solid article on building more complex memory from scratch it has a lot of good specific techniques in here across two high level approaches
thoughtful coding is talking to an ai first about what you’re going to ask the ai to build
it’s been so fun watching the build energy this past week, i had to start rebuilding babyagi again last night :)
>be openai >announce plans to take a cut of openai-aided discoveries >release prism, free tool for research teams >rate limit prism >announce prism+, more free credits, latest models >just promise small equity if you build co >openai becomes next yc >sama used to run yc

some thoughts on naming AI products naming is important, and has impact, especially on early virality. babyagi was originally called "task-driven autonomous agent" and don't think it would have performed as well autogpt had a more descriptive name, and had a lot...
these guys got @replit agent to build a fully local and offline AI voice assistant for mobile 👀

very excited to share a project i've been iterating on for 3+ years: an AI-powered startup "accelerator" inspired by my time at Techstars and seeing similar programs, I've wondered what an AI-first version could look like check out what i made this weekend...

i was cleaning my office and found this sketch from when I was building BabyFoxAGI (6th iteration) back in Sept 2023 the two pane approach was 6 months before Claude artifacts, and I think this general pattern is still pretty solid...
would be cool to see a PR review bot that automatically generated and tracked potential features to add, and scored each PR on which potential features become easier/harder to add later based on architectural changes in the PR
so chatgpt can help you with buying Shopify products, but the merchant will pay 4% more, which goes to openai (edited an earlier mistake) given current user base, Shopify merchants should be happy to get additional traffic combined w ads, kinda feels...

finally starting to see some longer solutions in KISS (prompting game), after disabling emojis and special characters. solutions still aren't quite like in the taboo game where you have to speak the hints, there are still a lot of acryonym/hacky solutions...
using @runanywhereAI, I have Mistral 7B running on a 5 year old phone (iPhone 13 pro) given phones getting more powerful and models getting smaller, how much longer til we have sufficient reasoning capacity on our phones that we don't rely...
we need an agents .md for everything coding agents got better when we collectively decided on a design pattern: point the coding agent to a single md file they got even better when we started organizing multiple md files, stemming from our...
can you learn generalizable skills playing video games? can AI? great article from @antimlabs on their experiments training AI by having them play flappy birds, snake, and more.

random thoughts on building this prompting game... one challenge is that every challenge so far has been solved in 1-3 characters. since i was using the same model to generate AND guess the random phrase, it was REALLY good at...

Fastest/cheapest inference for image generation? @FlyMy_AI gets added to @ArtificialAnlys /image as a provider, check out their speed vs cost running Qwen https://t.co/bIWEXQxuC1
our last 10 investments were: - 3 dev tools/infra - 3 vertical niche b2b - 2 deep tech - 1 consumer hardware - 1 general agent would be fun to look at some consumer AI this quarter

day #8 of the prompting game how many characters will it take you to get the AI to guess today's phrase? play at https://t.co/jmmguoWeQG https://t.co/jlL6jelBe9
before all these agents, i always had a python script that would traverse all the folders and files to concat into a single file so i could copy/paste easily into chatgpt
do you know anyone w a shit ton of video data? have them check out @videodb_io to build search on top of it, annotating at scale (eg. for training data), vibe-editing videos from clips, etc.
increasingly the capabilities of an agent often results in increased cost/time spent on easier tasks “oh, you just need a router to determine if it needs more or less capability” easier said than done
the idea of using traces so agents can better perform over time is easy to talk about, but hard to implement for anyone just getting started, this was a v1 of me trying this 1.5 yrs ago. you'll find it breaks...

what does an ai first legal service look like? @docdraftai pairs legal expertise from real human lawyers with ai to provide affordable solutions for customers they have a new /customer-stories page with case studies so i so pulled out stats to better...
you can effectively give an ai power of attorney if you set it up through an llc, but there still needs to be a human fiduciary
first-order ai literacy is knowing what to say to a model. prompts, tweaks, vibes, try again. second-order ai literacy is knowing how it hears you. fewer words, more shape.
okay i lied. the first prompt was, "here are my AI investments, write me a VC think piece about the state of AI incorporating these" and then we iterated. still, it reads pretty well, captures our thinking, and a good intro...
i almost didn’t open source babyagi
we did something similar at @untappedvc queued up chatgpt to create questions we talked through as a team, uploaded transcription of the meeting to generate a plan, converted it into a gantt chart, and then we do monthly meetings to talk...

my year w chatgpt: - agents, fund II, AI in VC - most likely to automate himself out of a job - 11k msgs sent across 1342 chats https://t.co/n5ibOtorvd