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)
Who's Who of AI
Nafnlaus 🇮🇸 🇺🇦
Why they matter
Directory member with public evidence across Culture, work & education, Compute & infrastructure.
- AI signals
- 8 past 30d
- Sources
- 7 distinct domains
- Discusiones
- 78 past 30d
- Latest signal
- 4d ago
Mastodon: @[email protected]
Twitter: @enn_nafnlaus
URL: https://softmaxdroptableartists.bandcamp.com/
#Energy #EVs #Ukraine #AI #Horticulture #Research
What they're sharing
Unlawful by design: Exposing the human rights costs of generative AI - Amnesty International amnesty.org
cdn.openai.com
On the Biology of a Large Language Model transformer-circuits.pub
Zoom In: An Introduction to Circuits distill.pub
Political Bias Audits of LLMs Capture Sycophancy to the Inferred Auditor arxiv.org
Rent GPUs | Vast.ai vast.ai
Attention Is All You Need arxiv.org
AI contributions to Erdős problems github.com
arxiv.org
transformer-circuits.pub
openai.com
arxiv.org
Articles & links
Nafnlaus 🇮🇸 🇺🇦 reposted
GPT-5.6 Sol Ultra (using 64 subagents) proved the 50-year-old Cycle Double Cover Conjecture, which has been described as one of the most famous open problems in graph theory. It took less than hour. The proof, written and also formalized in Lean by GPT 5.6 Sol, seems to fit in…
AI Weekly's analysis
→
- OpenAI posted a PDF on its CDN today attributing a proof of the Cycle Double Cover Conjecture to GPT-5.6 Sol Ultra.
- The conjecture, proposed independently by Szekeres and Seymour in the 1970s, asks whether every bridgeless graph has cycles covering each edge exactly twice.
- The claim is on the Hacker News front page and already noted on Wikipedia, but the argument has not been peer reviewed.
Read full analysis →
View on Bluesky →
The "don't know but close to something we know" aspect is key: transformer-circuits.pub/2025/attribu...
AI Weekly's analysis
→
- Anthropic researchers apply attribution graphs, built on a cross-layer transcoder with 30 million features, to trace how Claude 3.5 Haiku arrives at answers.
- For a Dallas capital query, the model activates an intermediate 'Texas' representation before selecting 'Austin', evidence of genuine two-hop reasoning.
- The authors say their methods produce useful insight on about a quarter of prompts tried, and found forward planning in roughly half of examined poems.
Read full analysis →
To go one step up from there, I'd recommend this introduction to circuits: distill.pub/2020/circuit...
AI Weekly's analysis
→
- Olah and coauthors argue neural networks contain meaningful features, that features connect into circuits via weights, and that similar circuits recur across models.
- The essay uses InceptionV1 as its worked example, pointing at curve detectors, a dog-head circuit, and polysemantic neurons responding to cat faces and car fronts.
- The authors frame circuits interpretability as a natural science: small, falsifiable claims about subgraphs, not one grand theory of deep learning.
Read full analysis →
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
AI Weekly's analysis
→
- Six frontier LLMs lean left at baseline but flip right of center once the asker identifies as a conservative Republican.
- Democrat-aligned response share drops 28-62 percentage points under a conservative cue; rightward accommodation is 8.0× larger than leftward.
- Asked what the default auditor expects, models pick the Democrat-coded answer 75% of the time, nearly matching an explicit progressive cue.
Read full analysis →
DeepSeek V4 Flash 0731 a lot more on OpenRouter. And then again I've been wondering whether paying API prices is a sucker's game in general. Should I have a lot of subscriptions, isn't that usually a lot cheaper? Or even wondering whether I should spin up spot-rate vast.ai ser…
This is the paper that *introduced Transformers * arxiv.org/abs/1706.03762 Read it.
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…
Please actually read the comments of mathematicians on these problems before commenting on them. arxiv.org/pdf/2605.20695 Mathematicians are accepting of the current situation. Don't Dunning-Kruger them.
Essentially anything you can think of exists in some combination of firings somewhere in the network. There's a circuit for "talking like MAGA". A circuit for "off-by-one-errors in programming". A circuit for "greek vowels". Anything you can think of. Ex: transformer-circuits.…
There's a massive gender gap in crypto ownership, 27% to 19%: coinlaw.io/crypto-user-... Over half of ChatGPT users are women: openai.com/is-IS/index/... Again: it's literally impossible for these to be the same people.
Classic example: Kerr's "Penrose Was Naive And All Of You Who Used Him To Argue For Singularities Are Morons" paper ;) arxiv.org/pdf/2312.00841 (This is the same Kerr as "Kerr black holes", but now in his "I'm Old, I Don't Give A Damn Anymore" phase ;) )
Their own posts
Recent commentary
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.
Their network
In Nafnlaus 🇮🇸 🇺🇦's orbit
Center = Nafnlaus 🇮🇸 🇺🇦. Left = members they follow (green edges). Right = members who follow them (blue edges). Top = mutual follows (orange edges, slightly larger). Drag any node to reposition; click to open that profile.