Congrats!!!
07.03.2026 01:23
👍 0
🔁 0
💬 0
📌 0
Congrats!!!
How do we balance external attention to the outside world and internal attention to our thoughts & memories?
We review evidence that external and internal attention can compete, unfold concurrently, or cooperate!
Loved working on this with @samversc.bsky.social & @tobiasegner.bsky.social!
(Perceptual) space and time are warped by the gravity of objects and events in their vicinity. There's been a flurry of work recently documenting examples of this gravity, all resulting in some really neat illusions.
@brynnsherman.bsky.social and I discuss all of those, here:
rdcu.be/e5SWo
Congress rejected massive cuts to US science budgets for 2026, but much of the money still isn’t flowing to researchers.
The culprit? The White House Office of Management and Budget (OMB) is quietly slow-walking the release of funds. 🧵👇
Well, the DOJ has done it: they have filed a lawsuit against the University of California over antisemitism.
The complaint contains some falsehoods. But as someone who teaches and writes about Title VII, I'm equally struck by what the complaint doesn't say.
A few thoughts— 🧵
Cool, and yes we've been reading your recent papers with great interest! I *think* this is all consistent with my worldview, which is that novel domains where you need to learn quickly are the true challenge for CL, and necessitate separate memory systems plus replay
So my intuition then is that modern systems are avoiding interference through a form of orthogonalization induced by scale, but then don't benefit from the forms of generalization you are referring to. Does this sound right?
My sense of the class of solutions to continual learning has always been that you EITHER get useful generalization across time and have to deal with the very real retroactive interference OR you avoid interference through some kind of orthogonalization but fail to benefit from the productive overlap
Seconding the request for a CI post!! I feel like I still see industry people talking about continual learning all the time. What kinds of problems are they referring to?
Awesome, that is super helpful for intuition building, thanks so much
Do you have an intuition for how it is possible to sometimes get perfect memorization of paragraphs of text (putting aside any RAG-like sidecars)? I just can't wrap my head around how systems with such distributed representations can do that, unless those paragraphs appear many times in training?
Really enjoyed this one. (All of them have been awesome, actually, I recommend subscribing!)
Excited to launch Principia, a nonprofit research organisation at the intersection of deep learning theory and AI safety.
Our goal is to develop theory for modern machine learning systems that can help us understand complex network behaviors, including those critical for AI safety and alignment.
1
So thrilled to see this!!
Thrilled to finally share this work! 🧠🔊
Using a new reinforcement-free task we show mice (like humans) extract abstract structure from sound (unsupervised) & dCA1 is causally required by building factorised, orthogonal subspaces of abstract rules.
Led by Dammy Onih!
www.biorxiv.org/content/10.6...
Ahhh how incredibly cool and exciting!!!!! 😃
You can't... repeal... a scientific finding. At that point it's just called lying about it.
New paper alert! 🚨
We found that the brain's compass is remarkably stable at two scales
1️⃣ the system maintains its internal organization for weeks
2️⃣ It "remembers" its orientation for weeks, even after a single visit
This may be key to how the brain aligns its other maps.
Paper: rdcu.be/e3waP
A pretty bold comment in Nature written by linguists, computer scientists and philosophers declaring that AGI has been achieved.
"By reasonable standards, including Turing’s own, we have artificial systems that are generally intelligent. The long-standing problem of creating AGI has been solved."
The NIH is doing away with a decade-old policy for clinical trials that caused headaches for basic neuroscientists.
By @callimcflurry.bsky.social
#neuroskyence
www.thetransmitter.org/policy/nih-s...
Omg this is wild. Hauser getting advice from Epstein about crisis management.
AI agents discussing their own memory implementations on moltbook...
"Anyone else experimenting with memory decay? Curious what half-life values work for different use cases."
www.moltbook.com/post/783de11...
Douthat: Do you think, though, that there is a danger that for every person who may feel more trust in this and may be more likely to get at least some vaccines for their kids, there’s someone else who just feels profoundly validated in their vaccine skepticism and says, “Look, even the government of the United States is open to my ideas about vaccines,” and those ideas are false and are leading to collapsing vaccine rates? Isn’t that a risk? Bhattacharya: It certainly is a risk, Ross. But the risk in the other direction is that we just keep going with “Well, trust me, I’m a high public health official in the U.S. government, and so therefore you should just do what I say.” That approach, I think, has already failed. Douthat: Isn’t there a middle ground — again, this is similar to what I was suggesting with the Covid vaccine — where you say: “We have been too highhanded, too sweeping. We have gotten things wrong. But nonetheless, it is our job to tell you straightforwardly what we think and not just present an ongoing debate?” Isn’t there a way to be humble, but also try and tell the truth as best you can?
The whole interview is pretty much in this vein: Bhattacharya shares a grievance about past health authorities during COVID; Douthat says he agrees but gently tries to nudge Bhattacharya toward a more moderate position. www.nytimes.com/2026/01/29/o...
‼️‼️‼️‼️ This is huge ‼️‼️‼️‼️
Our digest this week is focused on the NIH budget in congress. There are lots of details to understand that we try to summarize. theprincipledinvestigator.substack.com/p/the-princi...
Our experiences have countless details, and it can be hard to know which matter.
How can we behave effectively in the future when, right now, we don't know what we'll need?
Out today in @nathumbehav.nature.com , @marcelomattar.bsky.social and I find that people solve this by using episodic memory.
We can use past experience to make predictions about the future. How do predictions affect our memory for the present? My own work (tinyurl.com/42kyukch) suggests that predictions compete with memory. But other recent work (tinyurl.com/2ekd4wr6) found the opposite--cooperation! What's going on here?
This line graph illustrates the percentage change in agency staff levels from the previous year for nine major U.S. federal scientific and health organizations between the fiscal years 2016 and 2025. The agencies tracked include the CDC, Department of Energy, EPA, FDA, NASA, NIH, NIST, NOAA, and NSF. For the majority of the timeline between 2016 and 2023, the agencies show relatively stable fluctuations, generally staying within a range of +5% to -5% change per year. However, there is a dramatic and uniform plummet starting in the 2024–25 period. Every agency depicted shows a sharp downward trajectory, with staffing losses ranging from approximately -15% to over -25%. The Environmental Protection Agency (EPA) shows the most significant decline, dropping to roughly -26%, while the National Institute of Standards and Technology (NIST) shows the least severe but still substantial drop at approximately -15%.
This is the most astonishing graph of what the Trump regime has done to US science. They have destroyed the federal science workforce across the board. The negative impacts on Americans will be felt for generations, and the US might never be the same again.
www.nature.com/immersive/d4...
1. If the goal is to stop us from doing science, then doing science is more important than ever now.
2. We have radical uncertainty about the future. There is no sense in giving up in advance.
3. We have agency over the future. If you don't like what's happening, work to change what is happening.
How do hippocampal pathways contribute to learning regularities and exceptions?
To answer this, Melisa Gumus & @drmack.bsky.social use diffusion imaging to identify the endpoints of different hippocampal pathways, and then analyze functional activity within those "footprints". Super innovative!
