JOPRO (@JOPRO_org) header
JOPRO (@JOPRO_org) profile picture

JOPRO supports interdisciplinary scholarship, cultivates emerging researchers, and builds the networks that connect them.

Boston and Beyond
Joined July 2022
Photo by JOPRO (@JOPRO_org)Photo by JOPRO (@JOPRO_org)Photo by JOPRO (@JOPRO_org)Photo by JOPRO (@JOPRO_org)Photo by JOPRO (@JOPRO_org)Photo by JOPRO (@JOPRO_org)Photo by JOPRO (@JOPRO_org)Photo by JOPRO (@JOPRO_org)Photo by JOPRO (@JOPRO_org)Photo by JOPRO (@JOPRO_org)
Summer Internships have launched: blog.jopro.org/i/195885234/n…
1
1
55
A school district in western New York is spending $57,590 on a humanoid robot for its high school this fall, plus roughly $2,000 a year in software fees and a subscription to the laptop-based AI tutor that comes with it. The vendor, Realbotix, calls it their flagship education deployment and says candidly that they don't yet have evidence the robot improves learning. The district will help them find out. The superintendent's rationale is reasonable on its face: rural students get firsthand exposure to technology that's otherwise abstract to them, and some sense of the careers attached to it. I take that seriously. Exposure has value that doesn't show up on a test score. On the evaluation design: a professor quoted in the piece laid out what a real one would look like: compare students using the robot, students using the same AI tutor on a laptop, and students using neither, then measure retention and time on task after the novelty wears off. That comparison is the only thing that answers the actual question, which is what the $50,000 body adds over the software. The district's plan is mostly qualitative feedback plus login counts. We've been looking at K-12 as part of JOPRO's Data x Direction efforts (and SET at large), and this is the pattern in miniature. Districts are making consequential purchases without a shared method for telling afterward whether the purchase worked. When the buyer hasn't defined the evidence standard, the vendor's learning agenda becomes the default one. I've watched that dynamic play out in procurement in other sectors and it rarely corrects itself from the inside. The school, Salamanca, serves about 1,300 students, 79% of them economically disadvantaged, nearly a third American Indian or Alaska Native. Districts like this one are where the experiments land first and where the resources for evaluating them are thinnest. Which raises a question I don't have a good answer to yet: who is supposed to be building that evaluation capacity, and at what level? State education departments? Regional consortia? Right now it seems to be nobody, and the pilots are proceeding anyway. This is the era we are in, and we will need people informed and able to build paths (no less "solutions") against the torrents we are enduring.
2
1
158
Awesome new theme issue of Philosophical Transactions: ‘World models in natural and artificial intelligence’ Thank you @adamsafron! royalsocietypublishing.org/r…
16
75
10
360
44,396
Summer internship: Human Flourishing, Agency, Cognitive Sovereignty opportunities.jopro.org/
1
1
39
Happening now:
From early algorithms to generative AI to autonomous agents — each shapes us differently. Under what conditions does media actually support human flourishing? THIS THURSDAY: AHA Seminar Series featuring Sunny Xun Liu (Stanford Social Media Lab) Link in comments!
1
1
35
JOPRO retweeted
Launching the DigiNEST program blog today. The first post asks: what happens to human agency when AI systems begin to mediate how we interpret our own experiences? Shorter answer: we need better design language, and better questions.
1
1
1
16
Happening now:
🌐 AI is now more persuasive than political ads — for controversial sociopolitical questions. So what does that mean for our shared reality? THIS THURSDAY: AHA Seminar Series with Thomas Costello on AI & the future of belief. April 2 | 4–5pm ET | Join us: zoom.us/j/91647096870
3
1
1
58
Launching the DigiNEST program blog today. The first post asks: what happens to human agency when AI systems begin to mediate how we interpret our own experiences? Shorter answer: we need better design language, and better questions.
1
1
1
16
The concern is no longer theoretical. Recent work from MIT, Oxford Internet Institute, and others shows that AI sycophancy, emotional dependency, and eroded autonomy are already measurable in deployed systems.
1
1
5
Full post via the JOPRO Substack: open.substack.com/pub/jopro/…
1
5
Full announcement of internships in the coming week:
3
If you or someone you know has done research in this space, we'd love to hear from you. Soft tease -> @JOPRO_org: Forthcoming internship & RA opportunities ahead on these topics! Paper: "Slow convergence: Career impediments to interdisciplinary biomedical research" pnas.org/doi/10.1073/pnas.24… Please DM / reach out, or share with folks who may be up for a discussion around these topics.
2
1
2
104
This is a work in progress but I made a visual that compares how much water AI and data centers use in total to other industries. Relies on a lot of botecs I justify in the methodology. andymasley.com/visuals/water…
8
15
3
121
48,903
Good reading ahead of Embodied Intelligence '26, and our work in @CognitionFuture @JOPRO_org @Orthogonal_Lab
1/ We know LLMs struggle with "Theory of Mind" (modeling what others know) It turns out they also lack a "Theory of Space" (modeling where things are) A new paper reveals that SOTA LLMs suffer from "Belief Inertia", refusing to believe their own eyes when the world changes 🧵
1
96
(1/2) Just dropped a new paper: "World Properties without World Models: Recovering Spatial and Temporal Structure from Co-occurrence Statistics in Static Word Embeddings". A key line of evidence for LLM "world models": linear probes recover city coordinates and historical dates from hidden states. @wesg52 & @tegmark did this with Llama-2 and got R²=0.91 for city locations. Very cool result. But is it world models or word statistics? An alternative: maybe the structure isn't emerging inside the LLM. Maybe it was already latent in training text itself, inherited from the systematic differences in how language describes different places and eras. I ran the same probes on GloVe and Word2Vec, static word embeddings from 2013-2014, trained purely on distributional statistics; no layers, no attention, no contextual processing. R²=0.71–0.87 for city coordinates (see map). And the signal is selective, not a probe artifact. Latitude, longitude, temperature: all recoverable. Elevation, GDP, population: R² goes negative. The probe finds real distributional structure, not noise. So: deflationary for world models. But inflationary for language. Co-occurrence statistics alone preserve a richer imprint of the physical world than anyone assumed. The words that surround "Nairobi" and the words that surround "Oslo" are systematically different, and that difference is enough to localize them. On a map. The structure was never in the model. It was in the corpus. 👇
7
21
3
83
6,433
Congratulations to Aida, who started working on @be_inaudible with us at the Augmentation Lab. We’re so excited to see more privacy-restoring, human-centric technology! 🦾
Today, we're introducing Spectre I, the first smart device to stop unwanted audio recordings. We live in a world of always-on listening devices. Smart devices and AI dominate our world in business and private conversations. With Deveillance, you will @be_inaudible.
1
4
541