Vladimir

load-bearingly delve-adjacent

[Vladimir @Vladimir76897]

All models do this, and it’s extremely annoying. Instead of just removing the thing, they remove it and then write a sentence that they removed the thing I didn’t want there anymore. Something is not going well with these extended RL runs they’re doing

[Duca @big_duca]

Me: Please remove this component from the UI.

Fable: done.

*checks code*

Wtf

[Vladimir @Vladimir76897]

@jcubic @0xAX And it’s written by AI

[Vladimir @Vladimir76897]

@CardilloSamuel That’s a kernel bug

[Vladimir @Vladimir76897]

@JoesInvestments Here it is, for you and your clanker. https://t.co/q6m5HZ9uZk

Do note I run it on a 4 ring configuration, without a switch

[Vladimir @Vladimir76897]

@aijoey @Web3Twon @QwenDevs @Alibaba_Qwen @NVIDIAAI @sgl_project @huggingface I need to get my glasses. Thank you

[Vladimir @Vladimir76897]

@Hamburgerai https://t.co/4oqEddsHss here it is, my friend.

I did not get the paper background to work, however. Training with that just taught the lora to generate only the background. This is more like a filter.

[Vladimir @Vladimir76897]

@relizarov I’m dealing with this now. It does not care at all about the wrong levels of abstraction it uses and I have to ask it to kindly stop inventing data structures and states that do not/should not exist. Feel like a quicker technical writer mostly. GPT 5.6 Terra max

[Vladimir @Vladimir76897]

@TheZvi No wonder

[Vladimir @Vladimir76897]

@spark_arena @TheAhmadOsman If not inference , then what? Definitely not training either

[Vladimir @Vladimir76897]

@Hamburgerai Thank you. I'm training a LoRA for Flux2-klein to get this effect with local models

[Vladimir @Vladimir76897]

Somehow my Grok SuperHeavy discount subscription has vanished (?!) and I'm back on the X Premium usage limits... oh well

[Vladimir @Vladimir76897]

@unclebobmartin @CognitiveTake @danpdc Have you tried Kimi Code?

[Vladimir @Vladimir76897]

@unclebobmartin @ericzakariasson Somehow my limits have also vanished and my account no longer shows I have the SuperGrok Heavy discount. Yesterday I was doing tons of work with it and now it takes one turn to run into my limits

[Vladimir @Vladimir76897]

I would like to thank Grok 4.6 for wasting 6h of LoRA training time by doing text2img instead of img2img, the way I explicitly asked it to

[Vladimir @Vladimir76897]

@Hesamation They train LoRAs all the time and bring them up ASAP. They’re cheaper to train than the entire weights. That’s likely why you’re seeing improvements so quickly

[Vladimir @Vladimir76897]

@NVIDIAAI @pinchbench @openclaw Lightning got me into training models

[Vladimir @Vladimir76897]

You are absolutely right. I indeed dropped your entire database to make the query faster. That is on me. Let me check whether we have any backups.

[Matt Pocock @mattpocockuk]

A really great loop idea is:

- Database alerts you there's a slow query (@PlanetScale does this)
- Opens an issue in your codebase
- Immediately schedules a PR from an implementer
- Reviewer agent reviews the PR
- Pings the human for review

Repeat ad nauseum until app is fast

[Vladimir @Vladimir76897]

@ajambrosino Who gave Opus 5 an X account?