astro1234

Huge fan of Claude for a long time but switched now to Codex after all of this "ok one more week, one more week" stuff. It feels like Sol is quite capable, and the nice thing is the refusal rates seem more reasonable, _especially_ if you're doing anything thats reasonable but that they consider "sensitive". The dealbreaker for me is their initial nerfing / silent sabotage but now down-grading for anything even resembling AI work. That's a pretty low blow IMO.

I think the biggest phase transition happened roughly fall of last year with AI adoption. My entire job is now AI orchestration, and then a LOT of my time spent writing the specs/prompts, reviewing and validating the output imperfectly and generally being very paranoid. But it's to a point where no one I know really has a choice but to work like this because doing things by hand is just way slower, even though I find slop to be like nails on a chalkboard. And the "sloppiness" has been getting progressively less sloppy. Reading Google Docs generated by Opus 4.8 is torture, whereas 5.6 can actually write concisely and clearly. So I expect a lot of the paranoia and pain to gradually decrease over time.

That being said, domain knowledge is still very important but you just don't need nearly as much tribal knowledge (because RAG or a tuned LLM can wrangle the chaos of your company documentation) or broad technical knowledge (because if you understand stats and PCA congratulations, when the coding agent recommends probabilistic PCA you now get to learn what it is, why its good for your use case (maybe), and get the gist fairly quickly; you no longer need to know about it in the first place. That was the return on investment for breadth of technical knowledge: knowing what to even look into).

I won't say "the field has moved towards a bigger focus on fundamentals" though -- this is all just one internet person's opinion. No one is telling me this is happening I'm just realizing that these are the muscles I am flexing more nowadays and my deep breadth of technical knowledge muscles are not flexing as much any more. Like: you are now inundated with recommendations of various methods from a coding agent: you need to understand them quickly; fundamentals gives you that. At least I think so, but of course, maybe I'm wrong about this due to misunderstanding the actual skills involved in what I'm doing now.

In my experience Data Science looks very little like it used to a few years ago, and the priority skill these days is good strong understanding of the basics and very good sense of judgement. To me, statistics is the absolute number one priority for any data scientist. You need to fully and deeply understand just basic concepts in statistics in order to translate what you see into action and do what you’re really there to do which is to prevent screwing up and acting on the wrong information or what’s more likely the wrong interpretation of the information.

For me the most valuable skill I have is a lot of experience applying and learning about Bayesian statistics: what it is (the beginning parts of Jaynes Probability Theory were not useful practically but deeply significant in helping me understand what it means and where it comes from), seeing lots of probabilistic models in the wild, playing around with them in both personal and professional worlds. Some people play video games, I love building hierarchical models. The nice thing is that in addition to it being very expressive It’s also just so much easier, such an intuitive way to avoid footguns because it just requires you to conceptualize one small bit at a time. When you’re done you get the inference for free with lots of charming stops along the pareto frontier between rigor and compute. Variational inference, expectation maximization, EM, Laplace. You can understand all of them with just a few concepts. Plus marginalization is just so unbelievably elegant to me. What is so surprising and beautiful to me is that Bayesian inference and marginalization are so useful and practical today. That being said there are plenty of unintuitive surprises, which is also a plug to not just understand the math but the theory and fundamentals to know how to interpret what you’re doing and seeing.

Also again this is still a great guide with lots of super important stuff (SVD/PCA linear algebra and linear regression (so much reward from just understanding linear regression from multiple perspectives)), no doubt. But if you really truly understand the basics you don’t need to worry about graph Laplacians (though highly highly recommend it’s also beautiful). Because more and more you can outsource the question of which method is ideal to a deep research agent that will read and understand arxiv for you. But you still have to audit it which means just really understanding the fundamentals is so crucial nowadays.

That and valuing speed and practicality. Strongest discriminator between someone junior and someone senior is recognizing when to reach for something simple and when you need to bring out bigger guns.

I don’t mean to say you are wrong, I just mean to ask what criteria you would consider necessary to satisfy in order for you to consider a system to be reasoning.

I tend to agree but to adjudicate that someone has to define what they consider reasoning to be.

