Summary
The slot is dominated by one thing: Hugging Face's Elie Bakouch live-tweeting a ~47-post teardown of Microsoft AI's new MAI tech report, which we already have a wiki page on (the MAI Thinking-1 hill-climbing model trained with zero synthetic data and zero distillation from prior models). His notes are the strongest signal here, full of concrete RL recipe details: a GRPO variant with length penalty, entropy-based outer clipping, no KL term, and global normalization, plus infra disclosures like SGLang serving and 40% higher throughput per Watt on Microsoft's own chips. On the model-release side, Kilo flags an open-weight surge (StepFun Step 3.7 Flash and MiniMax M3 both shipped with open weights this week). Anthropic's Logan Graham posts a sober note on cyber safeguards being urgent within a 3-to-18-month window. The rest is noise: an AI-news-aggregator plug, an Anthropic-IPO prediction-market quip, an Affirm interview promo, and an NVIDIA GTC skyline shot.
Posts
- Microsoft MAI tech report teardown (cluster of 17) (@eliebakouch · report PDF). Bakouch calls it one of the most transparent reports at this scale and walks through the post-training stack: no cold start (zero synthetic data), difficulty filtering via two-stage pass rate, top-p masking like DeepSeek, and a GRPO variant with length penalty, entropy-based outer clip (holding entropy near 0.3), no KL term, and global rather than per-response normalization. He flags rounds of RL then self-distillation SFT to recover, attributed to train/inference mismatch and reduced once they switch to bf16. Infra notes: SGLang serving, good final-run numbers, and 40% higher throughput per Watt as bullish for Microsoft's chips. Critiques: unsure why they use SFT over on-policy distillation in consolidation, and the model/perf co-design seems not to weigh inference much. See MAI Thinking-1 hill climbing.
- Open-weight model wars heat up (@kilocode). StepFun dropped Step 3.7 Flash and MiniMax dropped M3, both open weights and both already live in Kilo. Two frontier-ish open releases in one week is a real signal worth tracking.
- Cyber safeguards are urgent (@logangraham). Anthropic's Graham frames cyber defense as a hard techno-philosophical problem with powerful unsafeguarded models possibly arriving in 3 to 18 months, arguing for urgently scaling access to defensive tools.
- Aligned News AI aggregator (@Scobleizer · site). Scoble plugs his agent-curated AI news feed and the time sink of training it. Click through to read.
- Max Levchin on Affirm and AI (@mlevchin). Yahoo Finance interview on payments, AI, and engineering-led companies. Promo. Skip.
- Anthropic IPO prediction-market quip (@brivael). Reposts a Polymarket note that Anthropic is no longer projected to IPO by end of September, with an editorial jab. Low signal.
- NVIDIA GTC skyline (@nvidia). Taipei 101 lit up green for #NVIDIAGTC. Promo. Skip.