BoorishBears

It's as hard a problem as coding: OpenAI could put all their resources in it and still be in dimensionally the same place as coding is today (much progress but not entirely solved).

So diverting significant resources is just going to harm the main goal and kind of suck at the side goal too.

4.5 is a great example of this: it wasn't perfect but it was better than every 5 family model at things like tone and writing, and despite being priced extremely ruthlessly and already finished training, they couldn't afford to tie up the resources to host it.

It's no mystery: solving LLM voice requires making the model worse at other things they value above all other things

They can't clarify it.

If you say the quiet part out loud, the lemmings will take that as approval.

"Only thoughtful application of AI allowed" -> "Of course my usage of AI was thoughtful! My custom pi harness with 1M context project brain is very thoughtful!"

Not sure why you think this wouldn't be something they did on principle either though. Most open source projects are built on someone's personal principles: that's what motivates them to keep working.

(hence the recommendation someone's free to fork it and do the work themselves)

That's not my read, at least not if you give them enough credit to use their software.

My read is any LLM use should be opaque to them. You should be able to explain your bug find in your own words, have formatted as the project expects, not have any LLM generated comments (which trend towards being useless anyways)

The wording of the post is very clever, if you're green you read it and you say "I'm banned because I use Claude". If you're experienced enough you can read the message between the lines: "if you use LLMs, it should still have gone through your brain before it reaches mine, so I'm not turned into a meat puppet by proxy"

Argon is the best writing model that's been released in a while

It's sad to see Carbon releasing at the same pricing because that implies Google is planning to do the thing everyone is doing: release promising model for subjective preference, RL-maxx a V2, replace promising V1 with V2 because it benchmarks better

Muse did it, Kimi did it, Deepseek did it, Tencent did it. No one seems to care abojt subjective preference anymore.

They have unwatermarked text from the same base models, human text, and watermarked text.

If they couldn't train a reliable detector with that they wouldn't exist.

If Pangram was snake oil (it's not but could be more mindfully marketed), it's definitely not now that every major lab has added watermarking.

You wouldn't even need to be able to read the information the watermark encodes to recognize it exists in a piece of text

My last shred of skepticism has been killed watching people try to imply this doesn't prove Anthropic (and likely OpenAI) were actually much better off than most people assumed and will certainly stay multi-trillion dollar companies

I'm surprised no one has mentioned online: most people don't want to risk console bans, and that immediately makes jailbreaking their Switch a non-starter

At this point most people getting a 3DS are probably doing so specifically to jailbreak it

There 3090 is from the era of overclocking, anything but the reference design should be fine even on air cooling (especially if you have a 3 connector model)

My EVGA came with stock firmware that started at a 420W and allowed up 500W, of course 500W did nothing for gaming

We reached inverse hype: where the folks jaded by hype are so jaded that they start to hype the same low value content they normally deride... as long as it speaks counter to any hyped topic.

Ironically, that technique usually pans out since a lot of useless fads get hyped. But every once in a while it causes people to completely and utterly miss the point.

Have you tried SFT to guide it early to more efficient trajectories? I know it hurts generalization and I've seen folks claim that it hurts final performance, but for my limited budget I literally will not do rubric based RL without SFT to shape the thinking to the rubric anymore.

It's so much more compute efficient for my goals, though I do deal with subjective preference which makes exploration during training much harder

Maybe this is me being the wrong persona since I clicked for the model and not the platform, but literally the moment my eye caught

> Agent observability is crowded, so instead of a bullet list of how Laminar differs from X, we’ll explain how we’re solving agent observability from first principles

my brain shut down and bounced minus the comment because it's a launch and maybe you'd want feedback.

I'd be using this for post-training in a custom harness: so what you focused on during post-training and what it excels at are the most interesting things