Research & Papers
-
Old CNN beats GPT-6 and Claude at skin cancer ID
Research & Papers Specialized fine-tuned models still outperform general LLMs on high-stakes clinical image tasks.
-
12 AI Models Share the Same 'Controversial' Opinion About Humans
Research & Papers Convergent outputs across competing AI systems raise questions about shared training data bias.
-
Dev trains 360M model to simulate a terminal
Research & Papers Shows world-model techniques scaling down to text UIs with cheap, self-generated training data.
-
SlopCodeBench Oct 2026: Astra leads at 16.3%
Research & Papers Shows frontier coding models still fail over 83% of strict real-world code tasks.
-
Researcher Seeds Web With 50K Fake Stats, Gets Cited by ChatGPT
Research & Papers Shows LLMs actively amplify misinformation when trained on stat-heavy SEO bait.
-
HybridAgents: Latent Thinking, Text Acting
Research & Papers Merging latent reasoning with tool-use could redefine efficiency benchmarks for agentic AI.
-
Deriving Transformers from Associative Memory
Research & Papers Shows attention has deep theoretical roots, which matters for principled architecture research.
-
Claude cracks NIST curve seed mystery in 30 min
Research & Papers Resolves a foundational trust question about one of the world's most-used crypto standards.
-
AI benchmarks miss 90% of US work hours
Research & Papers Without broad benchmark coverage, capability claims about AI in real work remain largely untestable.
-
Goodfire: Neural Networks Are Already Clean
Research & Papers If correct, this reframes the entire sparse autoencoder research agenda in AI interpretability.
-
TimeTeller's saliva test maps your circadian clock
Research & Papers Circadian-based medicine could change how cancer therapy is dosed and scheduled.