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 ).

Oh. I thought it'd be proprietary, but here it is!

https://huggingface.co/collections/talkie-lm/talkie-13b

https://github.com/talkie-lm/talkie

A quick dump into an LLM suggests its not strictly compatible with any popular arch, but similar to Qwen2 or Llama 13B.

...This could be interesting to "combine" with a larger MoE, perhaps to rewrite responses sentence by sentence? A "desloppifier" if you will.

I tried to get it working in llama cpp by hacking to make it look like llama arch, but it has some weird scale values and the vocab didn't want to work ☹️

https://old.reddit.com/r/LocalLLaMA/comments/1sz1qer/mistralaimistralmedium35128b_hugging_face/

Actually horrible to talk to. The prose is pure AI slop, has an incredibly irritating "helpful but quirky" tone, and hallucinated over 6 times in a short "free" conversation that got capped after less than 10 messages. Even the web search failed multiple times. I tried it in LeChat and the Agent Mode.

Oh dear.

But for cleaning up https://www.sacred-texts.com/eso/index.htm ; you could set Gemma-4 with puppeteer on it.

Sorry, completely missed this yesterday - might be time for a new doom thread as this one seems to have gone funky and not showing in notifications 😟

What is "puppeteer"? Tried searching but seems several things related to LLMs and scraping...

I tried to get it working in llama cpp by hacking to make it look like llama arch, but it has some weird scale values and the vocab didn't want to work ☹️

There's already a couple of attempts, see:

https://huggingface.co/models?sort=modified&search=talkie

https://huggingface.co/sol-wy/talkie-1930-13b-it-q5

https://huggingface.co/dtestnyrr/talkie-1930-13b-it-gptq-int4

Even an Transformers conversion from Huggingface Staff: https://huggingface.co/lewtun/talkie-1930-13b-it-hf/tree/main

And this one, where they seem to have hacked it to Qwen3 arch, but didn't include whatever custom extra code is needed? https://huggingface.co/wave-on-discord/talkie_converted/blob/main/config.json

What is "puppeteer"? Tried searching but seems several things related to LLMs and scraping...

It's a nodejs library that lets you run a headless google-chrome. It was designed for test automatomation, but you can wrap it in an mcp server to let LLMs read and post online.

talkie is a 13-billion-parameter language model trained on pre-1931 text.
Create a headless puppeteer session and ask the model a few questions here: https://talkie-lm.com/chat
sleep for 20 seconds to allow the model to respond, then take a screenshot for me.

Oh dear.

That negative reaction to the model being dense :(
If they don't like it, nobody is forcing them to use it!

a short "free" conversation that got capped after less than 10 messages.

Looks like that's via Le Chat or something. Those issues are probably just the system prompt.
I'll wait for an ik_llama.cpp implementation before trying it myself.

What is "puppeteer"? Tried searching but seems several things related to LLMs and scraping...

It's a nodejs library that lets you run a headless google-chrome. It was designed for test automatomation, but you can wrap it in an mcp server to let LLMs read and post online.

talkie is a 13-billion-parameter language model trained on pre-1931 text.
Create a headless puppeteer session and ask the model a few questions here: https://talkie-lm.com/chat
sleep for 20 seconds to allow the model to respond, then take a screenshot for me.

I see - thanks!

Oh dear.

That negative reaction to the model being dense :(
If they don't like it, nobody is forcing them to use it!

Yeah, it's the same as the Reddit posts that are all sarcastic gripes about not being able to run the new 1T models on a pair of 3090 🙄

The biggest problem is not using any copyright data has likely killed the writing ability and doubtful if fine-tuners will be able to make anything useful out of it ☹️