1) DeepSeek r1 is real with important nuances. Most important is the fact that r1 is so much cheaper and more efficient to inference than o1, not from the $6m training figure. r1 costs 93% less to *use* than o1 per each API, can be run locally on a high end work station and does not seem to have hit any rate limits which is wild. Simple math is that every 1b active parameters requires 1 gb of RAM in FP8, so r1 requires 37 gb of RAM. Batching massively lowers costs and more compute increases tokens/second so still advantages to inference in the cloud. Would also note that there are true geopolitical dynamics at play here and I don’t think it is a coincidence that this came out right after “Stargate.” RIP, $500 billion - we hardly even knew you.
Real: 1) It is/was the #1 download in the relevant App Store category. Obviously ahead of ChatGPT; something neither Gemini nor Claude was able to accomplish. 2) It is comparable to o1 from a quality perspective although lags o3. 3) There were real algorithmic breakthroughs that led to it being dramatically more efficient both to train and inference. Training in FP8, MLA and multi-token prediction are significant. 4) It is easy to verify that the r1 training run only cost $6m. While this is literally true, it is also *deeply* misleading. 5) Even their hardware architecture is novel and I will note that they use PCI-Express for scale up.
Nuance: 1) The $6m does not include “costs associated with prior research and ablation experiments on architectures, algorithms and data” per the technical paper. “Other than that Mrs. Lincoln, how was the play?” This means that it is possible to train an r1 quality model with a $6m run *if* a lab has already spent hundreds of millions of dollars on prior research and has access to much larger clusters. Deepseek obviously has way more than 2048 H800s; one of their earlier papers referenced a cluster of 10k A100s. An equivalently smart team can’t just spin up a 2000 GPU cluster and train r1 from scratch with $6m. Roughly 20% of Nvidia’s revenue goes through Singapore. 20% of Nvidia’s GPUs are probably not in Singapore despite their best efforts. 2) There was a lot of distillation - i.e. it is unlikely they could have trained this without unhindered access to GPT-4o and o1. As @altcap pointed out to me yesterday, kinda funny to restrict access to leading edge GPUs and not do anything about China’s ability to distill leading edge American models - obviously defeats the purpose of the export restrictions. Why buy the cow when you can get the milk for free?
Real: 1) It is/was the #1 download in the relevant App Store category. Obviously ahead of ChatGPT; something neither Gemini nor Claude was able to accomplish. 2) It is comparable to o1 from a quality perspective although lags o3. 3) There were real algorithmic breakthroughs that led to it being dramatically more efficient both to train and inference. Training in FP8, MLA and multi-token prediction are significant. 4) It is easy to verify that the r1 training run only cost $6m. While this is literally true, it is also *deeply* misleading. 5) Even their hardware architecture is novel and I will note that they use PCI-Express for scale up.
Nuance: 1) The $6m does not include “costs associated with prior research and ablation experiments on architectures, algorithms and data” per the technical paper. “Other than that Mrs. Lincoln, how was the play?” This means that it is possible to train an r1 quality model with a $6m run *if* a lab has already spent hundreds of millions of dollars on prior research and has access to much larger clusters. Deepseek obviously has way more than 2048 H800s; one of their earlier papers referenced a cluster of 10k A100s. An equivalently smart team can’t just spin up a 2000 GPU cluster and train r1 from scratch with $6m. Roughly 20% of Nvidia’s revenue goes through Singapore. 20% of Nvidia’s GPUs are probably not in Singapore despite their best efforts. 2) There was a lot of distillation - i.e. it is unlikely they could have trained this without unhindered access to GPT-4o and o1. As @altcap pointed out to me yesterday, kinda funny to restrict access to leading edge GPUs and not do anything about China’s ability to distill leading edge American models - obviously defeats the purpose of the export restrictions. Why buy the cow when you can get the milk for free?
2) Conclusions: 1) Lowering the cost to train will increase the ROI on AI. 2) There is no world where this is positive for training capex or the “power” theme in the near term. 3) The biggest risk to the current “AI infrastructure” winners across tech, industrials, utilities and energy is that a distilled version of r1 can be run locally at the edge on a high end work station (someone referenced a Mac Studio Pro). That means that a similar model will run on a superphone in circa 2 years. If inference moves to the edge because it is “good enough,” we are living in a very different world with very different winners - i.e. the biggest PC and smartphone upgrade cycle we have ever seen. Compute has oscillated between centralization and decentralization for a long time. 4) ASI is really, really close and no one really knows what the economic returns to superintelligence will be. If a $100 billion reasoning model trained on 100k plus Blackwells (o5, Gemini 3, Grok 4) is curing cancer and inventing warp drives, then the returns to ASI will be really high and training capex and power consumption will steadily grow; Dyson Spheres will be back to being best explanation for Fermi’s paradox. I hope the returns to ASI are high - would be so awesome. 5) This is all really good for the companies that *use* AI: software, internet, etc. 6) From an economic perspective, this massively increases the value of distribution and *unique* data - YouTube, Facebook, Instagram and X. 7) American labs are likely to stop releasing their leading edge models to prevent the distillation that was so essential to r1, although the cat may already be entirely out of the bag on this front. i.e. r1 may be enough to train r2, etc.
Grok-3 looms large and might significantly impact the above conclusions. This will be the first significant test of scaling laws for pre-training arguably since GPT-4. In the same way that it took several weeks to turn v3 into r1 via RL, it will likely take several weeks to run the RL necessary to improve Grok-3’s reasoning capabilities. The better the base model, the better the reasoning model should be as the three scaling laws are multiplicative - pre-training, RL during post-training and test-time compute during inference (a function of the RL). Grok-3 has already shown it can do tasks beyond o1 - see the Tesseract demo - how far beyond is going to be important. To paraphrase an anonymous Orc from “The Two Towers,” meat might be back on the menu very shortly. Time will tell and “when the facts, I change my mind.”
Grok-3 looms large and might significantly impact the above conclusions. This will be the first significant test of scaling laws for pre-training arguably since GPT-4. In the same way that it took several weeks to turn v3 into r1 via RL, it will likely take several weeks to run the RL necessary to improve Grok-3’s reasoning capabilities. The better the base model, the better the reasoning model should be as the three scaling laws are multiplicative - pre-training, RL during post-training and test-time compute during inference (a function of the RL). Grok-3 has already shown it can do tasks beyond o1 - see the Tesseract demo - how far beyond is going to be important. To paraphrase an anonymous Orc from “The Two Towers,” meat might be back on the menu very shortly. Time will tell and “when the facts, I change my mind.”
• • •
Missing some Tweet in this thread? You can try to
force a refresh
