10 November 2025

41 new episodes (27 hours listening) published by the 38 podcasts on the list. 2 highlight clips below (5 mins total).

AI 10 November 2025 · 2 clips

Interconnects

Richard Bian (product and growth lead), Chen Liang (algorithm engineer) and Ziqi Liu (research lead), of the Ant Ling team at Inclusion AI, Ant Group

Episode

Inclusion AI is Ant Group's model lab, begun in February 2025 and shipping a trillion-parameter model inside eight months, with Ling releases in April, July and September. The product lead spent 11 years in the United States at Microsoft and Square before returning to China during COVID; the research lead has been at Ant Group about eight years.

Hosted by

Nathan Lambert

12 November 2025 · 1h 17m · 2 clips below

1 of 2 explainer

Ant Group's Ling team took DeepSeek's blockwise FP8 training recipe, measured it running slower than bfloat16 in places, and traced the loss to quantisation and dequantisation overhead.

Quantising weights and inputs on the way into each matrix multiply and converting the result back costs enough to eat the gain the lower precision was meant to deliver. Their fix was to fuse the gating function with the quantisation step in the mixture-of-experts layer so the two run as one pass, which matters because that layer runs its activation across every expert.

3m 23s · from 25:51
My observations
- DeepSeek published the recipe and this team ran it as described, then had to profile the training and fuse operations itself before FP8 got ahead of the precision it was supposed to replace.
- Two labs ran the same published procedure and got different throughput, with the gap sitting in implementation rather than in hardware, which is worth holding against any compute forecast that treats a published efficiency gain as available to everyone. A board reading an efficiency argument inside a capex thesis is reading an assumption about who can execute.
- They profiled their own implementation on their own hardware, and nothing here shows the original numbers were wrong on the stack that produced them.
- compute
- ai-infrastructure
2 of 2 framework

Inclusion AI frames open weights as the strategy available to a lab that is behind, and notes that the leaders in this game are not the open ones.

The team reaches for poker: whoever is ahead on chips plays their own hand and that is understandable, while a player who is behind minimises mistakes by following a direction others have already proven. Later in the conversation they say there is no guarantee they would do the same if they were in front.

1m 20s · from 12:52
My observations
- Inclusion AI's openness follows from its position rather than its mission on this telling, which means the supply of open weights out of China can stop when the position changes. An organisation that has standardised on those weights is exposed to somebody else's competitive strategy.
- The team says the decision came from above its pay grade, and the same episode ends with an explicit call to work with them, so read the framing as a lab describing itself.
- Whether a Chinese lab that reaches the front keeps releasing weights would separate this explanation from the ideological one, and that test may not be far off.
- open-models
- ai-geopolitics

10 November 2025

AI 10 November 2025 · 2 clips

Ant Group's Ling team took DeepSeek's blockwise FP8 training recipe, measured it running slower than bfloat16 in places, and traced the loss to quantisation and dequantisation overhead.

Inclusion AI frames open weights as the strategy available to a lab that is behind, and notes that the leaders in this game are not the open ones.