Who's Who of AI

B! 🐝 Cavello (they/them)

58 trust critic @b-cavello.bsky.social · 685 followers
activist, aspiring-polymath, problematic feminist working to better this world @b_cavello@mastodon.publicinterest.town avatar: Portrait of B against colorful background banner: Collage of creative commons illustrations
What they're sharing

Articles & links

↻ B! 🐝 Cavello (they/them) reposted
@hwaight.bsky.social

I’m excited to share a new paper in Nature that shows how large language models launder the strategic rhetoric of authoritarian states. Paper here: www.nature.com/articles/s41.... A thread.

State media control influences large language models | Nature nature.com
AI Weekly's analysis →
  • Chinese state-media content appears in typical LLM training sets at roughly 41 times the rate of Chinese-language Wikipedia.
  • Across 37 countries, models prompted in the local language produce more regime-favorable responses in countries with lower press freedom.
  • A pretraining experiment with just 6,400 state-scripted documents pushed an open-weight model to pro-government responses nearly 80 percent of the time.
Read full analysis →
 View on Bluesky →

Cool resource from @emilymbender.bsky.social and @nannainie.bsky.social for folks looking to use more precise and less anthropomorphic language. This piece includes a bunch of examples for substitutions to support clearer communication! buttondown.com/maiht3k/arch...

How to talk about buttondown.com
AI Weekly's analysis →
  • The guide covers six language categories: cognition, emotion, communication, agency, human-role analogies, and biological metaphors.
  • Proposed substitutions include 'probabilistic automation' for 'artificial intelligence' and 'conversation simulator' for 'chatbot.'
  • Rephrasing 'ChatGPT assisted students' as 'the students used ChatGPT' is given as an example of restoring human accountability.
Read full analysis →
View on Bluesky · ♥ 44 ↻ 14 ↩ 1 · 9 from the directory shared this · 10d ago
↻ B! 🐝 Cavello (they/them) reposted
Kanishka Misra @kanishka.bsky.social

I will be at #ACL2026 from July 2--7! I will be giving a keynote at CDL workshop on controlled rearing and hypothesis generation from language models! Tianyang Xu (first author) and I will present work on cross-modal generalization in VLMs on July 7! Paper: aclanthology.org/20…

Cross-Modal Taxonomic Generalization in (Vision-) Language Models aclanthology.org
AI Weekly's analysis →
  • Frozen Qwen3 and Llama 3.2 language models, paired with DINOv2 or SigLIP image encoders, predicted object hypernyms they never saw during training.
  • On hypernymy questions alone, Qwen3-0.6B scored 78.5 F1 and Qwen3-1.7B reached 88.5 F1, against a 46.7 majority-label baseline.
  • The generalization broke when researchers shuffled image-label pairs across categories, dropping average visual coherence from 0.27 to 0.12.
Read full analysis →
 View on Bluesky →
↻ B! 🐝 Cavello (they/them) reposted
sorelle @friedler.net

Are you curious about the environmental impact of AI? Install the AI Impact Tracker to estimate the environmental footprint of your ChatGPT usage, including energy, water, and CO2. chromewebstore.google.com/detail/ai-im... #FAccT2026

AI Impact Tracker - Chrome Web Store chromewebstore.google.com
AI Weekly's analysis →
  • AI Impact Tracker is a new Chrome extension that quantifies the carbon, water, and energy usage tied to a user's ChatGPT activity.
  • It measures output tokens and combines them with coarse location data to model localized power grid and water resource impacts.
  • The listing shows version 1.0.1, updated June 25, 2026, three users and no ratings, with aggregate data sent to aiimpacttracker.cs.haverford.edu.
Read full analysis →
 View on Bluesky →
↻ B! 🐝 Cavello (they/them) reposted
@kottke.org

CrankGPT. “Just a hand crank, a little computer, and a small stack of speech and language models running locally. Provided the electronics are kept dry and at a reasonable temperature, there’s no reason this thing won’t still work in a thousand years.” [squeezlabs.github.io]

CrankGPT — fully offline, human-powered local AI squeezlabs.github.io
AI Weekly's analysis →
  • CrankGPT runs a full voice-interactive AI pipeline on a Raspberry Pi 5 with 8GB RAM, powered solely by a 20W hand-crank generator.
  • Cold-start to functional conversation takes roughly 30 seconds; time to first token ranges from 0.8 to 2.9 seconds depending on model size.
  • Memory bandwidth, not raw compute, is the primary bottleneck for on-device LLM inference, with DDR5 hardware achieving 29-58% faster token generation than DDR4.
Read full analysis →
 View on Bluesky →
↻ B! 🐝 Cavello (they/them) reposted
@loreleikelly.bsky.social

👀 discuss! a plus one for accountability & access, but potentially bad if it becomes part of blind populist financially incentivized race over the cliff. apnews.com/article/bern...

AP Exclusive: Bernie Sanders unveils plan to give the public direct ownership of AI companies apnews.com
AI Weekly's analysis →
  • Sanders proposes a one-time 50% stock tax on AI companies with at least $200 million in annual AI sales, paid in shares rather than cash.
  • The transferred equity would seed a sovereign wealth fund Sanders estimates at roughly $7 trillion, managed by a seven-person independent commission.
  • A 5% annual dividend from the fund would send more than $1,000 a year to every American, with the remainder going to health care, education and housing.
Read full analysis →
 View on Bluesky →
↻ B! 🐝 Cavello (they/them) reposted
@baricks.bsky.social

Excellent new research from @cdt.org @mluria.bsky.social analyzes the dark patterns applicable to AI chatbots, including general-purpose systems (ChatGPT, Gemini, Claude) and “companion” platforms (Replika, Character.AI): cdt.org/insights/dar...

cdt.org View on Bluesky →
↻ B! 🐝 Cavello (they/them) reposted
logan koepke @jlkoepke.bsky.social

yes, and — even worse: today AI systems today enforce *existing austerity* of our public benefits systems and social safety net, helping facilitate determinations of who is worthy and not of care, food, unemployment, etc. only made worse by OBBBA's Medicaid cuts + work require…

Benefits Tech Advocacy Hub btah.org View on Bluesky →
Their own posts

Recent commentary

AI fairness folks, help me out? I know that back circa... 2018 or so, there were some handy interactives that had demonstrations of how different definitions of fairness could lead to different model behavior. I'm struggling to track down the links now. Can you point me?

View on Bluesky · ♥ 0 ↻ 0 ↩ 1 · 50d ago
Their network

In B! 🐝 Cavello (they/them)'s orbit

Center = B! 🐝 Cavello (they/them). 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.