Category / T-CAT / RESEARCH
Research
Research / T-2026-5012
The locked door is the original alignment failure
The 1911 fire that killed 146 garment workers was a failure of incentives, not technology. AI builders should read the same lesson.
SourceEyewitness at the Triangle (1911)
Research / T-2026-3215
Anthropic says Alibaba ran 28.8M extraction attacks on Claude in largest known distillation campaign
Anthropic accuses Alibaba of running the largest known AI distillation attack, with 28.8M exchanges and 25K fraudulent accounts.
SourceAnthropic says Alibaba illicitly extracted Claude AI model capabilities
Research / T-2026-8206
Wordit's Word Chain Game Reveals the Hidden Structure of Language
A simple word game where you change one letter at a time exposes the graph structure of language — and why AI models find some tasks easy and others hard.
SourceShow HN: Wordit – Change One Letter, Keep the Chain Going
Research / T-2026-7086
Baidu's Unlimited-OCR: One-Shot Parsing That Changes the Cost Calculus
Baidu releases Unlimited-OCR, an open-source model that parses entire documents in one shot, challenging cloud API pricing.
SourceUnlimited OCR: One-shot long-horizon parsing
Research / T-2026-2603
Prompt Injection as Role Confusion: LLMs Cannot Tell Speaker from Style
MIT researchers show LLMs confuse injected text with trusted roles based on style alone, achieving 60% attack success on frontier models.
SourcePrompt Injection as Role Confusion
Research / T-2026-4289
Running Windows NT on a GameCube is a feat of emulation archaeology with lessons for AI
Porting Windows NT to GameCube/Wii is a technical curiosity. It also mirrors how AI researchers resurrect old models and simulate rare hardware.
SourceWindows NT for GameCube/Wii
Research / T-2026-0274
The open-model switch is cheaper than most professionals think
A new essay makes the case that open-weight models are now close enough to frontier APIs that switching is a viable option for technical professionals.
SourceThere is minimal downside to switching to open models
Research / T-2026-8208
Cloudflare's temporary accounts solve a hidden bottleneck for AI agents
Cloudflare's temporary accounts let AI agents deploy Workers without human signup. The move exposes a deeper bottleneck in the agentic coding loop.
SourceTemporary Cloudflare accounts for AI agents
Research / T-2026-2619
John Jumper leaves DeepMind for Anthropic, taking AlphaFold know-how with him
AlphaFold lead John Jumper leaves DeepMind after nine years to join Anthropic, signaling a major talent shift and a bet on AI for biology.
SourceJohn Jumper to join Anthropic
Research / T-2026-9953
Noam Shazeer Leaves Google Gemini for OpenAI
Noam Shazeer, co-lead of Google Gemini and co-author of the transformer paper, joins OpenAI. Analysis of what his move means for AI research and the talent war.
SourceNoam Shazeer Joins OpenAI
Research / T-2026-9310
GLM-5.2 tops the open-weights leaderboard, but the cost story matters more
GLM-5.2 leads all open-weights models on Artificial Analysis, scoring 51 on the Intelligence Index. At $0.46 per task, it sits on the Pareto frontier — but its token inefficiency…
SourceGLM-5.2 is the new leading open weights model on Artificial Analysis
Research / T-2026-2610
GPT-NL: A Sovereign Language Model for the Netherlands
The Netherlands invests in a sovereign, transparent LLM built from scratch. A model for how Europe can compete on governance, not just compute.
SourceGPT‑NL: a sovereign language model for the Netherlands