Weekly YouTube Digest - Jul 26-Aug 1, 2026

This week's fast take

GPT-5.6 just made itself better...

• Berman says GPT-5.6 Luna received an 80% price cut, while GPT-5.6 Terra received a 20% cut; he says GPT-5.6 Soul instead gained a fast mode that runs 2.5 times faster for twice the price 1.
• His main accounting rule is cost per completed task, not price per token: a cheaper model that uses twice as many tokens may cost the same in practice 1.
• The video attributes the improvement to GPT-5.6 Soul helping find production GPU-kernel changes and better speculative decoding; Berman repeats figures of 20% lower serving costs and 15% better token-generation efficiency 1.
• Using the Artificial Analysis intelligence index, he says GPT-5.6 Luna Max is slightly ahead of GLM 5.2 Max and costs about 6 cents per completed task, versus roughly 25–28 cents for GLM 5.2 Max 1.
• Berman labels the broader recursive-self-improvement story as partly his own thinking, including the speculation that large models may become the tool for making cheaper models that are harder to displace 1.

I'm disappointed

• Berman defines open source as making a model available to use, modify, inspect, and share, and argues that this reduces dependence on a small number of closed providers 2.
• His economic claim is that open-weight models push value toward chips, data centers, infrastructure, and applications instead of letting the model layer keep all the margin 2.
• He presents Kimi K3 as evidence for the open-model case, saying Chinese labs are producing models that can approach closed systems while putting pressure on their prices; that comparison is his claim, not a fresh test in the video 2.
• The safety case against open weights is stated clearly: safeguards can be removed, access cannot be revoked, misuse becomes easier, and responsibility is harder to assign 2.
• Berman's conclusion is that open-weight AI should not be banned, while also acknowledging that Anthropic's later letter did not call for a blanket ban; he remains critical of its broader stance 2.

Kimi K3 Just Broke The Economics Of AI

• The speaker describes Kimi K3 as an open-weights model with 2.8 trillion parameters, free to download and own, while noting that running the full model locally is out of reach for most people 3.
• The video argues that cheaper API access could push token prices down even for people who never run Kimi K3, and that smaller distilled versions may make its capabilities easier to use 3.
• It introduces Kimi Delta Attention as a way to keep an updated memory rather than rereading every earlier item, with older information gradually fading 3.
• It describes attention residuals as preserving earlier versions of a representation across layers, so later layers can use the change history instead of only the latest state 3.
• The speaker says combining the two ideas improves scaling efficiency over Kimi K2, but the spoken multiplier is unclear in the transcript; do not turn that line into a precise benchmark without checking the paper 3.

Gemini Robotics 2 brings whole body intelligence to robots

• The video presents Gemini Robotics 2 as a generalist robotics model, meant to make one robot useful across many tasks instead of limiting it to a narrow pre-programmed sequence 4.
• The release is organized around three capabilities: whole-body control, dexterous manipulation, and multi-robot collaboration 4.
• The dexterity examples go beyond pick-and-place, including tasks such as screwing in a light bulb and handling a trash bag, where many joints must coordinate at once 4.
• For collaboration, the video says each robot runs its own copy of the same stack and the robots coordinate through their own reasoning while working on one task 4.
• The stated goal is generality: reacting when the scene changes, rather than replaying a fixed motion; the video does not provide a broad deployment evaluation 4.