OpenAI's new gpt-oss is now on Poe! gpt-oss-120b is a state-of-the-art open-source model that delivers exceptional performance in complex reasoning, multi-step problem-solving, and analysis. (1/2)
RESEARCH
-
Building Intuition Through Targeted ML Experiments Over Sweeps
By
–
For training experiments I try to avoid sweeps of all kinds. Instead I aim to build a strong intuition by running targeted experiments one at a time.
-

AI Applied to Major Global Challenge Shows Promise
By
–
Amazing to see this push to apply AI to one of our most pressing problems. I hope we can see a lot more of this.
-
Synthetic Data Quality Reduces Compute Requirements
By
–
Do you think the synthetic data is relevant here? What I mean is that some effective compute has been embedded into the data to generate these high quality tokens, so you don't need as much compute to get value out of them
-
Sam Altman’s AI Model Ambitions Lack Self-Awareness
By
–
sam altman wants to build a model that’s rated 3200 on codeforces but has no idea who Sam Altman is and i think that’s beautiful
-
Claude System Prompt Connection Philosophy Explored
By
–
My favorite part of the Claude system prompt is how it ends: "Claude is now being connected with a person."
-
Evaluation Methods for Prompt Engineering Changes
By
–
I would love so much to see the evals you used to figure this one out (or indeed any evals for any of these prompt changes)
-
GPT-OSS Model Performance: Coding Excellence Mixed with Factual Hallucinations
By
–
i’ve spent the last couple hours talking to gpt-oss and can safely say it’s unlike any model i’ve tested one second it’s coding for me at a professional level, the next it’s making up basic facts and clinging to them no matter what i say something very strange is going on
-

Google Research and UCSC achieve genome assembly accuracy improvements
By
–
Some quite exciting improvements in genome assembly accuracy from @GoogleResearch and @ucsc Genomics Institute!
-
GPT-5 Achieves 90% on Public Benchmarks, Promising Results
By
–
Once again about the presumed GPT-5 benchmarks: testing was done with the public simpleBenchDatasets. These indicate that around 90% is achieved. That is excellent – and in all probability GPT-5 will be an outstanding model. I don't think there's any doubt about that now.
