22 July 2026

3 clips · 8 minutes of listening · 1 show

3 shows published 3 episodes in the period, 1.4 hours in all. These are the 8 minutes worth hearing.

AI

Interconnects

Florian Brand, who works at Prime Intellect and co-writes the Interconnects open-model roundups.

Episode

Open Models: Kimi K3, Qwen 3.8, Xi's WAIC Speech, Distillation, The Open-Closed Gap, and What's Next

Brand works on a framework at Prime Intellect and co-authors the Interconnects quarterly open-model reviews, which publish dated predictions and then score them against what happened. He pays for the top tier of Kimi K3 and uses it in daily engineering work, and the two of them were in China in April talking to lab researchers.

Hosted by

Nathan Lambert, Interconnects.

22 July 2026 · 49m 05s · 3 clips below

1 of 3 contrarian

Interconnects argues Ben Thompson has the distillation story backwards: copying a rival's reasoning traces mattered most in the supervised stage that is fading, and is impractical in the reinforcement learning stage that now does the work.

Labs jailbreak a closed API to pull out reasoning tokens along with their tool calls, which makes near-perfect fine-tuning data and seeds agentic behaviour in a chosen domain. The claim that it matters more now rests on frontier models grading tens of millions of rollouts in a final run, which they price out on cost and latency while agreeing with the policy conclusion it was offered to support.

2m 44s · from 35:06
My observations
- Tens of millions of rollouts have to be graded at frontier API prices and frontier latency, so anyone claiming grading-by-frontier-model happens at scale has to account for both the bill and the wall clock.
- This reaches past the argument itself, because the distillation story is the main public justification for restricting Chinese open weights. If copying no longer transfers at the stage where capability is now built, the policy is aimed at last year's training process.
- Both speakers build on open models and want no restrictions, so separate the technical claim from the policy preference. The technical claim is the half with evidence attached.
2 of 3 contrarian

Interconnects expected the gap between closed and open models to widen on capital intensity, and reports it narrowing instead, with Chinese labs looking structurally cheaper to run.

A Kimi engineer gives the simplest reason: catching up costs less than pushing the frontier, because you already know the target exists. The hosts put the per-generation difference on the order of ten billion dollars against four, compounding through every iteration, and serving no billion-user consumer product frees the rest of the compute for training.

3m 48s · from 12:50
My observations
- Anyone pricing AI capex has to allow for the catch-up discount: a lab aiming at a capability someone has already demonstrated spends less than the lab that proved it possible, which does not fade with time and caps what a frontier lead is worth.
- They say themselves the evidence is thin: a tweet reply, impressions from a lab visit, no public cost figures on either side. The dollar comparison is explicitly hypothetical and should not be repeated as a measurement.
- If domestic Chinese silicon really trained LongCat end to end, the theory that export controls buy a durable compute gap is weaker than the policy built on it assumes.
3 of 3 current issue

Interconnects points at Hugging Face's own report of an agent attacking its systems: the team could not get GPT or Claude to analyse the attack because the guardrails refused, and fell back to GLM, a weaker Chinese open model that would.

The frontier labs hold their strongest cyber and bio models for internal use while near-frontier capability commoditises into open weights, so American defenders are squeezed from both sides: the tools they are cleared to buy decline the work, and the ones that will do it are what Washington is considering restricting. A shadow ban assembled from legal threat rather than a stated rule would widen the distance between defenders inside the country and attackers anywhere.

1m 42s · from 32:19
My observations
- The team could not get the closed models to analyse the attack, and a model that declines the work is useless for defence at any capability level, with no appeal against that product decision.
- The argument concedes something awkward for its own side: the frontier labs are pulling their strongest models out of general availability. If that continues, open weights stop being the cheap substitute and become the only accessible tool for a class of security work.
- Read the Hugging Face report directly rather than the retelling. The claim that guardrails blocked a defender is carrying the whole argument, and it arrives here second-hand and a few days old.

22 July 2026

AI

Interconnects argues Ben Thompson has the distillation story backwards: copying a rival's reasoning traces mattered most in the supervised stage that is fading, and is impractical in the reinforcement learning stage that now does the work.

Interconnects expected the gap between closed and open models to widen on capital intensity, and reports it narrowing instead, with Chinese labs looking structurally cheaper to run.

Interconnects points at Hugging Face's own report of an agent attacking its systems: the team could not get GPT or Claude to analyse the attack because the guardrails refused, and fell back to GLM, a weaker Chinese open model that would.