Nafnlaus 🇮🇸 🇺🇦

Background: www.amnesty.org/en/documents... (Also, for the record, a lot of what these companies do is actually buy bulk books and scan in their legally-acquired copies, which is unambiguously compliant with copyright law)

Unlawful by design: Exposing the human rights costs of generative AI - Amnesty International amnesty.org

The "don't know but close to something we know" aspect is key: transformer-circuits.pub/2025/attribu...

On the Biology of a Large Language Model transformer-circuits.pub

To go one step up from there, I'd recommend this introduction to circuits: distill.pub/2020/circuit...

Zoom In: An Introduction to Circuits distill.pub

As a reminder, LLMs recognize when they're being tested and have a tendency to answer based on what they think an alignment-tester wants to hear (for example, "racism is bad", etc) - and swapping cues about the politics of the tester changes the outputs. arxiv.org/abs/2604.27633

Political Bias Audits of LLMs Capture Sycophancy to the Inferred Auditor arxiv.org

[[ Googles vast.ai ]]

Rent GPUs | Vast.ai vast.ai

I used to work extensively with differential evolution (still do), which faces a similar "cheating" problem, even though just purely through randomness, rather than reasoning (a far more capable thing than randomness). Many random examples: arxiv.org/pdf/1803.03453

arxiv.org

With abliteration, you use sets of queries that trigger "bad" behavior (and sometimes "good" queries on similar topics that don't trigger it): huggingface.co/datasets/Naf... ... to find the vector that corresponds with the bad behavior, and then you weaken that vector w/weight…

Nafnlaus/ShrimpMoss_Chinese_Censorship_Abliteration at main huggingface.co

Yep. Once you have the weights, you can train out anything you see as misbehavior. I once made a dataset for abliterating Chinese censorship, for example: huggingface.co/datasets/Naf...

Nafnlaus/ShrimpMoss_Chinese_Censorship_Abliteration · Datasets at Hugging Face huggingface.co

This is the paper that *introduced Transformers * arxiv.org/abs/1706.03762 Read it.

Attention Is All You Need arxiv.org

A large chunk of the proofs solved by AI are not done by OpenAI or similar; most are done by individual mathematicians. github.com/teorth/erdos... There's little secrecy. Most fully disclose their prompts. Some do extremely elaborate prompts. Some, painfully simple, little mor…

AI contributions to Erdős problems github.com

In case anyone needs any free inference compute and wants to do it on the dime of the worst people around, apparently Truth Social's (Perplexity-based) "Truth Search AI" is neither rate limited nor topic restricted. It's been writing a thriller about Barney the Purple Dinosaur for like 10 minutes.

So apparently Tencent Cloud, which I just switched to from OpenRouter, has an off-by-1000 bug on their pricing for DeepSeek V4 Flash, and consumed my entire monthly budget in a single query. #FML

Anthropic's annualized revenue growth rate is pretty insane.

Price per 1M tokens today: 85-300x lower. Param counts vs. GPT-3: ~13% of the total parameters, ~3% of active parameters (est). It's honestly kind of staggering how quickly things advance.

Just thinking about how it took the art world 90-140 years to get over the concept of "photography as art" and wondering if it'll be the same way with AI. Early on artists almost universally agreed with Baudelaire with his critique of photographers as failed artists cheating to make souless slop.