Ouch! Poor ollama. Try ovllm instead, a python library* wrapping vllm for ease of use and pairing with dspy. Vllm is extremely performant for batch inference. Because of that, it takes a bit more time for models to be supported (a month at most), but when they are, they run

I love small models 🥰

We now have a channel within the DSPy Discord server where we can discuss doing DSPy program optimization with RL! Link below, all are welcome!

Should you RL your compound AI system or optimize its prompts? We think both! 🤯 A short preview of work co-led with Noah Ziems and Lakshya A Agrawal!👇

What a beautiful paper main page:

I am very happy to see those types of project happening. Those are some of the first applications of deep learning I was dreaming about while having a beer with my friends as master students in ecology learning about AI in 2017. AI is a major piece to the puzzle of

yesterday, I was eagerly refreshing for this view to update to gpt-5. today I eagerly refreshed to get it back when I turned legacy model back on.

gpt 5 was a good release for the masses a bad and confusing release for AI very advanced users.

This week’s releases have convinced me that there will be no single magical LLM that reaches human-level expertise at almost everything. Not because I think it’s impossible, but because it wouldn’t be efficient. Genie 3 is good at Genie-3 things. GPT-OSS is good at what it was

GPT-5 they promised vs GPT-5 they delivered.

I sincerely have grown to love Bash. The constraint that even building an array is an arduous process is wonderful for keeping system configuration scripts brain-dead simple.

We complained and laugh at model naming, turns out we want a very long list of models.

I do pretty well learning the name of humans and reaching out to the one I need when I need. I think I can do the same for AI.

Jeremy Howard Prompt engineering → Context engineering → Router engineering

router engineering is the new hot thing

Just 9 days from that review, and everything changed. Today, Grok 4 heavy is by far better then gemini deep think and gpt-5 pro and opus 4.1 and the task I am doing. very surprising.

Guys, there are a lot of new systems to build! Systems that can effectively leverage the new capabilities delivered by LLMs. Don't waste your time. Use DSPy. ps: I am careful in giving bold advice, but this one is easy to give. If calling AI from python, learn dspy.

This is disappointing, llama 4 (scout) cannot deduce and detect the totals on an invoice. only got 2/11 invoices correct on both values.. although, I have not yet optimized the prompt but not a great baseline!

I’m convinced of two things: 1. 2026 will be the year of specialized LLMs. 2. The only way to predict the future… is to build it yourself.