Last week in AI: Kimi K3, open source software, and compute constraints

Hey there,

"Open" dominated the headlines this week, let's unpack it:

1. ​Kimi K3 launched: Moonshot AI released Kimi K3, a soon-to-be-open-weight model that appears to be unusually good at software development. My X feed quickly filled with impressive UI demos the model had one-shotted. It’s great at UI. It also appears great at cybersecurity. It is not great at sales — but more on that below. Moonshot says the full model weights will be released by July 27, 2026.
2. ​Reflect Open: The popular note-taking app upended its product roadmap to work better with coding agents, open-sourcing the product and loosening its previous stance on encryption. The gravitational pull of agents is reshaping the rest of software in their image.
3. ​Kimi pauses new subscriptions: Demand for Kimi K3 overwhelmed Moonshot AI’s available capacity. That is another reminder that we are still deep in a global compute shortage, but I would not read too much into the subscription pause itself.
   ​
   AI labs spend roughly one-third of their compute on pretraining, one-third on reinforcement learning, and one-third on inference — a heuristic I first heard from Reiner Pope on Dwarkesh. Under that allocation, a 50% increase in inference demand would require a 25% cut to the compute available for pretraining/RL.
   ​
   For a lab releasing an open-weight model, limiting subscriptions and waiting for outside inference providers to come online is probably a better trade than slowing work on the next model.

Two quick updates from me:

• ​Kimi K3 isn't great at sales: K3 ranked 32 out of 49 model configs. It's not cheap either at 15 cents per call. The current configuration available on the Moonshot API is "max", so this model really ate up tokens. I think this means OpenAI and Anthropic are still in the lead when it comes to "general" intelligence, though coding might become a commodity quicker than the big three AI labs were counting on.
• ​A teardown of the Cerebras knowledge base: I read the Cerebras post and was immediately impressed. They solved a lot of hard problems elegantly. Read the original post. We've run into a lot of these building ours at Vercel. Will write up a full post, but here's 6 bullet points that cover the TL;DR.

As for next week, it's packed full of tech earnings announcements (Tesla, Alphabet, Intel, etc.). The X timeline is unsure how to react to the Kimi news, so I expect some bumps driven by a DeepSeek-like moment as investors debate terminal values of the big labs.
​
And last but not least: you'll get to keep Fable access (as long as you upgrade to a Max plan).

LFG,

Drew