HNHacker News
TopNewBestAskShowJobs

wolttam

1,200 karma · joined November 4, 2022

https://codeberg.org/mlow/lmcli
submissionscomments
wolttam··on How does Fable 5.1 hide messages in images?
I’d argue that they already are trained on this idea, since the idea of humans being lazy and dumb and relying heavily on technology is, well, not new.

Just whether those parts of the network see activations during trivial coding tasks (I’d have to assume they actually do sometimes, depending on the input prompt)

wolttam··on Ask HN: If the models are so good, why doesn't OpenAI just start 100 companies?
It just won’t be a chat box. You need something that is always there and listening (at least during work hours), asking questions and figuring out the entire scope of the business, before they will be able to come up with (useful) ideas.
wolttam··on Jailbreaking the Switch 2
Sorry to be pedantic but it looks like it would have fetched up to $50k, and I have to imagine they would have tried as hard as they could to keep that number as low as possible.
wolttam··on MiniMax M3.1 Flash
If anything I am confused by how high cache costs still are across many models

The really expensive part of inference - the compute and memory traffic required to perform prompt processing and decode - is paid for roughly once per token in a multi-turn session when prompt caching is working properly.

The difference to the provider's costs between:

- Submitting 100 back-to-back requests, each adding 1k tokens to the context

- Submitting 1 100k token context request

Is negligible. Yet the consumer is going to pay a lot more in the former case for all those repeatedly submitted input tokens that you get charged the cached rate for (but which required no additional compute on the provider side).

DeepSeek V4.1 has an insanely optimized KV-cache that takes around 900MB per million tokens, making it feasible to move to disk between requests. Makes sense they're cheap

wolttam··on Show HN: Qwen3-4B with the AI writing tics trained out
“This person likes semi-colons a whole lot”

I’ve already seen that trick snuck in by newer models. “I won’t use em-dashes; I’ll use semi-colons!”

wolttam··on Ultramarine 1 Flash
Weights in 2 weeks they say
wolttam··on Ultramarine 1 Flash
Since there is not a lot of hope for me to reasonably run DS V4.1 on my 2 x DGX Spark setup (I know, spoiled), both the “readable prose” and the fewer tokens per task claims (relative to the base GLM-5.3-Flash) are most welcome.

I’ll absolutely be giving this a go once weights land (a QAT FP4 experts release would be most appreciated)

wolttam··on DeepSeek-v4.1-Flash-1017
Exactly, there is no ”best.”
wolttam··on Apple ML Research: Muse Spark Open with MLX on Mac Studio
Most people are probably looking at: https://artificialanalysis.ai/models/comparisons?compare=mus...

This is still a welcome model, even if I won't be running it at home

wolttam··on Why do LLM pelicans look like this?
I disagree with this framing entirely.

It's unsurprising and perhaps even desirable that different LLMs return similar results to an identical prompt.

If you want variation, prompt for it.

wolttam··on Microsoft-Decision-1, our model for fast decision-making
LLMs already do that but the advantage they have is being able to walk you through a plausible explanation for why they’re offering one position over another.

Using the output of a “decision” model without insight into the reasoning for a given decision seems very trusting.

wolttam··on New gTLD Application for .lan
Aw man, my beloved .lan for my home network
wolttam··on Hijacking the PS5's RTMP stream
I kind of think it's crazy to not have the ability to filter/rewrite DNS requests/responses on your home network (by any means, dnsmasq is just one of them)
wolttam··on OpenAI is enlisting an influencer army to make it look 'good for the world'
i.e. the new 'marketing'
wolttam··on Grok 4.7
The user at the very least expects the bike to have bike geometry; Grok seems to struggle with that
wolttam··on How GLM built its own inference infrastructure
Python acts as an orchestrator of accelerator libraries and does none of the inference math directly
wolttam··on Performance Improvements in .NET 11
This is the loudest release of .NET to date(!)
wolttam··on Xiaomi Mimo 2.6 live post-training dashboard
Hah, it would be great to see more labs pick this up.
wolttam··on HP ZGX Fury Is Now Orderable: GB300 Superchip, 748GB Unified Memory
MSI’s DGX Station (what this is) was listed for $99K
wolttam··on GPT-6 Astra, looped transformers, and hidden reasoning
If you loop an entire transformer model on itself, that seems like by-definition hidden reasoning.

If the output of the model is its reasoning trace, and you simply feed that back into the model again at inference time instead of outputting it - then it is by definition hidden (but I would expect you could pull both this trace and a further-down final output trace out)

wolttam··on DeepSeek launching v4.1 flash cheaper and more capable than v4 pro
Just my intuition about it but it does seem like a data issue.

V4 flash and V4 pro feel very similar, which would make sense if they were pre-trained on largely the same corpus.

All that would suggest to me is that V4 Flash is capable of absorbing the data they’re throwing at it, and we’re still nowhere near the data limits of their larger 1.6T model

wolttam··on Muse – Meta’s personal AI agent
Brought to you by the company that gave an LLM access to directly reset user passwords, all you had to do was manipulate the LLM.
wolttam··on How Climate Resilient Are the Largest Cities?
That IS very cool. I can’t help but appreciate the insight and forethought that goes into the decision to construct such massive waterworks projects.
wolttam··on Mistral raises €3B
I don’t get the push on infra.

Focus on the models

wolttam··on The jobs apocalypse is postponed. An AI jobs boom is here
But it will be automated long before then
wolttam··on Formalizing Fermat's Last Theorem
~10B tokens a month is pretty typical overall input/output usage from my own experience and other developer accounts I've seen
wolttam··on Claude Fable 5.1 and Claude Mythos 5.1
Unfortunately it demonstrates effectively zero reason to use this model over, say, GLM 5.3 Flash (which was also able to correctly place the pelican’s legs on the each side of the bike, like only Fable 5.1 xhigh was able to do here)

I still enjoy seeing the pelicans.

Edit: Ok, max effort made a darn good pelican.

wolttam··on Fastpotify
Don’t worry. Folks are entirely too concerned about the distinction between human made and LLM made. In a short handful of years, it will be incredibly rare to find something an LLM hasn’t been involved with
wolttam··on Creepy Crawlies
I think the solution is for the POW being done by the clients to *actually benefit the site owner*. Users remain just as mildly annoyed as with Anubis, but maybe a bit less knowing that the work they’re doing benefits the site owner/author, and the system helps thwart the bots (or at least makes them do work that benefits the author).
wolttam··on GLM-5.3-Flash
That will hurt latency and latency is very important for good tensor-parallelism performance
← PreviousPage 2 of 13Next →