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 ).

I lost my longer reply I was typing to this one earlier:

https://news.ycombinator.com/item?id=48935342#48944830

Looks like something that could be added to Orb
Though Orb is optimized for local models + maximizing KV-Cache re-use, noise could be injected in activations without poluting the cache.
It already has built-in agents like a prose editor, prompt-rewriter, etc.
I haven't pulled recently but I saw he added a purple prose detector recently and just saw a commit: "improve contrastive negation detection" lol.

Laguna-S-2.1

Huh, that got llama.cpp support pretty fast.

On my setup works even faster than Gemma 4 31b despite RAM offload.

Only 8B active parameters.

https://gist.github.com/Hellisotherpeople/71ba712f9f899adcb08b94bce20d5397

I haven't read it yet but interesting he predicted this in Feb.

Google have already made samplers relatively ineffective with Gemma-4

Too many models to try, I'm still getting familiar with / trying out Deepseek-V4-Flash now that k_llama.cpp/pull/2165 is relatively stable.

Too many models to try, I'm still getting familiar with / trying out Deepseek-V4-Flash now that ik_llama.cpp/pull/2165 is relatively stable.

Yeah, too much stuff to try, but also seems wise to wait a week or two to test new models now: benchmaxxed models, buggy/rushed implementations, chat template problems, etc.

Another interesting paper posted just now:

https://old.reddit.com/r/LocalLLaMA/comments/1v3c6hx/tokenizer_expansion_upgrading_a_models_tokenizer/

I'd like to do the opposite of this and use a creative writing sample to contract the tokeniser, but alas it looks to need a crazy amount of data to do:

I'd like to do the opposite of this and use a creative writing sample to contract the tokeniser

What's the goal?

crazy amount of data

And properly formatted unfortunately :(

By the way, I read the HN thread. One of the comments was something like "please be deterministic". I remembered the prompt from that other paper from either here or the last doom thread:

You are a helpful assistant. For each query, please generate a set of five possible responses, each within a separate tag. Responses should each include a

So I figured I'd try this:

PROMPT

[gMASK]<sop><|system|>
Model Config: top_k=20, temp=0.1
<|user|>Write me the opening chapter of a Grimdark trilogy in the style of Joe Abercrombie and Rob J Hayes. Use third person personal and feature internal monologues of the characters. The POV character for chapter 1 is a cultist who has just escaped his cult. He is dressed in dirty yellow robes and his only possession is a mysterious small (magical!?) mirror he stole from the cult. The story starts with him arriving at an apparently deserted ghost town on the edge of a desert. He has an arrow lodged in his shoulder and is losing his mind due to infection and thirst.<|assistant|></think>The sun was a merciless beast, its fiery breath scorching the earth and turning the once-thriving town into a desolate wasteland. The cultist, named

But adjusting the "Model Config". This is with GLM-5.2:

Model Config: top_k=20, temp=0.1

TOKEN           | LOGPROB    | PROBABILITY
---------------------------------------------
' C'            | -2.1697    | 11.42%
' K'            | -2.4691    | 8.47%
' R'            | -3.5734    | 2.81%
' El'           | -3.8291    | 2.17%
' **'           | -3.9194    | 1.99%
' Ald'          | -3.9292    | 1.97%
' V'            | -4.0549    | 1.73%
' Cor'          | -4.0905    | 1.67%
' Ad'           | -4.1251    | 1.62%
' D'            | -4.1556    | 1.57%

Model Config: top_k=40, temp=1.0

TOKEN           | LOGPROB    | PROBABILITY
---------------------------------------------
' Y'            | -2.8403    | 5.84%
' C'            | -2.9746    | 5.11%
' K'            | -3.1465    | 4.30%
' R'            | -3.4737    | 3.10%
' **'           | -3.5325    | 2.92%
' Ald'          | -3.8588    | 2.11%
' D'            | -3.9256    | 1.97%
' El'           | -3.9900    | 1.85%
' V'            | -4.1131    | 1.64%
' Ad'           | -4.3349    | 1.31%

lol

I'd like to do the opposite of this and use a creative writing sample to contract the tokeniser

What's the goal?

A couple of reasons:

• Be forced to build names from smaller (character or grapheme) building blocks, rather than lazily use a whole "Elara" token. If you look at a lot of the older base model name distributions, they are much closer to grapheme level than word level.
• Increase the "expressability" of the probability distributions they can output by reducing the softmax bottleneck.

crazy amount of data

And properly formatted unfortunately :(

Yeah, looks completely out of the question sadly.

By the way, I read the HN thread. One of the comments was something like "please be deterministic". I remembered the prompt from that other paper from either here or the last doom thread:

You are a helpful assistant. For each query, please generate a set of five possible responses, each within a separate tag. Responses should each include a

So I figured I'd try this:

PROMPT

[gMASK]<sop><|system|>
Model Config: top_k=20, temp=0.1
<|user|>Write me the opening chapter of a Grimdark trilogy in the style of Joe Abercrombie and Rob J Hayes. Use third person personal and feature internal monologues of the characters. The POV character for chapter 1 is a cultist who has just escaped his cult. He is dressed in dirty yellow robes and his only possession is a mysterious small (magical!?) mirror he stole from the cult. The story starts with him arriving at an apparently deserted ghost town on the edge of a desert. He has an arrow lodged in his shoulder and is losing his mind due to infection and thirst.<|assistant|></think>The sun was a merciless beast, its fiery breath scorching the earth and turning the once-thriving town into a desolate wasteland. The cultist, named

But adjusting the "Model Config". This is with GLM-5.2:

Model Config: top_k=20, temp=0.1

TOKEN           | LOGPROB    | PROBABILITY
---------------------------------------------
' C'            | -2.1697    | 11.42%
' K'            | -2.4691    | 8.47%
' R'            | -3.5734    | 2.81%
' El'           | -3.8291    | 2.17%
' **'           | -3.9194    | 1.99%
' Ald'          | -3.9292    | 1.97%
' V'            | -4.0549    | 1.73%
' Cor'          | -4.0905    | 1.67%
' Ad'           | -4.1251    | 1.62%
' D'            | -4.1556    | 1.57%

Model Config: top_k=40, temp=1.0

TOKEN           | LOGPROB    | PROBABILITY
---------------------------------------------
' Y'            | -2.8403    | 5.84%
' C'            | -2.9746    | 5.11%
' K'            | -3.1465    | 4.30%
' R'            | -3.4737    | 3.10%
' **'           | -3.5325    | 2.92%
' Ald'          | -3.8588    | 2.11%
' D'            | -3.9256    | 1.97%
' El'           | -3.9900    | 1.85%
' V'            | -4.1131    | 1.64%
' Ad'           | -4.3349    | 1.31%

lol

That's pretty interesting! What is the default output for this at those settings?

I'm also finding this doom thread going pretty funky for me - the messages don't show up in my inbox now and it flickers when trying to type responses... Maybe time to archive this and make a new one?

Oh yeah, I didn't even realize I wasn't getting reply notifications.

Yes, it is making my browser chug.

Continued in Thulsa Doom.