The Future is Global

DeepSeek Primer

For those who haven't realized it yet, Deepseek-R1 is better than claude 3.5 and better than OpenAI o1-pro, better than Gemini. It is simply smarter -- a lot less stupid, more careful, more astute, more aware, more meta-aware, etc. We know that Anthropic and OpenAI and Meta are panicking. They should be. The bar is a lot higher now. The justification for keeping the sauce secret just seems a lot more absurd. None of the top secret sauce that those companies have been hyping up is worth anything now that there is a superior open source model. Let that sink in.

DeepSeek is like buying the most expensive house in the neighborhood for 10 million and a guy next month buys similar house next door for 200K. Very few understand this means but they will soon feel it.

OpenAI was founded 10 years ago, has 4,500 employees, and has raised $6.6 billion in capital. DeepSeek was founded less than 2 years ago, has 200 employees, and was developed for less than $10 million. How are these two companies now competitors?

Lots of hot takes on whether it's possible that DeepSeek made training 45x more efficient, but @doodlestein wrote a very clear explanation of how they did it. Once someone breaks it down, it's not hard to understand. Rough summary:

• Use 8 bit instead of 32 bit floating point numbers, which gives massive memory savings
• Compress the key-value indices which eat up much of the VRAM; they get 93% compression ratios
• Do multi-token prediction instead of single-token prediction which effectively doubles inference speed
• Mixture of Experts model decomposes a big model into small models that can run on consumer-grade GPUs

I spent the day learning exactly how DeepSeek trained at 1/30 the price, instead of working on my pitch deck. The tl;dr to everything, according to their papers:

Q: How did DeepSeek get around export restrictions?
A: They didn’t. They just tinkered around with their chips to make sure they handled memory as efficiently as possibly. They lucked out, and their perfectly optimized low-level code wasn’t actually held back by chip capacity.

Q: How did DeepSeek train so much more efficiently?
A: They used the formulas below to “predict” which tokens the model would activate. Then, they only trained these tokens. They need 95% fewer GPUs than Meta because for each token, they only trained 5% of their parameters.

Q: How is DeepSeek’s inference so much cheaper?
A: They compressed the KV cache. (This was a breakthrough they made a while ago.)

Q: How did they replicate o1?
A: Reinforcement learning. Take complicated questions that can be easily verified (either math or code). Update the model if correct.

Explainer: What's R1 & Everything Else?

Is AI making you dizzy? A lot of industry insiders are feeling the same. R1 just came out a few days ago out of nowhere, and then there’s o1 and o3, but no o2. Gosh! It’s hard to know what’s going on. This post aims to be a guide for recent AI develoments. It’s written for people who feel like they should know what’s going on, but don’t, because it’s insane out there.

NYT: How Chinese A.I. Start-Up DeepSeek Is Competing With Silicon Valley Giants

The company built a cheaper, competitive chatbot with fewer high-end computer chips than U.S. behemoths like Google and OpenAI, showing the limits of chip export control.

The Path to AGI (1h)

AI today is overhyped in the short term but massively underappreciated in the long term—what’s coming will be truly transformative.

Claude and the Future of Work (35 min)

As AI gets better, critical thinking will become one of the most valuable skills—people need to question what they see more than ever.

YC Podcast of the week (40 min)

AI and Decision-Making

The future is unknown and unknowable. The real skill is learning how to deal with what you don’t know.