Exciting news 🎉
Document Question Answering is now a first class citizen in @huggingface transformers! With just 3 lines of code, you can process any document like so:
Ankur Goyal
4,965 posts
Ankur Goyal
@ankrgyl
customer service @braintrust
- I no longer speculate about how various systems work. Go to github, clone locally, open cursor, and chat with the codebase to find out!
- 1/ Today we're thrilled to announce DocQuery, a new #opensource query engine for analyzing documents using large language models (LLMs). impira.com/blog/hey-machi…
- Amidst the chaos, we built something: an AI proxy that lets you use a variety of providers (OpenAI, Anthropic, LLaMa2, Mistral, and others) behind a single interface w/ caching & API key management.
- As models get more powerful, i find myself focusing more effort on context engineering, which is the task of bringing the right information (in the right format) to the LLM. Context engineering is hard because it is pervasive. You need to engineer every layer of the stack to
- it takes an insane amount of work to structure a software project so that an agent can make large, meaningful changes to it without careful review. i suspect that this flavor of "meta engineering" will be the new "software engineering"
- To anti leet code folks -- just spent the last day optimizing a hash table implementation which makes a user visible action in our product ~7x faster. Would not have been able to do that without knowledge of CS fundamentals (nor was today's AI).
- Excited to share that we've raised $36m from @martin_casado at @a16z along with @saammotamedi @GreylockVC @eladgil @basecasevc to further our mission of helping developers build AI products that work. A bit more on what we're up to 🧵
- mark my words door to door sales is making a comeback
- I deeply relate to this post -- I was "VPE" at 24 and "CEO" at 27. The advice I got was "stop coding", "hire PMs", "delegate everything". Listening to that crap was catastrophic. Now I spend all day talking to users, writing code/content, and collaborating with ICs.
- talked to another great team today who ripped out their ai framework the story is the same every time -- most of the value of an abstraction is abstracting across LLMs the rest eventually weighs you down
- highly recommend marrying someone great
- things that most serverless providers don’t handle that will be existential for AI in 2025: * Async execution (o1 frequently returns first byte after 1m) * Websockets (realtime audio) * Secure runtime code execution (AIs generate code, where can I run it?)
