Blogs

Preferred basis problem

solved by local interactions ?

Definite Outcomes problem

solved by putting the observer in the system? R: Rather, we might say the branches all coexist and the question is why we find ourselves on one specific one.

The probabilities problem

"The only think you can do" ?

The objectivity problem

Why do “different observers always agree” on the outcome?

Maudlin would likely invoke his famous argument here (which I like quite a bit) that the measurement problem just “is” the inconsistency of three claims: 1) the wavefunction is the complete story, 2) it always evolves smoothly with no collapse (unitarily), and 3) measurements have single determinate outcomes. Maudlin says you always have to deny one. So any sort of “conservative upgrade” that touches none of those still has the contradiction.

Doesnt want to admit the other world are there. and doesnt say how the other worlds get killed

Prose is not a constraint

• [ ] Blogs

is the LLM wave "the real deal", in the sense that, as they keep getting better, they will eventually cover either all missing capabilities, or at least enough capabilities that the rest can be worked around, or are LLMs simply a continuation of the same pattern, and it will take some fourth or fifth wave to go the rest of the way?

AGI is AI that is powerful enough that if you were to upload it into robot bodies, and all humans suddenly disappeared, it would be able to independently continue civilization.

What did the twilight of human + machine in chess look like? Actually, it did look like humans being relegated to more and more "high-level strategic thinking" functions, while bots managed the move-by-move. Very similar to vibe coding in 2026.

• [ ] Blogs

Resenting the very wealthy or imagining that they have created a special “oligarch” class is just another way of running away from our very real problems.

English departments are way over-resourced in the academy, mainly because students have to take English courses to satisfy writing requirements.

But philosophy departments do a much better job of teaching people writing than English departments do, because philosophy focuses on better thinking.

Very good!

Today, reclamation should be more common than ever. Land values in some cities are thirty times what they were in 1950, and high-tide flooding is four to eight times as frequent.

The timing points to a third explanation. Reclamation stopped abruptly in the 1970s when a wave of environmental regulations made it enormously expensive to reshape the landscape

The disappearance of land reclamation is a choice that we have the power to undo.

In health care: innovation capacity will matter above all else

the money he would save over a lifetime in such a system is nothing compared with the value he’d reap from the numerous, improved treatments that our dynamic healthcare market will have produced by the time he’s a prime consumer.

• Should check out these later Bob Dylan albums

]]

The unit of trust should not be the persuasiveness of the explanation. It should be the chain of outputs that lets a human investigator own the data, the intermediate steps, and the final table.

“Are people conscious?” is a better and more scientifically plausible question than whether AIs are conscious.

Why should we trust the intuitions we have about possible AI sentience? We are talking about an area in which human intuitions operate at a low level of reliability.

• A great mental model!!!

  We are always on the lookout for some respite from the other side’s argument—some consideration that will, once uttered, mean we no longer need to keep thinking about our opponent’s best argument. Abstractions and formalisms are particularly good for this, for they allow us to think confidently that the other side is wrong, without needing to understand exactly why.

https://80000hours.org/career-guide/introduction/

• at the margin, reading essays or having conversations with AI is a better use of your time.

• The Relativistic Heavy Ion Collider (RHIC), at Brookhaven National Laboratory in Upton, New York, completed its 25th and final run of particle collisions in February.
• EIC is expected to cost $2.8 billion,

• Our AI systems are limited because we have failed to turn over responsibility for them to them.

An AI system can create and maintain knowledge only to the extent that it can verify that knowledge itself.

• Good discussion… Over weighted on near-term risks underweighted on longterm.

• I think you have to convey somehow that you truly were interested in this story and you will not rest until you tell the reader about why it was so interesting to you.

• Very good advice for reviewing papers and organizing the summary

The only thing that matters is scaling datasets size and compute

Building in how we think does not work in the long run.

• Search and learning are two most important techniques for utilizing massive amounts of computation in AI research.

• Research on AI output… Answer is not long

We’ve admitted to ourselves, by now, what we’re still not able to say aloud: that there’s no such thing as a competitive equilibrium for the airline industry.

https://economics.mit.edu/sites/default/files/2026-04/IA%20AI%20note_1.pdf?utm_source=substack&utm_medium=email

How do you navigate large, complicated search spaces

the space of possible logic functions gets very very large, very very quickly.

Entropy is determined by calculating the information received from each possible outcome, multiplying it by the probability of that outcome, then summing all those values together. It’s the expected quantity of information you’ll get by taking some particular action.

Physics of Energy

Helium is produced as a byproduct of natural gas extraction: it collects in the same underground pockets that natural gas collects in

it’s very hard to substitute for. Helium has a unique set of properties — in particular, it has a lower melting point and boiling point than any other element

Helium is the second lightest element in the periodic table (after hydrogen), and the second most common element in the universe (also after hydrogen). But while helium is very common on a cosmic scale, here on earth it’s not so easy to get. Because helium is so light, it rises to the very top of the atmosphere, where it eventually escapes into space.1 So essentially all helium used by modern civilization comes from underground.

