Building apps. Open-source @ https://teapaw.org/t.co/fMgG5BQ6b5. Blog @ https://teapaw.org/t.co/nnueDVNZap 🚴 ☕
Portland ↔ Depoe Bay, OR
Joined December 2007
- Tweets6.9K
- Following1K
- Followers713
- Likes21.8K
Pinned Tweet
I'm excited to announce The Foundry Project!
I've been fine-tuning LLMs to reanimate the voices constitutional founders for debate and conversation starting with James Madison.
468K words of primary sources. A 5K-word character constitution built from his writings and biographies. ORPO preference training on Qwen 3-32B.
Open source. Docs, training results: github.com/seaberger/Foundry
I found that a Mac mini performs XGBoost modeling faster than any cloud server I could find including v5 droplets. Just setup a new M5 Pro mini to act as a MCP server for my upcoming beta app launch. Stoked!
I tried to use my Codex reset after 9pm PDT yesterday but @OpenAI @OpenAIDevs @thsottiaux I was denied and it disappeared unused. 😞 you didn’t say they expired before the end of the day and I saw no countdown or specific hourly deadline in the app.
Yes! Can’t wait to share what I’ve been building soon.
to build a cool product now, you really only need two ppl:
one dude with exceptional vision & taste who knows what should exist & can provide the right feedback loops for iteration.
& one dude who is absolutely relentless about making it real no matter what.
that’s it. & sometimes these two are the same.
I did an experiment in my project and Astra xhigh used 43% fewer steps and 49% fewer tokens than high on a 14 million token coding task. This is mainly saving input tokens because it achieves the task with fewer turns. Outcome quality was equivalent in my test.
So I’ve been using Opus-4.8 in Claude Code the past couple of days after running out of Fable credits. It’s just so less capable of the holding the wide variety of build threads and context that Fable does. And Fable KNOWS me and my goals somehow much better. I spend a ton of time getting Opus to generate markdown docs keeping track of these threads and using Sol for input.
Spent an hour this morning preparing Opus to perform a multi-threaded six hour build and test research session while I’m in the office today. No PR’s can be merged! Unlike Fable I started trusting it to merge some types of PRs if they passed several gates and hurdles. I am looking forward to reading the entire chat thread and findings when I get home about deep Monte Carlo optimizations
Had to switch my Claude Code orchestrator from @AnthropicAI Fable over to Opus-4.8 this morning due to maxed out Max account. Even 50% more allowance would have been enough for me. It just burns at such a high rate. But I cannot trust Opus as much and it struggles more.
I love my Fable orchestration platform. I ask a question about enrichment potential for features used to build a nested XGBoost nested Monte Carlo loop, and it’s like: dispatching two sonnet miners in parallel…. So cool, token saving, and works perfectly.
I want to keep using Fable @AnthropicAI after Sunday. Please don’t execute the planned rug pull.
Here's a follow-up from my earlier post asking Fable to aggressively conserve tokens.
I told it to route work to specialized sub-agents instead of doing everything itself:
• Fable → specs, synthesis, architecture, judgment
• Codex → coding, execution, testing, code review (on the OpenAI budget) using installed CLI directly
• Sonnet 5 → documentation sweeps, research & design reviews
It found this nice token saving workflow itself:
* Issues as shared agent memory: every Codex task reads the GitHub issue for context and posts its findings back there, letting agents chain work together without re-explaining the problem.
I started the day with only 4% of my weekly Fable allotment left with a reset happening this afternoon. I asked Fable to be “token usage aware” and to use tools I gave it (codex and sonnet-5 for specific types of tasks) to get as much as possible done. It came up with a couple of new ideas and we made several key breakthroughs in my codebase on that small allotment and queued it all up for this afternoon. Stoked! Will share the tips later today after I get back home from work.
Did a couple of epics with Fable and burned 50% of my full Max allotment and 80% of the one week Fable allotment in a day and a half. Most coding was done via sub agents using other models including codex. This model is great, though it’s not AGI. Token burn rate is too high @AnthropicAI!
Great ideas for using Fable here!
Start your engines...
10 Prompts I'll Run Immediately When Fable Turns Back On
danielmiessler.com/blog/prom…
So many @claudeai code API outages recently. More than I ever experienced before. Hope it can get resolved.
I’ve been working on a finance related project I wanted to launch for early beta in a few months. With Fable launched yesterday I feel like I have to move so much faster due to the two week window Anthropic said they will allow access to Max customers. Making every sprint count now. 😰
You can talk to Madison right now.
Ask him about executive power, privacy, or citizenship.
seaberger--foundry-chamber-g…
Thank you @modal and @charles_irl -- credits from AI courses you supported gave me the A100 compute to start this project.
