the hope is that “constrained distribution” is not quite the same as true OOD right like if i constrain decoding to valid json grammar, json is still within distribution, im just choosing a “top k” subject to my constraints
LLMS
-
Heavy Lifting with Subword Splits and Max Length Settings
By
–
Literally heavy lifting with subword splits and max_length sets and all
-
GPT-5 Emergent Capability Discovery Announced
By
–
I feel I discovered an emergent capability of GPT-5. I need to do some more testing, but if this proves to be correct, it's pretty wild.
-
MCP Servers: Convert Specs to Code Functions
By
–
This is an awesome approach to using MCP servers. We need more people to pay attention to this. Basically, convert a MCP spec into code functions, and have the model write code to call the functions, instead of typical tool-caling. I’ve been doing this with Python for a long
-
Leveraging AI Models for Code Generation Over Tool Calls
By
–
This is an awesome approach. I’ve been doing this with python for a long time now. You always want to lean into what the model is good at, and obviously, models have seen far more code than they have tool calls. Plus there are benefits like loops etc!
-
GPT-4.5 Search Limitations Frustrate Users
By
–
They do this with GPT-4.5 too by the way, any time search is needed
— Matt Shumer (@mattshumer_) 26 septembre 2025
So annoying https://t.co/FC1kGu9T4jThey do this with GPT-4.5 too by the way, any time search is needed So annoying
-
Models Available on Hugging Face with Multi-Framework Support
By
–
All the models are available on Hugging Face and compatible with Transformers, vLLM, llama.cpp, and ollama out of the box.
-
Automate Web Workflows Using LLMs at PyCon Kenya
By
–
I am at @PyconKenya and have a talk at 12:35PM today. I am talking about how to automate web workflows using LLMs. We going to do an AirPods giveaway. Join my talk and win one of those 😉
