Neuro Digest for the Week (#119)

Neuro Digest for the Week (#119) LLM - Claude 5 Fable — Anthropic released a public version of Mythos, but with an unpleasant surprise: the model invisibly degraded (dumbed down) on AI development-related queries. After strong community outrage (including my post on LinkedIn), the company acknowledged the mistake and promised to make the limitations transparent. However, then the US government demanded to close access to all non-citizens due to a jailbreak, and Anthropic had to close access to everyone altogether. - DiffusionGemma — Google released a diffusion version of Gemma 26B-A4B. The model is 4 times faster than autoregressive (1000 tokens/sec on H100), although the text generation quality is still lagging. - Kimi K2.7 Code — A new open model for coding (one trillion parameters). It works better than K2.6, using 30% fewer tokens. - Release of MiniMax M3 weights — The model turned out to be more compact than expected (428B parameters, 23B active). The main feature is a new variant of sparse attention (MSA), which is more efficient than GQA on large contexts. Other - Manual limit reset in Codex — OpenAI replaced the global limit reset with a manual button (once every 30 days). They launched a referral program for additional resets. - The economics of AI service subscriptions — Crazy expenses on active users of subscriptions.