Could be the model is getting confused because it has special training about a "search" tool baked into the model https://
github.com/openai/gpt-oss
?tab=readme-ov-file#browser
… Frustrating that it doesn't notice when the search tool hasn't been configured though
@simonw
-
Model Confusion with Search Tool Configuration in GPT
By
–
-
LLMs Cannot Accurately Describe Their Own Capabilities
By
–
My guess is that the fraction of humans who understand enough about LLMs to ask those questions overlaps pretty heavily with the fraction of humans who know that LLMs are almost universally terrible at accurately answering questions about themselves
-
Token Efficiency in System Prompts: Prioritizing Common User Queries
By
–
I don't think it would be worth using up valuable tokens in the system prompt to cover topics that most users would never think to ask about
-
GLM 4.5 Air outperforms gpt-oss-20b despite higher RAM usage
By
–
Not sure yet! GLM 4.5 Air on my laptop impressed me more than gpt-oss-20b but also used nearly 4x the RAM
-
Tool Calling Implementation Challenges in New Harmony Models
By
–
Sounds to me like it's tool calling, which is the bit of the model I've not got working yet myself – the new Harmony stuff means there are plenty of details that model serving tools need to nail down before I'll be confident if we know how well that actually works
-
H100 GPU Pricing Estimates and Model Card Analysis
By
–
I got those H100 pricing estimates from here: https://
thundercompute.com/blog/nvidia-h1
00-pricing
… More of my notes on the model card: -
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
-
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
… -
Running AI Models Locally and via APIs with LM Studio
By
–
My full piece on the models, including various pelicans on bicycles and detailed notes on running them on my own laptop using @lmstudio and @ollama – and via API using @openrouter and @cerebras
-
OpenAI Open Weights Models Rival Chinese AI Labs
By
–
Here's the concluding section of my write-up of the new OpenAI open weights models – they're really impressive, and I think they hold their own (or even beat) the models we've seen come out of the various Chinese AI labs over the past few months