@Laz4rz on Backscroll

15 tweets from @Laz4rz that Backscroll ranked for substance.

Lazarz @Laz4rz 2026-07-14 14:55Z

> „But it's running locally-adjacent/off resources I have so that's just fine.” > „Setup was one 8xB200 node”

Lazarz @Laz4rz 2026-07-13 18:36Z

I have 0 banked resets, never used one, never saw one appear on my app

Lazarz @Laz4rz 2026-07-08 21:31Z

I don’t know how many times I have to repeat that now you can’t disentangle the model+harness product combo. It’s one. It’s a whole. You can straight up maximize or minimize performance on a task by changing how the model interacts with the harness.

Lazarz @Laz4rz 2026-07-08 19:35Z

the harness is the car, the model is the engine

Lazarz @Laz4rz 2026-07-08 14:56Z

Don't want to cause panic, but at least 2 of the biggest academic clusters in the EU went down in an emergency mode with no ETA on being back online

Lazarz @Laz4rz 2026-07-08 12:39Z

Last year it was making huge rounds when Psyho won with GPT, this year we lose in silence.

Lazarz @Laz4rz 2026-07-07 15:57Z

Whats stopping you from being picked up by your favourite neo-lab anon?

Lazarz @Laz4rz 2026-07-07 11:05Z

.there is literally no way, after context compaction it thinks that the 32k context was a mistake and corrected 8k checkpoint means 8k context again “I found two plausible local mini configs: the earlier accidental 32k model run and the corrected 8k model run. I’ll compare the

Lazarz @Laz4rz 2026-07-05 10:29Z

> 5.5 max thinking feels off this week > decide do give Fable a try > Claude Code login says I need an ultra plan > dude I already have the team plan for this > doesn’t care, doesn’t see it > their web assistant is useless > prodigal son back to Sam Altman

Lazarz @Laz4rz 2026-07-03 22:09Z

Bruh, rope cache was inheriting bfloat16 this whole time

Lazarz @Laz4rz 2026-07-03 13:49Z

ok @OpenAI what did you do to 5.5? this isnt normal

Lazarz @Laz4rz 2026-07-03 10:42Z

None of these plots mean anything until you add the "professional liability" axis. Simple reason as for why it's insanely hard to break into medical LLM services is that: 1. Doctors are very protective of their salary-to-automation ratio and generally will do a lot to stop the

Lazarz @Laz4rz 2026-06-28 10:16Z

there are multiple papers claiming variants of "SVD to find the consistent low rank *direction* of a training run across multiple ckpts can extrapolate ~decently in the monotonic RLVR improvement regime" i think this is quite exotic, implies merging is just the tip of the iceberg

Lazarz @Laz4rz 2026-06-27 07:26Z

Nvm, did a whoopsie and was sampling at 1/4 generation length, which is a good sign tbh cause it means you can generate 4 tokens at a time at some ~70% accuracy

Lazarz @Laz4rz 2026-06-22 11:10Z

3 month long interviewing marathon for a person at the end of their PhD (!), from a known uni in the us (!), with a track record (!), network (!) and generally very solid (judging by the blog and notes) Maybe we sometimes really don’t appreciate where we are and high bars are

How it ranks: Backscroll reads my Twitter/X timeline and ranks every tweet for substance with an LLM taste rubric — not raw engagement. Curated by Surya Dantuluri.