Published some thoughts on how we may be able to unlock “AI coworkers”. This will be a new class of AI native product across domains where precision is of paramount importance and correctness can be quantified. I'm excited about the startups building towards this future, unlocking digital employees for enterprises writ large!

@aashaysanghvi_ As a former competitive player, I’m bummed I don’t know about this group

@xprunie I’ve been getting glimpses of it recently.

RT @vipulved: Rolling out a new inference stack for DeepSeek R1 @togethercompute that gets up to 110 t/s on the 671B parameter model! https…

RT @mamoonha: Glean’s growth is a clear signal of deep, sustained adoption in the workplace across industries. An intense focus on making c…

RT @jainarvind: I’m extremely proud to share that @glean surpassed $100 million in ARR in our last fiscal year in just over three years. I…

@santa_kaus @OpenAI 💪🏽

Feeling the AGI

Networking stocks took a beating from Deepseek, but people don’t appreciate that large scale inference needs much better front end networks!

Late to the party, but on cloud sneakers are the best.

RT @swyx: i am once again telling everyone who will listen that @Joubinmir’s podcast is the single best operator pod in the world delete Y…

RT @bernhardsson: Went to a founder dinner and half the people there were ex-Palantir. What kind of founder crack do they put in the water…

1. Feels like we need LLMs to get good at propositional calculus. Requires going beyond attention + RL. 2. Need to figure out the construction of concept graphs, where leaf nodes are facts / lemmas. Combined with 1, we should be able to do better than next token prediction in domains with incomplete verifiers.

RT @pk_iv: 2024 was the year of obsession. I only had one goal: make @browserbasehq the best choice for developers using AI to browse the w…

@shomikghosh21 Would add $NU as a candidate as well.

Bunch of extra work required to get o1 to work with tools:

Pricing agents that automate work is an interesting open question that will see more answers in 2025. You can: 1. Price at a discount to wage spend on an equivalent human. 2. Price per “outcome”. 3. Introduce “credits” that buy you a set amount of time / work. Among the factors in the input space, distribution of tokens burned to conduct work will be most interesting to observe per domain. The more variable the scope of the agent, the more variable the distribution -> which affects costs -> which affects pricing.

What a wild match. Ding blundered a dead drawn position with Rf2, and that's all she wrote. The youngest champ in history.