IBM Mixture of Experts on Jev

Neither, exactly. The entry copy says the segment weighs the vendor's speed and cost claims against TypeSafe's own caveats about how the benchmarks were built. That is scrutiny of the methodology, but it comes from show notes rather than a transcript.

The entry copy does not record a verdict, and its caveat says these are show notes, not a transcript of the conclusion. What the notes do record is the tradeoff: giving up text generation for fast structured decisions with calibrated probabilities.

A minority. The Jev segment is chaptered in the entry copy from roughly 11 to 29 minutes of a much longer episode whose other topics are separate model releases and a lunar model. Skip accordingly if you only care about decisions.

The panel's own, and IBM states the opinions belong to the participants rather than the company. That is useful for hearing skepticism, but it is the opinion of people who did not build it, with no measurements attached.

The episode is hosted by Tim Hwang with co-host David Zax, and the three guests are Kaoutar El Maghraoui, Principal Research Scientist for AI Native Systems; Gabe Goodhart, Chief Architect for AI Foundations; and Martin Keen, Master Inventor. The page is explicit that the opinions are the participants' own and do not reflect IBM.

The episode notes describe System One models as giving up the florid text generation in exchange for fast, structured decisions with calibrated probabilities. The pitch as stated on the page is frontier-level judgment on decision tasks with responses in hundreds of milliseconds, and the segment weighs the speed and cost claims against TypeSafe's own caveats about how its benchmarks were built.

The surrounding episode is a price-war roundup and a lunar model release. Anthropic shipped Claude Opus 5.5 at 20 percent cheaper than earlier Opus models, OpenAI launched GPT-6 Sol and Luna with Luna at half its predecessor's cost and ten cents per million input tokens, and NASA and IBM released an open-source Lunar Foundation Model trained on roughly two million image tiles from Lunar Reconnaissance Orbiter and other missions.