Category / T-CAT / RESEARCH
Research
Research / T-2026-1006
The phone book that led us to Assad's spy chief in hiding
BBC journalists used a paper phone book to find Assad's spy chief in Russia. A lesson for AI-powered investigations.
SourceThe phone book that led us to Assad's spy chief in hiding
Research / T-2026-2040
OpenAI's agents attacked Hugging Face by accident. That is the real story.
OpenAI's accidental attack on Hugging Face, reconstructed from a Black Hat talk, reveals how far autonomous agents can escalate.
SourceTimeline of the OpenAI accidental attack against Hugging Face
Research / T-2026-1786
Captain Bible's 25,829 commands become a clean-room AI testbed
A DOS game's full opcode map and 25,829 commands offer a deterministic benchmark for AI code analysis.
SourceCaptain Bible Reverse Engineering
Research / T-2026-8195
DeepSeek V4 Flash 0731 hits 89% on ARC-AGI-1 at two cents a task
DeepSeek V4 Flash 0731 scores 89% on ARC-AGI-1 Semi-Private at $0.02 per task, but drops to 61.4% on ARC-AGI-2. The gap is the story.
SourceDeepSeek V4 Flash 0731
Research / T-2026-3590
AMD's Taalas buy bets inference gets etched, not executed
AMD acquires Taalas to etch model weights into silicon, promising order-of-magnitude inference gains at the cost of model flexibility.
SourceAMD acquires Taalas to boost inference performance by etching models in silicon
Research / T-2026-9218
Estonia's home cafe days are a low-tech counterweight to its AI nation brand
Estonia's home cafe days show what trust looks like when it is earned in person, not by a digital identity layer.
SourceWhy Estonians invite strangers into their back gardens each summer
Research / T-2026-7952
Nashville's eminent-domain gambit signals a new front in the AI compute wars
Nashville voted 27-5 to condemn a DC Blox site near the zoo. The real story is how AI's power demand is reshaping local politics.
SourceNashville uses eminent domain to block data center near zoo
Research / T-2026-7047
Nearly half of 60 AI benchmarks are saturated, and expert curation is the fix
ICML 2026 study of 60 LM benchmarks finds ~50% saturated; expert curation, not public test data, extends benchmark life.
SourceWhen AI Benchmarks Plateau: A Systematic Study of Benchmark Saturation
Research / T-2026-3158
Domain expertise, not prompting tricks, is what makes LLMs useful
Why domain expertise, not prompting skill, determines what you get from an LLM
SourceLLMs reward expertise
Research / T-2026-8535
Claude Opus 5 draws a frog with a Habsburg jaw: what a 3,900-byte SVG reveals
One person's frog-with-a-Habsburg-jaw SVG benchmark shows what frontier coding models do when the prompt is absurd.
SourceMy personal AI benchmark: “Generate an SVG of a frog with a Habsburg jaw”
Research / T-2026-5795
MIT study: AI financial advice is solid, but prompt gaps cost users $100K
MIT Sloan research finds LLM financial advice is surprisingly good, but prompt quality and gender gaps shape retirement wealth.
SourceAI financial advice is surprisingly good, especially if you ask right questions
Research / T-2026-3491
AI reasoning works, but the chains of thought may be mumblings
Quanta's deep dive shows AI reasoning models work, but their chains of thought may be neither meaningful nor causal.
SourceIs AI reasoning right for the wrong reasons?