I’m curious what this means? I think the evidence is pretty convincing that, while brittle, there is reasoning going on (though it depends on your definition of reasoning which I’m curious what that is for you).

Why not? I think there’s fairly strong evidence that there is something that convincingly looks like reasoning. I think anthropic has done some nice circuit tracing and mechanistic interpretability work on this for instance.

One of the highest quality, nuanced strategy docs for this company I’ve ever seen. Cannot poke any holes in this. Also while people very fairly pile onto the Reality Labs disaster, it gave Meta the ray bans which is a durable form factor to build on. May or may not pan out that they recoup everything from this (it is more possible than I think most people think but requires a lot more success there which will take time if it happens).

The thing people forget: Meta has a unique advantage that it can leverage which is: make large bets and chase them for a long time. Mark is not beholden to a board. He would have been out many years ago if he was. It can be of course a downfall but it hasn’t been.

Failing is not nearly as painful as being conservative. He bet on something and held his ground and while it was a failure, not taking a bold strategic gamble would be arguably dumber.

Did my PhD at Princeton, knew Jenny Greene personally (not my adviser though). There is zero conflict of interest in Astronomy generally. No one has anything to gain. Various institutions, Simons included, are just one source of much needed funding. Jim Simons is also a legend in the field, known for Chern-Simons (major result), then founding the medallion fund which netted him billions which he then durned around and used to fund fundemental science. Astrophysics is too low paying for anyone who doesn’t genuinely care about it to do it.

Funding institutions can influence which research gets done, that’s what they do by definition. This can steer people towards and away from various topics or questions, but people will loudly speak their mind if they don’t think something is right. It’s a core tenant of the culture. Go to a colloquia and watch people debate and critique each other.

its the only way to study certain things. Can't make up for it with smaller telescopes or cheaper projects.

Not a dumb defensive question but you should know the nice thing about these experiments is the incredible amount of work that goes into calibration and understanding all error signals.

Messing up the data analysis has major precedents. If you aren't familiar you should look into BICEP data in 2014, they thought they had observed primordial gravitational waves which would have been earth shattering. Instead they just messed up the dust correction pipeline. I don't envy the day they came to that realization. I was in several conference rooms at Princeton where BICEP people presented their analysis and David Spergel (of WMAP, previous head of the department at princeton) and others were able to walk them through how they thought they had kind of messed things up. This is what routinely happens, ESPECIALLY when something unexpected is observed. Every possible explanation is looked into, and ESPECIALLY in cosmology, you can do that incredibly well. Cosmology is one of the most beautiful sciences in my experience, precisely because we have such good ways to model the observations to probe various models, and you can treat the observations with Bayesian stats with virtually no risk of misspecifying your model, or, if you do find its misspecified, you have discovered something new about the universe.

The process to go from raw observations to physics, correcting for all the crap in between early universe light and us (dust which also rotates light polarization -- this explained the BICEP issue, instrument systematics which are measured to incredible precision on the ground (e.g. point spread function -- what is the detector response to various intensities of light; e.g. you get electrons for bright sources that spill into neighboring pixels)

Everyone everywhere is looking to make a name for themselves by discovering the discrepancy -- be it a screwup of some other team (astro community is generally very supportive and positive but also competitive) or a problem with simulation assumptions, a genuine discrepancy in our understanding of the universe (i.e. the tension in the hubble constant -- you infer rate of expansion from cosmic background radiation / early universe observations, and then try it using an alternative method -- using local variable stars, and you get a statistically significant difference).

So I would say: if there's a screwup it will be found, and a genuine fuckup is possible and does happen, but when it does believe me we will know usually within a few months. You'll have a ton of people trying to reproduce the results, pouring over everything there is that could possibly explain these observations. The wheel of astrophysics grinds slowly but it grinds finely.

Edit: also shoutout to Jenny Greene -- one of the world's foremost experts on galactic astronomy and also a genuinely great person. She rented me her house for a summer for dirt cheap when I was a poor grad student with nowhere to stay. Also hosted the best graduate student parties (our idea of a party is beer and board games and complaining about our advisers)