Fernando Freire

Fernando Freire

@dogonthehorizon.com

I do things and sometimes they work out! Founding Engineer @ teem.finance; PLT/FP/Haskell Journeyperson otherwise.

110 followers519 following63 posts

Love this framing of AI-assisted dev. I've been thinking for a week or so now that our small (3) dev team needs a shared prompt library to generate consistently. We're all using Sonnet3.5, but in slightly diff ways. cursor.directory et. al. seem like good starting points.

annievella.comThe SDLC Strikes Back: Adapting to AI-Driven Development - Annie VellaEarlier this year, Lovable celebrated their biggest milestone yet - more than 12,000 (!) new projects created in a single day. The very next day, they went down. The irony? Their success became their ...

The prompt engineered approach is likely useful for a variety of requirements, not just perf opt. I've had good success w/ the LLM by having it output code that answers a series of questions as well. Excellent read!

Max Woolf@minimaxir.bsky.social

New year, new blog post: I had a random question, what happens when LLMs are prompted to write better code, again and again? Do they actually write better code? The answer is yes*! minimaxir.com/2025/01/writ...

Lot of hype around ghostty. Other than sane defaults, biggest change for me is seeing big perf increase while scrolling large tsx files in neovim. Kitty struggled, ghostty has no problem. Jury is still out on whether it warrants the hype 😝

Anyone in the #neovim sphere working on a model context protocol plugin? I want avante.nvim to leverage mcp servers. I tried wrapping the python sdk in a rplugin but the experience of installing/updating leaves a lot to be desired. Writing a lua mcp client seems like the next best option?

Discharged these headphones too much to the point they are undervolted. Taking the extremely reasonable step of macguyvering a manual charge until they have 3v to start the normal charging process 😒😒😒 the last pair of helping hands are my partner's hah. What an adventure.

Fernando Freire reposted

A novel agent architecture that simulates the attitudes and behaviors of 1,052 real individuals 85% accurately --applying LLMs to qualitative interviews about their lives, then measuring how well these agents replicate the attitudes and behaviors of the individuals that they represent.

We have this homebrew agent framework that we've been building at teem.finance but we've been slow in adding new learnings to older parts of the framework. Just spent some time today refactoring a dynamic task planner and saw 90% reduction in cost, 30% speed improvement. 🧵

View full thread

- 90% reduction in cost using 4o-mini - 30% improvement for p50 generation, 50% for p99 - 15% reduction in output tokens All possible because we can measure the changes we're making (currently Langsmith) and see behavior change across iterations.

The two big takeaways: - Evals are great and help measure improvements in probabilistic systems - Structured outputs (in whatever form they take, but we like instructor) help harness probabilistic systems in deterministic ways Feels pretty good!

We have this homebrew agent framework that we've been building at teem.finance but we've been slow in adding new learnings to older parts of the framework. Just spent some time today refactoring a dynamic task planner and saw 90% reduction in cost, 30% speed improvement. 🧵

View full thread

Like all projects, tho, requirements change. We've since added support for a dozen more task types, the goals themselves are getting more complex, and the tasks now ask for data from dependents in order to succeed. As we scale up the platform, this gets more expensive and less deterministic.

And since starting with instructor we've learned a lot! Better prompting for smaller models, implementing validators, and we're no longer 0 to 1; we have data we can use to inform the context of future runs of the agents. We did all this and managed to move to 4o-mini in the process, so we now see:

python.useinstructor.comEnhancing AI Validations with Pydantic's Framework - InstructorLearn how to improve AI outputs using Pydantic for validation and reasking techniques.

We have this homebrew agent framework that we've been building at teem.finance but we've been slow in adding new learnings to older parts of the framework. Just spent some time today refactoring a dynamic task planner and saw 90% reduction in cost, 30% speed improvement. 🧵

The v1 planner is inspired by examples in `instructor`. Given some goal, it plans tasks for the goal given a schema of available tasks. This works well when the goals are simple and the tasks available help accomplish the task. The best frontier models can usually figure out the right thing to do.

github.comGitHub - instructor-ai/instructor: structured outputs for llmsstructured outputs for llms . Contribute to instructor-ai/instructor development by creating an account on GitHub.

Fernando Freire reposted

entropix: qwen vs llama when using entropix, qwen 2.5 7B coder seems to produce much clearer entropy paths for entropix to follow, vs llama 3.1 8B this makes qwen+entropix a great combination for writing code

@thomas.lockney.net is great systems thinker, leader, and engineer; you'd benefit greatly from having him in your camp 😃

Thomas Lockney@tlockney.net

Hey folks, I'm looking for my next role with a remote friendly organization. I have extensive experience in product and platform engineering and have been in engineering leadership roles for the last decade. More details on my background here: https://linkedin.com/in/thomaslockney

Combining vim-fugitive & avante is enabling me to write first-draft commit messages based on the changeset from `git commit --verbose`. Not the biggest time saver but it's definitely reinforcing the idea that LLMs can do some non-0% of the work with close senior eng review.

Fernando Freire reposted

> It's scary how versatile/productive your terminal, and specifically Neovim, can get. In one screen: 1. data integration/dbt code 2. analysis of SQL queries 3. db connections/browser 4. result of queries 5. docker build 6. dbt run 7. postgres 8. more windows/sessions (tmux) More dotfiles.ssp.sh.

Gave Avante a try and it's pretty cool. Ran it on a medium nextjs project and it wrote comments, added simple functions, and searched for things across the project. Tried 4o-mini first and didn't work, sonnet 3.5 is. Need to tweak prompts so 4o works, otherwise it'll be too expensive for daily use 😬

Another day messing with avante and it's pretty slick. Using the `@codebase` marker with a project using shadcn gives me a "v0 at home" vibe. Assuming today was a good sample it should come out to ~$40/mo which is still a bit. I'm sure 4o-mini or haiku 3.5 could be used with some elbow grease

You are up to date.