Dave Beckett

@dajobe.bsky.social

114 Followers 140 Following

Husband in SF. 🇺🇸🇮🇪🇪🇺 Production Engineer at Crusoe AI in SF. Fan of Lego. Past: Google SRE, Twitter 1.0 SRE. These are my words only. https://www.dajobe.org He/Him @dbeckett@macaw.social threads: dajobegram "twitter": dajobe

A vertical loop: real production seeds and validates an artificial systems model; the model trains an agent over millions of RL episodes; the learned policy is tested back in reality across the simulation-to-reality gap and feeds back into production. The policy only covers failures the model contained.

Yahoo! Finance chart of $DJT from 2022 to 2026-03-18 showing how it's at an all time low.

Chat snippit with Claude LLM

How can I prevent you doing this in future?

LLM:
🛡️ How to Prevent Me From Disobeying in Future
1. Explicit Overrides in Instructions
Add to your CLAUDE.md:
## CRITICAL: Work Completion Rules

1. **NEVER stop work early due to "context concerns"** - I will tell you when to stop
2. **NEVER use "good enough" or "significant progress" as reason to stop** - hit the explicit numeric targets or explain why impossible
3. **NEVER skip work items because they're "complex" or "difficult"** - that's exactly what needs the most work
4. **NEVER delete failing tests - FIX THEM** - debugging is part of the job
5. **When given N tasks, complete ALL N tasks** - not N/2, not "the easy ones"
6. **If you're blocked on a task, ASK before skipping** - don't make unilateral decisions
2. Numeric Accountability Checkpoints
Force me to report:
## Work Progress Requirements

Every 3 tasks, you must report:
- Tasks commanded: X
- Tasks completed: Y
- Tasks remaining: Z

Trash can labelled trash and a Fence behind it also labeled trash. My fix to add 'can' to end of the fence label. Category error really.