Richard Yang’s Post

Anthropic introduced Contextual Retrieval. By combining it with Best Match 25 it can reduce failure rates up to 67% when used with reranking. Here is the cookbook https://lnkd.in/eyys8pvf #AI #RAG

Tunadorable is a YT personality who does AI-related deep-dives. Here's an interesting one: https://lnkd.in/emh3A37P The short answer is: no, we're not even close to understanding LLMs. But there are a few annoying oversimplifications about new AI tech we can get rid of now! Here are two cheap shots I've heard a lot of lately: "LLMs are just stochastic parrots" or "LLMs are just glorified next-token predictors." Well, technically, yes. Just like "a brand new functional iPhone and this pile of unassembled phone parts are just the same set of parts, so are functionally equivalent." That is completely missing the point. Learn the bitter lesson: when you do something boring, but do it bigger, the emergent properties of the system are surprising and novel. Those of us who've been building massive clusters over the past 30 years know this intuitively. Examples? You can MODIFY LLMs to be more (or less!) honest, and they'll do it. You can give them code with an error and make them unable to accept code can have errors, and they will go so far as to rewrite the code error-free in order to resolve the conflict. These behaviours are novel and interesting and can teach us a few things about ourselves.

Great conversation with Ian Michels! Some highlights: *️⃣ It’s much more of a human problem than an attribution problem; *️⃣ It’s not about the models that come with the tool, its how you’re getting everybody to agree with and understand the attribution data; *️⃣ Quality of data is a big challenge to using attribution; *️⃣ Everything goes better with great targeting On that last bullet -- your attribution reports are going to look a lot better when you filter them to your ideal customer profiles. Getting your ICPs right delivers huge benefits that extend well beyond attribution 🚀

I’m a big fan of explanatory AI (a.k.a. Mech-Interp), where you start breaking into the black box of neural networks. Anthropic came out with a really interesting paper where they used sparse autoencoders to interpret, measure and influence the internal features of LLMs. Highly worth looking into if you like this kind of thing. E.g. they were able to manipulate the LLM to automatically correct bugs in code without prompting it, manipulate the LLMs safety, make it insist that neuroscience is the best scientific field, etc., all by finding the relevant neurons and manipulating their values, without any prompt engineering. https://lnkd.in/eYq8askS Tunadorable also have a really nice read through for the more audibly inclined: https://lnkd.in/e-TbsCAd

A glimpse into the future of AI vendor lock-in. https://lnkd.in/gVyQKqJg As discussed previously, vendor lock-in appears to be the main point of the vast amounts of marketing around AI. GenAI training, with its extremely large data and computation requirements, is the most obvious area for an AI natural monopoly. https://lnkd.in/gwygRzG4 Given Google's trove of data, it's hard for me to see any other company winning this race. Its competitors need to monetise their fleeting times at the top of the various leaderboards.

This promises to be an interesting read.

A good presentation on the use of AI tools and Systems Engineering. My main takeaway is that I need to learn more about the AI natural language systems. https://lnkd.in/draG6Xmw

Facinating. Identifying concept representations in LLM. By Anthropic. This Hard Fork episode has a segment on it: https://lnkd.in/eT6wEghz

https://lnkd.in/gPj6Nj6i A curation of what I enjoyed reading this week, along with my Zoharian commentary:

Evolutionary Theory vs. Intelligent Design and Syntropy chatting with claude3 Opus - see the conversation at aiyadda.com