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 ).

Yeah, but I think the point about modern LLMs all being much less "guidable" is spot on though.

I first noticed this with miqu:70b and it's got worse and worse since then... Perhaps the problem is the long-context extension data causing this? I remember some of the older 4k-context models would take such crazy and creative turns in stories, but the recent models all seem to have some sort of "boring/safe" attractor state that they want to converge towards ☹️

PSA - If anyone else here had > 1TB free private models/datasets "grandfathered in" from before the limits a year ago, be aware it seems like they've just removed that and set it back to 1TB for Pro users.
(Just means a higher bill or time to purge I guess)

This looks an interesting model:

https://old.reddit.com/r/LocalLLaMA/comments/1puv3de/k2v2_70b_and_creative_writing/

https://pastebin.com/YBwTE8Be

Is it just me or is new GLM 4.7 kinda meh?

Merry Christmas, by the way!

Is it just me, or is the new GLM 4.7 kind of meh?

Merry Christmas, by the way!

Merry Christmas, Chuck!

If you like ambient/immersive details, it does so in spades. You're going to learn about the grain pattern, age of the lacquer, and history of the wooden table in the corner of the room. I would say it writes marginally better depending on what you like.
However, my refusal tests actually had refusals. Haven't had that with 45 or 46. It spent over 4,000 tokens debating with itself over whether to respond at one point. Seems easy to overcome, but never had to before.
My biggest subjective test with any new model is whether I actually want to play through my old scenarios again. The answer for me is I didn't see enough delta from 46 to care.

@BigHuggyD @ChuckMcSneed

Is it just me or is new GLM 4.7 kinda meh?

I haven't had a proper chance to play with it creatively yet, my rig is busy training control vectors for it. For coding/work (via ClaudeCode and OpenWebUI), it's a huge step-up vs GLM-4.6 for me. It gets things right the first time more often.

P.S. I saw on Reddit, apparently the API has those prompt injections like Anthropic do, steering away from intellectual property. So API testing is probably pointless.

It seems to produce more Claude-3 slop rather than Gemini slop this time, though that's my "vibe check" and the slop forensics bench will reveal that more I guess.

I also saw a "screenshot of a Reddit thread" where Z-AI said they're targeting Roleplay/Sillytavern!

You're going to learn about the grain pattern, age of the lacquer, and history of the wooden table in the corner of the room.

lol GLM-4.6 is like that as well. Control-Vectors worked well for this with GLM-4.6

My overall impression is that it will need more nuanced prompting than GLM-4.6 and I suspect I'll probably be swapping between this, and Devstral-2 for a while.

Merry Christmas guys!

P.S. (I read this and confirmed it myself), some of the Unsloth quants are pretty bad and consistently trigger refusals, while others don't.