Will Open AI’s o1 Reasoning model really change the world?

OpenAI launched Strawberry – another name for its Model 01 – on September 12, including the fully-featured o1-preview and the cheaper o1-mini, to demonstrate how AI can work by breaking down a query into step-by-step -Step-steps can greatly improve reasoning. This chain of thought works more like we humans do, breaking down a problem into bite-sized pieces and tackling (calculating) those pieces one at a time. o1 is particularly useful for STEM and math problems and has received a lot of interest.

Is o1 that good?

OpenAI o1 is definitely impressive. While most would argue that it is not yet general AI, it scores 83% on the Mathematics Olympiad and operates at essentially Ph.D.-level accuracy, up from 13% for GPT4.0.

But o1 has three problems: 1. it is very slow; 2. it is four times more expensive ($60 per million tokens versus $15 for GPT4); and 3. it’s just text. Solving the last problem will not be that difficult for OpenAI. And it can make progress on the first two, but solving the problem step by step and then iterating to find the “best” answer just requires a lot of computation. This is not an easy solution.

Cerebras announces o1-like reasoning on Lama 3.3-70B

Currently, the first two issues are being addressed by Cerebras Systems (Disclosure: Cerebras is a customer of my company, Cambrian-ai Research). Cerebras announced today at the annual NeurIps conference that it has released an open source version of Llama called CePO (Cerebras Planning and Optimization) based on Llama 3.3-70B, which adds O1-like reasoning to Llama models. Cerebras claims that this is the first and only reasoning model that can run in real time, providing answers to complex questions in seconds instead of the minutes required by OpenAI’s o1 model.

CePO makes Metas Llama 70B more accurate than Llama 405B and more accurate than GPT4 for many problems. Cerebras has not yet launched this new model, but may add it to its inference services portfolio announced earlier this year.

Cerebras also showed performance data where its CePO outperformed the current Llama-3.3 70B model and even the Llama 3.1 405B model.

What is our opinion?

Reasoning style 01 is quite powerful, but requires far more computational effort than GPT 4.0. Based on the current OpenAI pricing model, it likely requires around four times more computing power. This is where Cerebras’ wafer-scale engine comes into play, with tons of cores and SRAM memory on the Frisbee-sized “chip.”

Disclosures: This article reflects the author’s opinion and should not be construed as a recommendation to buy or invest in any of the companies mentioned. Cambrian-AI Research is fortunate to count many, if not most, semiconductor companies among our clients, including Blaize, BrainChip, Cadence Design, Cerebras Systems, D-Matrix, Eliyan, Esperanto, Flex, GML, Groq, IBM, Intel, Nvidia, Qualcomm Technologies, Si-Five, SiMa.ai, Synopsys, Ventana Microsystems, Tenstorrent and numerous investment customers. We have no ownership interests in any of the companies mentioned in this article, nor do we plan to do so in the near future. For more information, see our website.