Tweets
2025-03-23T17:14:43.245499 | 🔗
Bill Gurley’s got it right on his theory of triple-counted revenue and VC and hyperscaler subsidized compute for AI wrappers. He doesn’t fall out companies. But I will call out all coding AI companies… Low barriers to entry and squeezed on both ends by hyperscaler and marketing BG2, e28, 1:15:00
2025-02-24T22:00:08.384489 | 🔗
To summarize... Elon's approach at Twitter and US Govt. is shocking to software engineers and federal employees because it's the first time they faced a ruthless PE buyout (essentially). It is not coincidental that this is coming at a time when knowledge work is being commoditized by AI.
2025-02-12T01:27:06.831345 | 🔗
"Identity Initializations" - controlnet and applied to whisper sidecar
2025-02-10T11:07:53.805033 | 🔗
Only one party appeals to common sense anymore and thats the party that doesn't make any sense.
2025-02-09T18:52:18.807990 | 🔗
Rgd. MAGA Indian American hate. I wonder if the Caste System gives a permission mechanism for this? Seems odd with Vivek, Nikki, and Usha so prominent. And non-R — Harris.
2025-01-30T23:39:35.563684 | 🔗
ai is analytic
2025-01-30T05:53:34.899686 | 🔗
welp the one quantitative benchmark i tested deepseek (70b distilled) vs llama 3.3 on, llama got a 55, deepseek got 33... higher is better.
2025-01-27T21:19:35.190637 | 🔗
Really like Karpathy's point rgd. RL. I most often feel the need to fight the idea that there are "emergent capabilities" in imitation learning, e.g. pretraining. However the difficulty is that getting good RL data requires either simulators or fast real world data labelling which poses a challenge for things like drug discovery. "Last thought. Not sure if this is obvious. There are two major types of learning, in both children and in deep learning. There is 1) imitation learning (watch and repeat, i.e. pretraining, supervised finetuning), and 2) trial-and-error learning (reinforcement learning). My favorite simple example is AlphaGo - 1) is learning by imitating expert players, 2) is reinforcement learning to win the game. Almost every single shocking result of deep learning, and the source of all *magic* is always 2. 2 is significantly significantly more powerful. 2 is what surprises you. 2 is when the paddle learns to hit the ball behind the blocks in Breakout. 2 is when AlphaGo beats even Lee Sedol. And 2 is the "aha moment" when the DeepSeek (or o1 etc.) discovers that it works well to re-evaluate your assumptions, backtrack, try something else, etc. It's the solving strategies you see this model use in its chain of thought. It's how it goes back and forth thinking to itself. These thoughts are *emergent* (!!!) and this is actually seriously incredible, impressive and new (as in publicly available and documented etc.). The model could never learn this with 1 (by imitation), because the cognition of the model and the cognition of the human labeler is different. The human would never know to correctly annotate these kinds of solving strategies and what they should even look like. They have to be discovered during reinforcement learning as empirically and statistically useful towards a final outcome."
2025-01-26T23:07:04.662904 | 🔗
An llms.txt for Modal https://gist.github.com/davidzqhuang/4a00758b42db0e3f57b50c7782c3e396