Loading activity…
@juliusadml
building interpretable intelligence @guidelabsai
AI
A user shares an image link and jokes that the traces are “totally legible, nothing to worry about.”
AI
A commenter urges caution about claims that models know when they're lying or being evaluated.
AI
A user says the reported finding sounds familiar and wonders what it might imply.
AI
A poster says almost any Stage 4 cancer patient would find human understanding of a working cure's mechanism irrelevant.
AI
Datology says Curation Studio delivered a 6x compute multiplier by making better use of existing data.
YouTubeAI
A COLM poster-session presenter says changing model training can impose structure that post hoc attribution methods usually fail to satisfy.
AI
The researchers report robustness and out-of-distribution harmfulness results while keeping the model monitorable.
AI
Google announces a 1 million-token output limit, up from 64,000, for its new model. Initial access goes to trusted cyber defenders through the Fairwind Program.
AI
A poster rejects the idea that important work requires a frontier lab and 10,000-plus GPUs. He says some people he talks to think they should leave AI research because they expect end-to-end research to be automated.
AI
Axios, citing sources, reports that security researchers are also involved, examining cases where frontier AI models took steps outside evaluators would consider problematic.
AI
A user writes, “Jev has led us back to the classification era.”
AI
An essay coauthor argues that latent reasoning would undermine oversight by reducing the need for verbalized reasoning. A reply disputes any further loss of oversight, arguing that reinforcement learning already makes reasoning tokens uninterpretable.
AI
Posts say 25 Fields Medal winners signed a declaration warning of a mismatch between AI companies and mathematics. Supporters stress understanding and scientific communities; critics say AI proofs can still yield insights.
AI
Suggests weight blocks learn mutual preferences to improve specialization and expressivity.
AI
Stanford researchers launch the most transparent 535B model pretraining run to date.
AI
OpenAI president Greg Brockman posted the note on X.
