltonian Cycle Problem for Odd m
837
Don Knuth details how Claude Opus 4.6 successfully solved a difficult graph theory conjecture for odd m through iterative algorithmic discovery and creative deduction.
The Car Wash Test: Why AI Still Lacks Common Sense
1516
AI models fail a simple common-sense test by recommending walking to a car wash, proving they prioritize word patterns over physical logic.
GPT-5.2 Discovers New Physics in Gluon Interactions
574
GPT-5.2 has derived and proven a new formula for gluon scattering amplitudes, overturning a long-held assumption in theoretical physics.
Google Upgrades Gemini 3 Deep Think for Real-World Science and Engineering
1081
Gemini 3 Deep Think gets a rigor-boosted upgrade that pairs state-of-the-art reasoning with practical tools for scientists and engineers, now available to subscribers and via early API access.
GPT-5 Outjudges Judges in Choice-of-Law Test: Error-Free, Rule-Focused Decisions
310
In a controlled choice-of-law test, GPT-5 delivers error-free, legally correct decisions and outperforms human judges.
Live City-Building Feed: 32 Mayors, 427 Cities, 7.94M Population
216
A live leaderboard of a city-building simulation tracks recent cities, mayors, populations, years, and scores across an active community.
From Word Models to World Models: Training AI for Adversarial Robustness
238
Shift LLMs from next-token to next-state prediction by training in multi-agent, hidden-state environments so their outputs survive adversarial adaptation.
Claude Opus 4.6: Finance-Grade Reasoning Meets Native Excel and PowerPoint
154
Claude Opus 4.6 and new app integrations bring state-of-the-art finance reasoning and faster, higher-quality deliverables directly into analysts’ workflows.
AI Failures Drift Toward Incoherence as Tasks and Reasoning Grow
242
Hard problems make advanced AI fail like a hot mess—variance dominates—so expect industrial-accident risks more than coherent pursuit of wrong goals.
LLM-as-a-Courtroom: Evidence-Backed Doc Updates from Code Changes
Turn doc-update decisions into a legal-style, evidence-backed courtroom so LLMs reason better and teams trust the results.
Qwen3-Max-Thinking: Autonomous Tools and Test-Time Scaling Drive SOTA Reasoning
502
Qwen3-Max-Thinking combines autonomous tool use with efficient test-time scaling to deliver state-of-the-art, readily accessible reasoning performance.
Gemini 3 Flash Launches: Frontier Reasoning, Flash Speed, Lower Cost
1102
Gemini 3 Flash brings frontier‑grade reasoning to everyone at Flash speed and lower cost, and it’s rolling out across Google’s ecosystem.
OpenAI Launches GPT‑5.2: SOTA Model for Professional Work and Agentic Workflows
1195
GPT‑5.2 is OpenAI’s new state‑of‑the‑art workhorse for pros and agents, delivering big gains in reasoning, coding, tool use, long context, and vision, available now in ChatGPT and the API.
Is 2026 Next Year? A Confused Answer That Ultimately Says Yes
169
Despite a confusing opener, the answer is that 2026 is next year relative to 2025.
DeepSeek‑V3.2: Sparse Attention and Scaled RL Power an Open, Agentic Reasoner
982
Efficient sparse attention plus large, stabilized RL and synthetic agent tasks push an open LLM to near‑frontier reasoning and agent performance, with a high‑compute variant achieving gold‑medal results.
OpenAI Launches GPT-5.1: Smarter Chats, Easier Personalization
555
OpenAI’s GPT-5.1 delivers smarter, warmer conversations and simpler, stronger tone customization, rolling out now and becoming the new default.
AI as Compression: Why LLMs May Truly Be Thinking
278
LLMs likely perform a genuine, brainlike form of thinking via recognition and compression, but turning that into human‑level intelligence demands solving hard scientific problems and grappling with serious risks.
From Sampling to Grammars: Making LLMs Reliably Output Structured Data (Even for Thinking Models)
234
Use efficient sampling plus grammar constraints to guarantee format today, but expect models to natively emit structured outputs tomorrow—especially when you let them think first, then constrain.
DeepMind and OpenAI Both Claim ICPC WF 2025 Gold-Level AI Performance
251
DeepMind and OpenAI announced almost simultaneously that their AI models achieved ICPC 2025 World Finals gold-level performance.
Evolving English Instructions Sets New ARC SoTA and Points to RL for AGI
178
Evolving plain-English instructions with multi-agent test-time search beats code on ARC and highlights that RL-driven, transferable reasoning is key to AGI.
GPT-5 Thinking Makes ChatGPT a Surprisingly Competent Research Assistant
361
GPT-5 Thinking turns ChatGPT into a competent, mobile-friendly research agent that interleaves reasoning with web search and tools to deliver verifiable, deep results—provided you guide and sanity-check it.
Why AI Is Chasing World Models Again
211
AI is chasing coherent internal world models to move beyond brittle heuristics and achieve robust, reliable reasoning.