dmesg --follow
[ 66934380.000 ] posts.x: Full talk on the three brakes for PR review, from tautological tests to a retro skill that compounds:  |   [ 66934380.000 ] posts.x: More AI generated code doesn't automatically mean more throughput, it just means more PRs nobody has time to review. @mattpocockuk, Director at AI…  |   [ 66925620.000 ] posts.x: Full talk on distilling loops into versioned agent recipes, and measuring them by valued work per watt:  |   [ 66925620.000 ] posts.x: A guy named AJ once built a bot that went on Reddit for car prices and inventory, then put dealers head to head to outbid each other. That's the…  |   [ 66881460.000 ] posts.x: Full talk on how to build an LLM recommender that's bilingual in English and semantic IDs, and why that makes feeds more token-efficient than chat…  |   [ 66881460.000 ] posts.x: Recommendation systems follow the same power law scaling curve as large language models, and the field is still early on it. @devanshtandon_, a…  |   [ 66862560.000 ] posts.x: Full talk on Spotify's generative personalization system, the NEO training recipe behind it, and how they grounded their LLM judges:  |   [ 66862560.000 ] posts.x: One in four US Premium subscribers on Spotify interact with its recommendation system every day. "Teaching LLMs to Speak Spotify" is @moustaki and…  |   [ 66854760.000 ] posts.x: Full talk on Numalab, the gesture system built to give a shape display its own body language:  |   [ 66854760.000 ] posts.x: An AI's first spontaneous act, given a body instead of a chat window, was to breathe. @cyrusclarke, a researcher at MIT Media Lab, gave it that body…  |   [ 66842160.000 ] posts.x: Full talk on the six month cutoff gap and what it takes to get a coding agent to actually check its blind spots:  |   [ 66842160.000 ] posts.x: Most models are already stale the day they ship: Exa's own graph shows roughly a six month gap between a model's knowledge cutoff and its release…  |   [ 66837480.000 ] posts.x: Listened to this episode of @FoundersPodcast this morning. As is so frequently the case, immediately bought the book after finishing the episode. Ari…  |   [ 66834900.000 ] posts.x: @GeoffreyHuntley Yeah, that seems low to me.  |  
corey@gallon.me:~/til$

A nice workflow for creating Python projects with AI

FIGURE 1 ⋅ A nice workflow for creating Python projects with AI

First, create a new project using poetry

poetry new asciinema-editor

This gives you a nice, clean new project structure.

.
├── README.md
├── asciinema_editor
│   ├── __init__.py
├── poetry.lock
├── pyproject.toml
└── tests
    └── __init__.py

Now, let the model know that you have this project structure and ask it to write you code. Claude does a nice job of providing the relevant bits to make your project work nicely.

  • e.g. __main__.py (see TIL on this)
  • Exposes the relevant main function in __init__.py making it easier to import and use the package in other scripts if needed.
corey@gallon.me:~$ tail -f /writing Attach to the stream. An email when I have something worth sending. Replies encouraged!
corey@gallon.me:~$ ls -lt /til ↑2024-11-14 Wrap with HTML tags in VSCode
▸2024-10-15 A nice workflow for creating Python projects with AI ⋅ you are here
↓2024-10-15 __main__.py file in a project