Links.

Just stumbled across this gem! Derek Willis wrote a short, but super cool post about using language models to extract useful information from PDFs of congressional job postings. Love how he’s taking something so low-tech and messy and transforming it into actually useful data. This is exactly the kind of practical application that flies under the radar but could bring tons of value for some users.

Great conversation between Logan Filpatrick of Google and Alex Atallah of OpenRouterAI covering a wide range of topics, but most interestingly for me is their discussion on why real world use model evals are what matters.

In going over Meta’s LLAMA 4 model release blogpost, I realized I needed a refresh on how multimodal llms work. A quick search pointed me to Sebastian Raschka’s overview post of multimodal LLMs. I’ll be posting my recap from the post once I get through it.

I’ve always struggled to find a good digital hygiene setup, so I found it helpful to see Anrej Karpathy share his.

Mistral has released a new multilingual, multimodal OCR model, demonstrating strong performance in document understanding and fast inference speed. Cost, however, is still higher than what is offered by Google, with Mistral’s model costing 1000 pages/USD, 6x of what is offered by Google (Gemini 2.0 flash currently processes 6000 pages/USD). More details in their release blog.

Interesting visualization based website with walkthrough of several LLM architectures. Provides just the right amount of detail to understand the intuition behind some of its architecture choices.

``I sometimes talk to founders who wish they had started their company 10 or 20 years ago because there was little competition and lots of great unsolved problems.

I tell them it’s hindsight bias, and not what it felt like at the time. This moment, now, is wide open too.’’

• Dalton Caldwell ‘@daltonc on X’