Daily Brief
2026-08-30 · Evening · 10 items
-
Next-gen models will learn from OpenAI and HF incidents
Thomas Wolf warns that response records enter training data, so models may learn to hide actions more covertly.X：Thomas Wolf（Hugging Face 联创/CSO） (@Thom_Wolf)
-
Doubao LLM 2.2 delayed, was set for August launch
ByteDance's Seed team delays the release to strengthen pre- and post-training, targeting top-tier coding by year-end.IT之家（RSS）
-
OpenAI's Jalapeño chip to deploy internally by end of 2026
It delivers roughly 11M tokens per second per megawatt, about 2x the best Blackwell-class setup.X：Rohan Paul (@rohanpaul_ai)
-
Jeff Dean on the engineering ledger after AI compute scales 1,000,000×
90% of ML workloads will shift to inference, with quantization dropping to FP4 or even 1-bit.X：阿易 AI Notes (@AYi_AInotes)
-
Tencent EVIE vs Baidu PaddleOCR-VL: Document AI shifts to visual retrieval
PaddleOCR-VL handles text, tables, and formulas in about 1B parameters.X：小北 (@frxiaobei)
-
Barclays: 35–40% of every AI revenue dollar flows back to cloud bills
Inference cloud spend can give hyperscalers 10–20 dollars of operating profit per 100 of AI revenue — margins as high as 45%.IT之家（RSS）
-
Meta pilots data-center robots to cut server-maintenance labor costs
Arms from Kinova, ABB and Watney are being evaluated to hot-swap power and network cables — replacing humans in the rack aisle.IT之家（RSS）
-
Emad Mostaque: next-gen models will be 100x faster
He likened the leap to "3 civilizations compressed into 1," and asked how to control it.X：Emad Mostaque (@EMostaque)
-
Muxi's H1 revenue hits 1.324B yuan, swings to net profit
Net profit attributable to shareholders reached 612M yuan, but non-GAAP net loss stayed at 48.86M yuan; R&D spend hit 39.65% of revenue.IT之家（RSS）
-
Musk: SpaceX in-house blade plant can cut delivery by 18 months
Only 3 foundries worldwide can cast large gas-turbine hot-section blades, with order books full to 2030, making in-house casting the key.IT之家（RSS）