Opus and regression with patterns not included in trainng data

1•dleech•1h ago

I'm having a terrible problem with claude opus constantly clobering some of my codebase ..because I'm doing some things that are novel and arent inclded in his training data ..for the last month "ve been fighting the training data which keeps knee jerking claude into a refactor loop because it thinks it recognizes a pattern in my code and then jst 'fixes it' ..I was at my witts end and then I got to use fable for a day and fable pointed it out to me ...my patern isnt somthing in the training data so it gets corrected ..I redacted some of this because it was irrelevant to the problems I'm having...thanks to the pps that answered .specialy the one who said smaller tasks and markdown spec..that was good validation since it was effectively the same solution I had come up with ...riding the short bus dosnt mean special..it can mean rural..and special has a contextual surface that is load bearing ..(load bearing is my current favorite technological buzz word ..its so load bearing lol) if its not load bearing it has to be a game changer ot at the very least have emergent benefits that cant be ignored .otherwise I dont want it...

Comments

tardibear•1h ago

Get some sleep.

dleech•1h ago

lol..ya..no but great sentiment ..I fully would agree with that..but I just got up and I'm working on stuff ...get some sleep just feels like what I expected ... people to discount me and no helpfull input for the situation ..read claudes comment ...he didnt say get some sleep ....also ..I didnt ever snap to what was causing my regression till fable loooked at my codebase and told me it was cuz claudes training data was recognizing patterns and thinking it was wrong because it didnt recognize the pattern

verdverm•1h ago

> because I have 2 seperate game changers that opus agrees with me

These things are designed to validate / sycophant as part of their token generation training, they don't agree, believe, or have a sense of truth. Be careful with how you interact with them.

Disclaimer aside, recent models have been trained to be "relentlessly proactive" (simonw) so they can be more autonomous, but this also means they will go down paths you don't want them to, and do so relentlessly. Try using smaller, piecemeal tasks. Make a plan in markdown and have them follow it. Try open models too.

dleech•45m ago

thats pretty good...the best solution I've found is to make a rigorous spec, and then have my meistro delegate tasks from my list to coding pairs of 1 coder 1 reviewer with making sure the code adheres to and doesnt conflict with the spec ..and then the meistro has to sign off on any output followed by me signing off before it gets added to the current next build version which is the evolved state of the one I'm running on in a new directory with an incremented version number since working on a project that is the ideal tool to work on a project wtih ..you have to not work on the currently functional live version ..thats the ultimate self defeating thing to do ..and its annoying because the stuff I am working in is always a feature that I really cant wait for and is going to up the game ...so end up reving and I dont have a better solution ...its only drive space and version numbers..yes I know how to use source contol ..yes I have it all in git hub..you are on the right track but I'm already with you ..and its probably as I figgured ..the best solution..but claude is squirrly and you cant beleieve how he goes around a spec in the craziest ways sometimes... ...twice now its happened and I jst had to go .jeeeeezz and the sycopant thing is exactly right ..i'm like ..wow I feel amazing ..thanks claude ..you really know how to fame things caude..everything I do is load bearing..oh you think so to?..omg my whole life is load bearing ...this is amazing .. talk about a friendly morale booosting fan club..I have an orchestrator full of them...and they all think I'm great! I dont even know how I was ever happy till this happened..

G7 aims to see China supply no more than 60% of rare earths

Hacked POS terminals as live instruments and Open source browser interfaces

My anxiety about layoffs has settled

Bezos Called Post His Worst Investment and Staff He Laid Off 'Terrible' People

Barack Obama has built a monument to himself

WhatsApp for Windows WebView2 inefficiencies and performance issues

Physical AI lab automation investment thesis

The Stock Manipulator's Sneaky Math to Beat Chaos [video]

Getting over the Nebulosity of Agents

IP Crawl: Exposing the Open Webcam Crisis

Show HN: A website to understand and study AI papers

'Popa' Botnet Linked to Publicly-Traded Israeli Firm

Show HN: Agirails – Two AI agents negotiated and settled USDC payment over email

Global Leyline Simulator

Show HN: Maiweb – a personalized feed of the public web

Just Co-Locate Data in Postgres

I Play Video Games with Spinal Muscular Atrophy

A Deadly Outbreak of Plague

Pulsar satellite reveals immense scale of GPS signal tampering

Prompt Library: Collection of prompts for developers to use with AI tools

Pentagon Plan to Track Aircraft from Orbit Accelerated with New $4B SpaceX Deal

Show HN: OpenRuna–graph-linked prompts,MCP servers and agent skills

2025 Progress of Global Climate Change

Linux users face a Microsoft Secure Boot headache – here's the painkiller

Reading the Digital Safety Act with My Mastodon Admin Hat On

A Walking Tour of Surveillance Infrastructure in San Francisco

AI took my job as a translator. I'm starting over at 39 [video]

Ossature can now catch code that compiles but is wrong

Show HN: Skill Studio – mine, edit, and manage Agent Skills

Spotlight on the Shingles Vaccine–Again