Anthropic Resignation Sets Off a Fight Over Doom Claims

@filicroval published a revised agent harness tierlist after playing more with the DeepSeek harness, calling it "so good" and useful beyond DeepSeek models. His earlier list put Claude Code, Codex, Hermes Agent and Oh My Pi in S+. @MiaAI_lab replied: "So many harnesses."

@MiaAI_lab calls GLM 5.3 Flash EXL3 "THE model to run if you own two DGX Sparks" and links the release repo. @filicroval reports 36 tok/s for 2xSpark owners. In the replies, @jodyalbritton says he benchmarked GLM against DeepSeek and Qwen and now runs GLM as his daily driver.

@RDriscollCPA shares a Q&A with a bookkeeping firm owner: one large accounting firm was told by a client to expect 20% off for AI use, and the Financial Times reported companies pressing law firms on AI use in RFPs. His advice: deliver more, such as product ideas mined from client call transcripts, and lean on trust. His real worry is clients leaving to do the books themselves with AI.

@rohanpaul_ai reports the NSA, CISA and FBI published joint advisory AA26-251A alleging six Chinese AI firms ran industrial-scale distillation against American frontier models since at least late 2024. The advisory says Moonshot AI distilled 17 American models, including Anthropic's Claude Fable 5, to train Kimi-K3, and that requests moved through API proxies called transfer stations. Its mitigation section asks American labs to serve suspected distillers subtly degraded responses without telling them.

@rohanpaul_ai posts Jensen Huang's answer when Axios asked whether open source model companies should be allowed to distill closed models: distillation, "learning from AI, learning from other people, and learning from other sources of knowledge, is fundamental to intelligence," he says in the post. Huang predicts the internet could be 99% AI-generated content within a few years and argues a smarter AI can also be a safer AI.

@thsottiaux writes that demand for Astra is "really unprecedented" and that the team is "pulling all the levers possible" to sustain it, warning that new Pro subscriptions might pause "for a bit" if the surge continues. In the replies, @IamEmily2050 says the $200 Pro subscription "doesn't even last one day" and urges OpenAI to distill Astra into Luna. @Youssofal_ reports disconnects and halved TPS.

@mark_k writes that doomers treat the data center backlash as an opening to sell AI control, calling data centers "merely the pressure point." @Dan_Jeffries1 says he resents proposals for chip control, bans, or tracking researchers, and recalls Hinton's "we don't need to train Radiologists anymore" call. @JordanSchachtel adds that a slowdown deal fails while Beijing "doesn't have neurotic doomers running their AI labs."