From the source
Nathan Lambert and Florian Brand used an Interconnects podcast to weigh why Kimi K3.1 Pro drew a mixed reaction despite top benchmark scores.
- Brand says the release felt different from earlier Kimi drops, partly because the company ran a bigger social media campaign that allegedly paid people to post about the model without disclosing it.
- Lambert calls that bad form and notes one person tweeted a screenshot of press instructions telling recipients not to say it was a paid promotion, which he says no one has corroborated.
- Brand says the model creator gets a few weeks of exclusive access before releasing weights, and that K3.1 Pro has about two weeks of exclusivity, with weights promised by November 4th.
- Lambert points to an Artificial Analysis number placing it ahead of GPT-6 Astra, but says other results look jagged, including Terminal-Bench Science, where he says it trails GPT-6 Astra by 20 points.
- Brand attributes the gap to Artificial Analysis benchmarks being short-horizon while ProgramBench is long and explorative, and to Kimi's RL training leaning exploitative rather than explorative.