Researchers open source Sky-T1, a “rational” AI model that can be trained for less than $450
Developing so-called inferential AI models has become easier and cheaper.
On Friday, NovaSky, a team of researchers from UC Berkeley’s Sky Computing Lab, released Sky-T1-32B-Preview, a logic model that rivals a logic model. The previous version of OpenAI’s o1 On a number of key criteria. The Sky-T1 appears to be the first truly open source thinking model, in the sense that it could be Duplicate from scratch; The team released the dataset they used to train it along with the necessary training code.
“It is remarkable that the Sky-T1-32B-Preview was trained for less than $450,” the team wrote in a message. Blog post“Demonstrating that it is possible to replicate high-level thinking abilities affordably and efficiently.”
$450 may not seem affordable. But it wasn’t long until the price of training a model with similar performance was determined They often ranged in the millions of dollars. Synthetic training data, or training data generated by other models, has helped reduce costs. Palmyra X 004, a model recently launched by AI company Writer, has been almost fully trained Synthetic dataIt is said that its development cost only $700,000.
Unlike most AI systems, inference models actively verify facts, which… It helps them avoid some of the pitfalls that models usually stumble upon. Heuristic models take a little longer – typically seconds to minutes – to arrive at solutions than a typical non-heuristic model. The upside is that they tend to be more reliable in areas such as physics, science, and mathematics.
The NovaSky team says it used another logic model, Alibaba QwQ-32B-Previewto create the initial training data for Sky-T1, then “formatize” the data mixture and leverage OpenAI gpt-4o-mini To restructure data into a more executable format. Training a Sky-T1 with 32 billion parameters took about 19 hours using a cluster of 8 Nvidia H100 GPUs. (The parameters roughly correspond to the model’s problem-solving skills.)
According to the NovaSky team, the Sky-T1 performed better than the early preview version of o1 on MATH500, a set of “competition-level” sports challenges. The model also outperforms o1preview on a set of challenging problems from LiveCodeBench, a coding evaluation.
However, Sky-T1 falls short of the o1 preview on GPQA-Diamond, which contains questions related to physics, biology and chemistry that a PhD graduate would be expected to know.
It’s also important to note that OpenAI’s GA version of o1 It is a more powerful model than the preview version of o1, and OpenAI is expected to release a better-performing logic model, o3in the coming weeks.
But the NovaSky team says the Sky-T1 represents just the beginning of their journey to develop open source models with forward-thinking capabilities.
“Going forward, we will focus on developing more efficient models that maintain strong logic performance and exploring advanced technologies that further enhance the models’ efficiency and accuracy at test time,” the team wrote in the post. “Follow us as we make progress on these exciting initiatives.”
