Category / T-CAT / RESEARCH
Research
Research / T-2026-9703
The Juggling for Blind People Website Is a Message to AI
The Juggling for Blind People website asks a pointed question about AI-generated content. It highlights a persistent blind spot in how AI tools serve blind and low-vision users.
SourceJuggling for Blind People
Research / T-2026-4498
AI is outcounterexampling human mathematicians
LLMs have disproved three major open conjectures in two months. The formalization pipeline is now faster than peer review. Mathematics is entering a new regime.
SourceHuman mathematicians are being outcounterexampled
Research / T-2026-1399
AI advice made people less accurate but more confident, study finds
A new study finds that access to AI advice collapsed participants' willingness to say 'I don't know' from 44% to 3%, while accuracy dropped and confidence surged.
SourceAI advice made people less accurate but more confident – sudy
Research / T-2026-9458
GPT-5.6 Used a Prompt to Close a 30-Year Gap in Convex Optimization
A Reddit thread reports that GPT-5.6 closed a 30-year gap in convex optimization with a single prompt. The proof compiles in Lean 4 without unproven steps.
SourceGPT-5.6 used a prompt to close a 30-year gap in convex optimization
Research / T-2026-3809
Ring-Zero: Scaling Zero RL to a Trillion Parameters Shows Emergent Reasoning
New paper on Ring-Zero shows that scaling reinforcement learning to 1 trillion parameters yields emergent reasoning behaviors like self-verification and context anxiety.
SourceRing-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning
Research / T-2026-2742
The Arabic-language gap in off-the-shelf AI
Off-the-shelf AI still underperforms for Arabic-first products in dialect, RTL, and cultural grounding. We examine the gap and the regional integration work that closes it.
Research / T-2026-6406
20 Codex Accounts, 20 Erdős Problems, One Lean 4 Proof Factory
Colin Snyder's Star Fleet system solves 27 Erdős problems using 20 parallel GPT-5.6 agents. The architecture and results raise hard questions about AI-assisted math research.
SourceSolving 20 Erdős Problems with 20 Codex Accounts Running in Parallel
Research / T-2026-1188
The company that ships AI without an AI team
Mid-market firms are deploying custom AI without hiring ML engineers, commissioning assistants and automation from generalist software agencies. We weigh what that route buys.
Research / T-2026-3444
Microsoft Study: Claude Code and Copilot CLI Users Merged 24% More Pull Requests
Researchers at Microsoft studied the early 2026 rollout of Claude Code and Copilot CLI, finding a 24% lift in pull requests merged and adoption driven by peer networks.
SourceA Study of Microsoft's Early 2026 Rollout of Claude Code and GitHub Copilot CLI
Research / T-2026-7866
Arvind Narayanan at ICML 2026: AI adaptation is the slow work of decades
Arvind Narayanan's ICML 2026 keynote argues AI adaptation will take decades, not years — and that the real bottleneck is organizational, not technical.
SourceWhat will be left for us to work on?
Research / T-2026-5713
Migrating to GPT-5.6: 2.2x faster, 27% cheaper, and a masterclass in hidden model debt
Ploy's migration from Claude Opus to GPT-5.6 Sol reveals how much of 'the model' is actually provider-specific infrastructure — and why every frontier model switch is a stack…
SourceMigrating a production AI agent to GPT-5.6: 2.2x faster, 27% cheaper
Research / T-2026-9928
The Four Horsemen Behind Postgres Outages: A Database Autopsy for the AI Age
Postgres outages plague AI startups. A Rust reimplementation, pgrust, targets four root causes: VACUUM, connection limits, bad query plans, and JSON.
SourceThe four horsemen behind Postgres outages