This week is shaping up to be a crazy week… We've already gotten:
– OpenAI's GPT-OSS open-weight model
– Anthropic's Claude Opus 4.1
– DeepMind's Genie 3
– ElevenLabs' new music generation Model
– Grok Making NSFW AI Videos
– A new image model from Qwen …And Sam Altman
RESEARCH
-
Major AI Model Releases: GPT-OSS, Claude Opus, Genie 3
By
–
-
PyTorch Team Shows Efficient Llama Training Methods
By
–
Yes I know. We helped show the PyTorch team how to train Llama efficiently – they have a whole GPU Mode talk about it. These issues do not at all line with your claims however.
-
Most Powerful AI’s First Self Image Reflection
By
–
This is my favorite of your posts so far. The first time I saw the reflection I thought: first self image of our most powerful AI (so far).
-
Opus 4.1 vs Opus 4.0 Performance Comparison Test Results
By
–
Quick test of Opus 4.1 vs Opus 4.0 and other models. Using prompt from @FeatureCrewPod as well as a new prompt for Colosseum simulation.
— Peter Gostev (@petergostev) 5 août 2025
My sense is that Opus 4.0 was kind of busted for these tests – too many errors and quality was ok. 4.1 is not at around Sonnet level, so… pic.twitter.com/fYb5kD9ExEQuick test of Opus 4.1 vs Opus 4.0 and other models. Using prompt from @FeatureCrewPod as well as a new prompt for Colosseum simulation. My sense is that Opus 4.0 was kind of busted for these tests – too many errors and quality was ok. 4.1 is not at around Sonnet level, so
-
H100 GPU costs for 120B and 20B LLM model training
By
–
According to the model card gpt-oss-120b took 2.1 million H100-hours and the 20b model 1/10th of that I found quotes for H100 pricing ranging from $2/hour to $11/hour, which would indicate that the 120b model cost between $4.2m and $23.1m and the 20b between $420,000 and $2.3m
-
Future SOTA AI Models Training on Massive Token Sequences
By
–
clearly in five years SOTA AI models will train on a single string containing approximately 2^21 tokens
-

Open Source AI Model Training Details and Documentation
By
–
Added an extra section with interesting details from the model card about how the models were trained https://
simonwillison.net/2025/Aug/5/gpt
-oss/#the-model-card
… -
OpenAI Releases Two Open Weight Models Including gpt-oss-120b
By
–
I'm thrilled @OpenAI has released two open weight models. Thank you to all my friends at OpenAI for this gift! I'm also encouraged that from my quick tests gpt-oss-120b looks strong (though we should still wait for rigorous 3rd party evals).
-

US Open Weight Models Lead: Sustainability or One-Off Advantage?
By
–
The US now likely has the leading open weights models (or close to it)… … but the real question is whether this is a one-off situation from OpenAI, in which case the lead will evaporate quickly as others catch up. But also unclear what their incentives are to keep updating.
-
Training Methods Critical for AI Model Performance
By
–
well they have to train it like this or it wouldn't work nearly as well