AI, data, agents, reliability, context, and technical architecture.
90% of time when people say "Retrieval-augmented generation" they mean that the index is built using an embedding model like OpenAI's text-embedding-002 and a vector database like Chroma, but it doesn't have to be this way. Retrieval is a long-standing problem in computer science -- a couple of PhD students made
Retrieval-augmented generation, or RAG, is a fancy term hiding a simple idea: Problem: LLMs can reason, but they don't have the most relevant facts about your situation. They don't know the location of your user, or the most relevant passage from the knowledge base, or what the current list of customers is in their CRM.
Apps with no Memory are boring. Compare a static website from the 90s with any SaaS or social network or phone app: the former knows nothing about you, the latter knows a lot. From UI preferences (dark or light mode?) to basic personal data (name?) to your friend list to what things you engage with the most, today's
The stages of an LLM app seem to go like this: * Hardcode the first prompt, get the end-to-end app working. * Realise that the answers are bad. * Do some prompt engineering. * Realise the answers are still bad. * Do some more prompt engineering. * Discover vector databases!!!1 * Dump a ton of data as plain strings into the vector db for
A couple weeks after AutoGPT came out we tried to make it actually usable. If you don't know yet, it looks amazing on first glance, but then completely fails because it creates elaborate plans that are completely unnecessary. Even asking it to do something simple like "find the turning circle of a Volvo V60" mostly takes
AutoGPT (repo) went viral on Github and looks impressive on Twitter, but almost never works. In the process of trying to improve it I dug into how it works. Really there are two important parts to AutoGPT: a plan-and-execute workflow, and looped Chain-of-thought (CoT). Plan-and-execute Say the user input is "Find one funny
Software 1.0 -- the non-AI, non-ML sort -- extensively uses testing to validate things work. These tests are basically hand-written rules and assertion. For example, a regular expression can be easily tested with strings that should and should not give a match. In software 2.0, and specifically supervised learning, the program is automatically learned from the dataset.
GPT-4 supports images as an optional input, according to OpenAI's press release. As far as I can tell, only one company has access. Which makes you wonder: how can you get multimodality support already today? There are basically two ways for adding image support to an LLM: 1. Train a vision encoder that makes the image digestible
Chain-of-thought reasoning is surprisingly powerful when combined with tools. It feels like a natural programming pattern of LLMs: thinking by writing. And it's easy to see the analogy to humans: verbalizing your thoughts in written (journalling) or spoken ("talking things through") form is a good way to actually make progress. To put another way:
Matt Rickard has a concise overview of Chain-of-thought: the design pattern of having an LLM think step by step. To summarize, the four mentioned approaches from simpler to more nuanced are: 1. Add "Let's think step-by-step" to the prompt. 2. Produce multiple solutions, have the LLM self-check on each one, pick
