Maurício Maia

@itsandrewgao There's also all the new bets "games"

in this essay i will discuss how tiktok killed mobile gaming👾🎮

@bentossell firecrawl.dev can scrape the docs and turn them into markdown. It has a LangChain integration that should make the second step easy to implement.

how do you download app documentation easily? (if you can?) like stripe, notion, posthog, etc I wanna download, then use ai (duh) on them. use case is often giving two lots of docs and saying how do i make these work together

@GregKamradt 140 companies and 1.7k active jobs rn. 8.5k jobs in the database

@mmaia oh cool, just for openai? Or others too?

I wanted to build a mini app that alerted me when there were new jobs at OpenAI New openings help you read between the lines of a company's growth The build ended up being bigger than I thought Here's the story: Started wanting to know when there was new jobs, but if you want to know what is new, you need to know what is old. So you need to store the old state (I didn't want to use created_at timestamps). So enter a postgres db (via supabase) But why do it just for OpenAI? I want to see other AI companies too Ok so now we need the concept of a company. I originally attached a job_board_token to each company, but had a hard time matching a stream of job board data with a company. So I separated streams from companies, then made an associations table But wait! Turns out that not all AI companies use the same job board provider. OpenAI uses Ashby, Anthropic used Lever, switched to Greenhouse Ok so now I need to create different fetchers and parsers for each job board But then I obviously wanted to config where notifications went and for specific companies So then I needed an events table, subscribers table that said who was listening for events (with filters) and notifications table But wait! Let's let the system grow, how do you find new job boards to track and scrape? I needed a way to make a proposal (a potential board to match), and confirm that it had valid data. So made an proposal service that just scanned the web for potential matches, then checked them against my job board providers But there isn't just one notification type, you'll have tweets (like the quote here), webhooks, emails, sms, etc. so you need to hook up all those channels Some light LLM work to summarize the job description (only for notifs, not for all jobs) Github actions scrapes 3K job boards every day (mostly tech/AI companies) And so here we are today Up next is to pull out the tools mentioned in each job to make a proper list of tools per companies

@vboykis Pocket could do it but...

Free product idea: a search engine that’s actually recommendations, ie I search for a post URL and it returns me “posts like this across the internet”. For example, I want to read more about how software teams function or dysfunction erikbern.com/2023/12/13/sim…

the rest of the owl

@simonw If it is a batch process, add another step to "reduce" the tags.

Anyone got any prompting tricks to help with consistent tagging? Ask an LLM for a JSON array of tags for some content and it will work... but if you repeat that against 100 pieces of content you'll get varying results because each document is considered independent of the others

@mhmazur I sense that if a model performs at X for some set of problems, people will try to make it perform at the same level for other sets of problems by just improving the prompt.

@mmaia For example, is there a prompt in the universe of possible prompts that could help PaLM 540B achieve a score of 80 instead of 57? How do we know?

LLM theory question: when evaluating an LLM using benchmarks, how much gain is possible by better prompting? Is there a theoretical limit or might it be possible to see massive improvements simply by using an exceptional prompt?

@borowis @mhmazur There're demonstrated cases where it's possible for GPT-3 to match GPT-4 by just improving the prompt.

@mhmazur I don't have quantifiable information, however I don't believe that i.e GPT-3 would be able to perform at GPT-4 level with better prompting with what I've seen and tried so far

I tried to cover the fundamentals of vector search and how to make search queries more efficient using @postgresql and @pgvector. Is there anything else I should have added? neon.tech/blog/understan…

1/ 🤖 You don’t need a specialized vector database. You just need PostgreSQL. Introducing Timescale Vector: PostgreSQL++ for AI Applications - 📈 3x ANN search performance vs Weaviate, 40%-1,500% boost for pgvector. - ⚙️ No need to learn and manage a separate vector database.

@OfficialLoganK @Kwebbelkop How about just the name of a city? E.g. "Paris", "Tokyo"

@Kwebbelkop Just visited Amsterdam, wonderful place!

Send me your DALL-E 3 prompts, I’ll share the results here 🧵👩‍🎨

@marclou I finally listened to the IH podcast with you. Finding out that B2B is not a founder fit was very relatable.

2018: I sold a software I didn't build 2020: I abandoned the startup 2023: It still makes $600/month Passive income is incredible

Interesting new position spotted at @sourcegraph: LLM Interaction Engineer aijobnetwork.com/jobs/sourcegra… #jobs #remotejobs

🗂️ Data from AIJobNetwork.com where I track jobs from selected AI companies. Analysis inspired by the amazing work of @AznWeng

📈 And the number o jobs published keep growing like crazy.

There are 9 open roles in the new OpenAI's Dublin office