Sat down with anam hira this morning to talk about Revyl (YC F24) which is part of the Groq Inc Startup Program. QA is one of the most overlooked but critical parts of the software stack and integrating it effectively into the development process is currently a mess. Revyl

Prompt Caching is live on. Groq Inc! Pin your prompts once, reuse them as much as you want, and stop paying for the same tokens over and over.

The whitepaper was cool, but nothing screams future like a Bitcoin ATM humming next to a Slurpee machine. Finance is decentralized—all the way to a Circle K.

Let’s switch the lingo from VC (Venture Capitalist) to CA (Capital Allocator). Sounds way more dope.

Who said robots need to have two legs and two arms?

Frontier Labs continuing to build thinking vs non-thinking tells me two things: 1. Speed matters—people, or consumers more specially, don’t want a model to think for 99% of query’s. They want an answer quickly. 2. Inference cost remains an issue—thinking models use an order of

Regardless of how GPT-5 performs, we’ve shipped major quality improvements to GPT-OSS running on Groq Inc. If you experienced issues in the past, I’d encourage you to give it another shot. The improvements are huge.

To add a bit to what Philip Deutch said on The All-In Podcast a few weeks ago—according to the same @anthropic paper, training a single state-of-the-art AI model by 2028–2030 will require a 5 GW data center. That’s twice New York City’s peak electricity demand to train a single model. For

I'm hiring interns to work with me Groq Inc. You'll be doing: evals, infra, and/or post-training Term: Fall or Winter, full-time What we're looking for: - evals/post-training experience OR really solid CS/Applied AI fundamentals - thrives in high-intensity environments and

After a long few weeks of walking from San Francisco, I have completed my pilgrimage to the Mecca of M&A TBPN

Ideas fade if no one writes them down.

Does RL work on a messy roommate? How much compute do I need?