robrenaud

I am also interested in how to do eval on an open source corporate search system. Privacy and information security make this challenging, right?

Deepseek is great. What fundamentally prevents an open source coalition from producing great AI systems for everyone?

> "Note that this s1 dataset is distillation. Every example is a thought trace generated by another model, Qwen2.5"

The traces are generated by Gemini Flash Thinking.

8 hours of H100 is probably more like $24 if you want any kind of reliability, rather than $6.

It's a joke about how Google has released/cancelled/renamed many messenging apps.

Do you understand why RL is better than SFT for training on reasoning traces?

Our genes are heavily evolved to live in calorie scarce environments. In those environments, high calorie foods are amazing. Our biology is built to find them incredibly rewarding.

Science and capitalism have created incredibly delicious foods that are nutritionally lacking, hyper optimized for (against?) our now mis-aligned reward system. In the west, calories are not scarce and the most delcious foods are far from the most nutritious. It will take a long time for our genes to catchup.

Mass producing delicious, cheap, but low nutrition food is profitable. Companies have gotten very good at it. That's the real big change.

My best learning was difficult, whole mind encompassing, and incredibly fun.

If you can get college students idle brains curiously contemplating the how and why of the subject, that's when the tuition is really worth it.

Are the market dynamics such that effective small companies grow, and ineffective small companies shrink? Is this bad?

I really do think hallucinated references are a thing of the past. Models will still make things up, but they won't make up references.

ChatGPT with web search does a good job of summarizing content.

I've been using self hosted langfuse via litellm in a juptyer notebook for a few weeks for some synthetic data experiments. It's been a nice/useful tool.

I've liked having the traces and scores in a unified browser based UI, it made sanity checking experiments way easier than doing the same thing inside the notebook.

The trace/generation retrieval API was brutally slow for bulk scanning operations, so I bypassed it and just queried the db directly. But the is the beauty of open source/self hosted code.

Gemini is much worse as a product than 4o or Claude. I recommend using it from Google AI studio rather than the official consumer facing interface. But for tasks with large audio/visual input, it's better than 4o or Claude.

Whether you want to deal with it being annoying is your call.

Do you fear that some big company will just host your system for cheaper than you can if you catch a lot of success?

That is, the same thing that Amazon did to Mongo will happen to you?

Do you think working in the open enables you to spend more time on engineering and less on sales and marketing?

> For example, combining a human-moderated knowledge graph with an LLM with RAG allows you to build "expert bots" that understand your business context / your codebase / your specific processes and act almost human-like similar to a coworker in your team.

I'd love to hear about this. I applied to YC WC 25 with research/insight/an initial researchy prototype built on top of GPT4+finetuning about something along this idea. Less powerful than you describe, but it also works without the human moderated KG.

Then those new models get distilled into smaller ones.

Raising the max intelligence of the models tends to raise the intelligence of all the models via distillation.

I just moved to SF a couple weeks ago. In city driving, I really like driving near Waymos as opposed to human cars. You quickly learn that the Waymos drive well and passively, which enables safe, assertive driving near them. Being able to condition on them not being aggressive and dumb (like some fraction of humans drivers, mostly young/male) makes driving easier.

> if advancing the SOTA AI is your dream, a product company may not be the right place.

Does Meta get in the way of this?

It's hard to compete with a company that is dead set on spending billions and seemingly wants to drive your SOTA AI product revenue to 0.

If you are OpenAI or Anthropic right now, it seems like trying to run a great restaurant at a reasonable price right next to a good (great?) restaurant that is serving everyone for free.

Is the encoder style arch better for representing classification tasks at a given compute budget than a causal LM?

Is this because the final represention in bert style models more globally focused, rather than being optimized for next token prediction?

The physics professor AI wouldn't have to solve novel problems in physics, it just has to be able to remove the confusion of physics undergrads.

As models improve, their representations converge. This even happens across modality for vision and text.

The reason to fine tune is to get a model that performs well on a specific task. It could lose 90 percent of it's knowledge and beat the unturned model at the narrow task at hand. That's the point, no?

Does this also apply to hiring as well as firing? Should companies hire more to make more money in times of growth? Is this a coherent position?

A good, automatically run, privacy preserving search engine that uses electronic medical records might be a valuable resource for busy doctors.

Newton.

Do EMRs improve the standard of care?

I got a couple patents while at Google. I sent a nice readable 4 page design doc that I wrote to a patent lawyer, and I got back 40 pages of nonsense that I basically didn't understand.

I wish there was some kind of readability requirement for patents , if they are to continue to exist.

It feels pretty good. I did some reading on some high quality posts about chess that I found through it.

What was the biggest thing you learned while implementing this? Was anything surprisingly difficult? Was there anything that worked better than you expected?

Maybe a combined approach beats either? Let some non-LLM reranker quickly spit out two results, and fill in the rest with the LLM.

Finding ways to reward people for building great open source software is something that we need more of. It's not lame at all.

Are there any strong LLMs trained without CUDA?

Model quality matters a ton too. They aren't serving OpenAI or Anthropic models, which are state of the art.