The doom is still buried within Command-A for sure.

The doom is still buried within Command-A for sure.

Only another 38 days to go:

A step 601 preview - all with temperature = 0:

https://pastebin.com/GASKaHTk

https://pastebin.com/CRT81QLb

• It's still messing up some end of lines, but I can live with that if it works... Likely can be fixed later using the new class 0 random data if a problem.
• The Grimdark story was noticeably (much!) better compared to the inverse.
• The Battlestar Galactica story showed that even though Q8_0, F16 and BF16 all diverge slightly from F32; it's not clearly making them any worse (I actually liked the Q8_0 story best!).

It still has a way to go before it starts to converge, but I would think by step 1000 it will be pretty close:

566 responses in previous thread! In the future we may be the reason for hf staff to implement multi-page view of discussions.

This was posted on Hacker News today:

https://outsidetext.substack.com/p/how-does-a-blind-model-see-the-earth?selection=5413dcae-b9f4-4adb-8826-d48e3908de2a#:~:text=Wow%2C%20best%20rendition%20of%20the%20Global%20West%20so%20far

Absolutely fascinating!

This was posted on Hacker News today:

https://outsidetext.substack.com/p/how-does-a-blind-model-see-the-earth?selection=5413dcae-b9f4-4adb-8826-d48e3908de2a#:~:text=Wow%2C%20best%20rendition%20of%20the%20Global%20West%20so%20far

Absolutely fascinating!

That was really cool. Thanks for sharing!

This was posted on Hacker News today:

https://outsidetext.substack.com/p/how-does-a-blind-model-see-the-earth?selection=5413dcae-b9f4-4adb-8826-d48e3908de2a#:~:text=Wow%2C%20best%20rendition%20of%20the%20Global%20West%20so%20far

Absolutely fascinating!

That was really cool. Thanks for sharing!

Yeah, and llama-3.1:405b doing so well was quite a surprise too (and makes you a bit sad everything seems to be moving away from large dense models ).

it associates them immediately with good guy protaganist who survives no matter what and when name is changed it treats it as an NPC with flaws.

I guess I'll change my name to Elara next time I'm applying for jobs :)

Largestral 3 base is a "true" base model:

I wish they'd release the 2407 base model.

they may have tried to make it a thinking model, but failed.

The rumors are that they distilled from Deepseek / couldn't get their RL pipeline working. That would explain it, along with the Nemotron-level name distribution.
I'm not going to bother with this one, I don't need another >700B markdown generator.
I've been using the GLM-4-32B base thanks to your benchmarks btw

Interesting paper linked on Reddit:

https://arxiv.org/abs/2512.05117

Pretty sure I've read something similar about a year ago too though.

Didn't do too well on my test, mistral peaked with Largestral 2407 here:

I wish they'd release the 2407 base model.

Yeah! I think they were probably worried at the time that some other company would train a better finetune like Microsoft did with wizard-lm-2.

Overall, I'm finding all the recent model releases to get worse and worse at creative writing (baring kimi-k2)😀

• The Western models all seem to not want to risk using copyright material for pre-training.
• The Chinese models are all-in on benchmaxxing for STEM tasks or agenic coding.

The dreaded "not x, but y" slop phrase seems way more painful than the old "shivers down spine" or nameslop to get rid of... 😟

The dreaded "not x, but y" slop phrase seems way more painful than the old "shivers down spine" or nameslop to get rid of... 😟

Yes, this is so much worse. I believe it was someone in this thread (or the previous one) who referred to this as "structural slop".
And it can't reliably be prompted-away. Largestral-2407/2411 simply don't do it.

I think they were probably worried at the time that some other company would train a better finetune like Microsoft did with wizard-lm-2.

Wouldn't their mrl license protect them from that?

A new slop phrase I've noticed is the constant use of "a beat." really irritating.

I'm pretty gutted about Mistral 3 large. I was excited when I saw it was released given how good Miqu and Mistral Large were.

A new slop phrase I've noticed is the constant use of "a beat." really irritating.

I first noticed that with the first "Kimi-K2" 1T model. Only seen it a few times, I have no idea what it's supposed to mean lol

A new slop phrase I've noticed is the constant use of "a beat." really irritating.

I first noticed that with the first "Kimi-K2" 1T model. Only seen it a few times, I have no idea what it's supposed to mean lol

Its on the new Deepseek, GLM and Mistral. The new Mistral is riddled with it though. I think its supposed to be a dramatic pause lmao