10^7-Dimensional LLM Memory, but Only If it Stays Sparse
A BDH seminar summary circulating in recent technical discussion frames LLM memory as a tradeoff between the familiar transformer KV…
-
-
Firefox Zero-Day: Mozilla Says Claude Mythos Found 271 Bugs
Mozilla said this week that its Firefox zero-day hardening work with an early version of Claude Mythos Preview helped identify…
-
A Formula From Another Field Opened Erdős Problem
Erdős problem #1196 now has a serious claimed solution, and the evidence ladder is unusually visible. Liam Price posted GPT-5.4…
-
A 14-Author Paper Tries to Make Deep Learning Theory a Science
A 14-author perspective paper posted to arXiv on April 23 argues that deep learning theory is starting to look less…
-
LLM Failure Modes Start in the Stack, Not the Chat
LLM failure modes are easiest to understand if you stop treating them as personality flaws, “the model lied,” “the chatbot…
-
66 Tokens Make a Diffusion Language Model Look Easy
A diffusion language model generates text by starting from masked or otherwise corrupted tokens and iteratively restoring them. In this…
-
Language-Agnostic Representations Show a Shared Semantic Workspace
The standard story is that LLMs work in words. They predict the next token, so surely their internal reasoning is…
-
The Abstraction Fallacy Makes Conscious AI Harder to Prove
Alexander Lerchner’s paper on conscious AI does something unusual: it does not start by asking whether today’s models seem conscious….
-
Claude Code Lost Its Thinking Budget: and Broke Workflows
A few weeks ago, one of the most useful facts about Claude Code stopped being visible. Not the code it…
