Small-scale study: 16 law professors were asked to judge short-answer Q&A practice materials written by one of the other 15 professors, or generated by an LLM (Gemini 2.5 Pro). They preferred the LLM materials in 75% of cases, with fairly strong inter-rater agreement. Has some…
Who's Who of AI
Mark J. Nelson
Comp. sci. prof. @ American University, Washington DC. AI & games researcher with miscellaneous other interests. https://www.kmjn.org/
What they're sharing
law.stanford.edu
CrankGPT — fully offline, human-powered local AI squeezlabs.github.io
compass.onlinelibrary.wiley.com
View on Bluesky →
How Many Submissions May an Author Make? A Harmonic Quota for Submissions under Coauthorship arxiv.org
China’s AI Boyfriend Business Is Taking On a Life of Its Own wired.com
Assistant Professors in Computer Science (5 positions) vacancies.maastrichtuniversity.nl
Notes from the AI Now Summit by Mistral koenvangilst.nl
View on Bluesky →
We Don't Understand Neural Networks At The Algorithmic Level kripken.github.io
View on Bluesky →
Shape the Impact of Games research: Become an Assistant or Associate Professor at the SDU Metaverse Lab fa-eosd-saasfaprod1.fa.ocs.oraclecloud.com
Postdoc on human-centric, collaborative AI tue.nl
Articles & links
finally some off-the-grid local AI
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 →
Mark J. Nelson reposted
Now published in open access! Your one-stop shop for the philosophy of language models. It's the spiritual descendant of our two-part preprint from 2024, fully updated. This should be particularly useful for anyone looking for an entry point into this rapidly growing field.
Dealing with too many submissions to TMLR by deriving "a Generalized Harmonic Quota Rule, a framework that subsumes the Harmonic Quota Rule and other natural quota rules" in a 12-page paper is a very The Machine Learning Community solution to this particular problem.
did you see this one a few months ago?
Many openings for Assistant Professors in Computer Science at Maastricht University! Areas of interest include Programming Languages and AI+PL/SE (among others). Deadline August 16.
Mark J. Nelson reposted
if you are wondering about Mistral koenvangilst.nl/lab/mistral-...
Mark J. Nelson reposted
New blogpost: We Don't Understand Neural Networks At The Algorithmic Level kripken.github.io/blog/neurosc... So, why is this even a question? Don't scientists agree on whether we understand LLMs?
Assistant/Associate Professor opening at the University of Southern Denmark in games & interactive technologies; focus areas Human-Computer/AI Interaction, data science, and AI. Deadline August 1.
Intriguing-looking postdoc on human-centric reinforcement learning, part of "an interdisciplinary project between artificial intelligence, social sciences, and societal partners". At TU Eindhoven with Hendrik Baier, deadline July 9. www.tue.nl/en/working-a...
Their own posts
Recent commentary
Underreported thing about Gemini (more than other LLMs I think) is that it's an ok replacement for Google Books, and fluidly multilingual. Like I can ask a question and request answers be based only on books by a specific academic (which are in Greek) and it will dig up relevant passages.
A reason I've pulled back from reviewing for big AI conferences is a feeling that I'm doing unpaid supervision of other people's PhD students. Too many ppl submitting 10+ papers to a single conference where I doubt the prof whose name is on the paper has done a thorough review & revision themselves.
An interesting thing about LLMs in Python is that they seem to broadly push code towards some kind of conventional wisdom about best practices, as judged maybe by whoever is setting up the posttraining recipes (I say "in Python" mostly because I notice that more strongly in Python).
A thing LLM coding sort of makes more feasible is making *smaller* personalized apps. Like instead of WMATA's big and annoying to use transit app, I'm trying out a custom little thing that just shows me the 2 bus lines I take. Lines hardcoded; stops hardcoded; no configuration; barely any interface.
Asked Gemini something about DC apartment logistics, and it recommended I "pop down to the front desk" to verify. Takeover of Google AI by DeepMind Londoners confirmed.
Isabelle/Isar has a nicely named proof method, blast, that tries to brute force, letting you concisely prove simple but tedious lemmas by just writing 'thus "Q" by blast', which'll go through if in fact blast can prove it. In good cases, LLMs feel kind of like that for me now: pip install by blast.
A pop AI book where the two people contributing blurbs were Al Gore and Sam Altman. A little microcosm of 2018.
Their network
In Mark J. Nelson's orbit
Center = Mark J. Nelson. 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.