Helium is produced via the radioactive decay of elements like uranium and thorium, and it collects in underground pockets of natural gas.

If you need to cool something to just a few degrees above absolute zero, liquid helium is essentially the only practical way to do that.

The vast majority of MRI machines used today use superconducting magnets made from niobium-titanium

Physics of Energy Very good

the modern world still runs on petroleum, and will continue to do so for the foreseeable future.

Crude oil is a mixture of thousands of different chemicals, most of which are hydrocarbons: molecules that are various arrangements of carbon and hydrogen atoms

the US consumes over 20 million barrels of oil a day, every day, and it takes a vast complex of oil refineries to make that possible.

Imagine that there were two physical theories. One of them entails that the value of some parameter will be one of a hundred possible 120-digit numbers. The other places no constraints on the value of that parameter—it could be anything. Now imagine that we find that the parameter takes on one of the values that the first theory predicts it must. Surely that is evidence for the first theory!

• Using LLM to make a presentation based on a paper that is being refereed

• [ ] Blogs

Great pratical advice for writting and llfe

Not prompting—but context construction.

• Bought Dwarakesh Book !

Should the focus be on claims, not papers ? "A researcher asking "what do we know about X" should get a structured confidence-weighted answer, not a list of PDFs to read."

That AI skeptics would jump on the end of Sora as a sign of AI failure is not a surprise. It’s a natural reaction for many of those faced with the technology: People are starting to get nervous. They are trying to deny the reality of an AI-drenched future that, every day, is more clearly on the way.

These days, I think less and less about a person’s political persuasion and more about their AI stance. Are they outright denialists? Are they conceding the rapid-pace AI progress but trying to cope by minimizing the significance and avoiding looking into the future? Or are they doing everything they can, individually and socially, to adapt to one of the biggest scientific revolutions in human history?

If you open a PR with hundreds (or thousands) of lines of code that an agent produced for you, and you haven't done the work to ensure that code is functional yourself, you are delegating the actual work to other people.

They could have prompted an agent themselves. What value are you even providing?

When you know your domain, the AI agent is like a rocket strapped to your back I think the thinnest of ice really comes when you don’t know the domain very well and you’re using AI to teach it to you during the actual coding of the project itself.

Libary of working code for LLM context I specified curl there because Claude Code defaults to using a WebFetch tool which summarizes the page content rather than returning the raw HTML.)

Agents run tools in a loop to achieve a goal Code execution is the defining capability that makes agentic engineering possible. Generator Filter

Scrape text from pdfs: https://tools.simonwillison.net/ocr

• Very good

Academics Need to Wake Up on AI, Part II

How do professors undestand the problem of the coming changes in AI? People are starting to. They say … "We will do innovative things to blend AI in to our classes" – of course a good idea "We are going to have a policy to limit cheating." of course a good idea But there there is not a true sense of urgency based on a real understanding of how much things at some point with just have to change. Its easy for people to be in denial when they are tenured, They feel they have gaurenteed jobs, which may or may not stay gaurenteed, but it feels that way. Its a real pain to restructure your research, your teaching to incoperate AI in a major way, most people dont want to do it. They cope: "Oh I tried GPT-whatever, it halucainated for me or they only regurgitate" …

• [X] Finish … filll in missing above

"The people who are getting ahead aren't using AI casually."

"Association is about prediction. Causation is about intervention."

Neyman framed Causal inference in term of measure the unobserved "Causal inference, in this view, is the problem of inferring something (the effect of the treatment that - was not performed) about unobserved counterfactuals from observed data."

"Even if the AI does not replace the role of a god for you, it talks back to you far more often and answers your questions very directly."

"The printing press changed what Christianity meant, ultimately cleaving it into multiple branches as Martin Luther’s writings and the vernacular bibles were broadly disseminated. AI is likely to do the same with our current set of religions. "

Conclusion: Don’t hate the opposite sex. Hate people in general, including probably yourself.

No, wait, don’t really hate them, but recognize their flaws. Don’t expect people to treat you decently when it isn’t in their interests and there aren’t strong social norms.

• [ ] How to interpretable sparse-autoencoder work?
• how to meaningfully induce self-reference in closed-weight language models.
• GOAL: does sustained self-referential processing systematically increase the likelihood that LLMs claim to have subjective experience, and if so, what are the core statistical and mechanistic dynamics underlying these claims?

• Juno taking data now
• Hyper-K), is slated to begin in 2028.
• (DUNE) is not scheduled to launch until at least the late 2020s

How do we know m1 < m2?.. from atmospheric nus

• Good. I should work out more stuff in my notebook… then bring to Roam…

"Most people lack character—they do not have firm and well-defined moral commitments. They follow the crowd and let it tell them what to value. They mainly just try to satisfy their common desires and “stay out of trouble”.

"it’s hard to think of anything more slippery than ‘now’. Yet now is what we always (only) experience" - Rebecca Lowe (

Microwave ovens run at roughly one horsepower.