HNHacker News
TopNewBestAskShowJobs

astrophysician

521 karma · joined May 3, 2019

astrophysicianhn@gmail.com
submissionscomments
astrophysician··on Ask HN: What difficult problems have you or your company solved using AI?
I work for a data science consulting company and previously in advertising, so I can say that there are (1) a number of very real problems that cannot realistically be solved _without_ "AI" (though what does that term even mean these days?), and (2) endless piles of "AI"-branded nonsense.

First of all, what do you consider AI? Deep learning? Hierarchical bayesian models? Linear regression? I would call all of these "AI" simply because I have seen all of these labeled as such at one point or another.

Does your business revolve around computer vision tasks ("how can I track people/cars/things in this image/video?")? You almost definitely need "AI" even if you have almost no historical data.

Does your business revolve around optimizing decisions based on large amounts of historical and relevant data (e.g. user interactions on your website, etc.)? You _probably_ need "AI" (but it depends).

There are an infinite number of ways to incompetently apply AI and upsell what you're doing to people who don't know the difference. It happens all the time and is a natural consequence of any hype cycle. But underneath all of the hype there is obvious substance (just maybe not at your company).

Concrete and very public examples of competent AI:

* Tesla's "auto"-pilot (computer vision/SLAM)

* Recommender systems that you see everywhere (Facebook Watch/News Feed, YouTube, Netflix, etc.) -- can be reinforcement learning based, factorization based, etc.

* Real-time bidding markets in advertising (DSP's like Google's DV360, Trade Desk, Xandr run optimized bidding "AI")

astrophysician··on Making better decisions with the Brier score
In my experience I’ve never found an instance where you would use Brier scores over cross entropy/Bernoulli/Binomial log likelihoods. Does anybody know a concrete example when you would prefer Brier??
astrophysician··on Particle mystery: physicists confirm the muon is more magnetic than predicted
Yea, some people are disappointed; some of the more interesting and exciting moments in physics are when we find out we're wrong, but not always. E.g. I will never forget the time and place I heard about the preliminary detection of primordial B-modes by BICEP (which turned out to be dust contamination) -- that was a predicted detection from canonical inflation models, as the Higgs was a standard prediction from the standard model (also a pretty exciting moment).

Not seeing something when we "expect" to not see anything (from the perspective of certain models) might be more boring, but it's definitely not a "waste" (again speaking purely from a physicist's standpoint).

We know the standard model is incomplete, but where and how are not well known. Not seeing evidence for new physics rules out certain models, and places upper/lower limits on others. It's progress either way.

astrophysician··on Particle mystery: physicists confirm the muon is more magnetic than predicted
From a physicists standpoint, not seeing something unexpected is not a waste at all.
astrophysician··on What Data Can’t Do
I agree -- as ML becomes increasingly easy to be applied by non-experts or people without a heavy math/stats background, I've seen an increasing volume of arguments against the data science profession (someone the other day called DS the "gate-keepers") but: there be dragons.

Anyone can use SOTA deep learning models today, but in my experience, it's more important to understand the answer to "what are the shortcomings/consequences of using a particular method to solve this problem?" "what is (or could be) biases in this dataset?", etc. It requires a non-trivial understanding of the underlying methodology and statistics to reliably answer these questions (or at least worry about them).

Can you apply deep reinforcement learning to your problem? Maybe. Should you? Well, it depends, and you should understand the pros and cons, which requires more than just the knowledge of how to make API calls. There are consequences to misusing ML/AI, and they may not even be obvious from offline testing and cross validation.

astrophysician··on Yes, experts will lie to you sometimes
> The link says: "We’re still learning how well COVID-19 vaccines keep people from spreading the disease", so don't make authoritative statements about post-vaccination spread unless you have a real source.

You know what, that's fair, I should say you "could" still, but again, I didn't make this up, here's the Mayo Clinic giving the same exact advice

> Keep in mind that if you're fully vaccinated, your risk of getting COVID-19 might be low. But if you become infected, you might spread the disease to others even if you don’t have signs or symptoms of COVID-19. This could be dangerous for people who are unvaccinated and at increased risk of severe illness from COVID-19. -- https://www.mayoclinic.org/diseases-conditions/coronavirus/e...

It seems like you feel that if there's any uncertainty, you should do the thing that makes you most comfortable? Again, I just ask why? Is wearing a mask that debilitating?

> What if I'm a parent of 4 kids and I bring my 4 kids with me? In fact, me going to McDonalds can and does spread heart disease to my progeny, which is a big reason obesity and heart disease are inter-generational problems. Kids learn bad eating habits from their parents.

I am genuinely confused by this example and how much you think this relates to a communicable disease: if you cause people to eat McDonalds, um....I guess yes you are spreading heart disease to them? How is that similar to someone being in the same physical space as you, who does not want to eat your proverbial McDonald's, but has to anyway because they have to breathe?

astrophysician··on Yes, experts will lie to you sometimes
I agree: in retrospect, in todays context, those statements are not good public health policy. And saying "they are not effective in preventing general public from catching coronavirus" may indeed be a false statement, even at the time. But, when masks were in short supply and when it wasn't really known how much protection they provide (so saying "hey masks might help" may give people a false sense of comfort thinking that they're protected and able to live as normal), those comments, while not the most objective, make way more sense.

But I think this "they lied to us, we shouldn't trust them" rhetoric is silly and dangerous. Again, I agree, it's definitely a lesson to learn. Public health professionals are human beings too and I think it's fair to say they are doing their jobs to the best of their ability.

astrophysician··on Yes, experts will lie to you sometimes
> You say this as if it's a fact, but as of 2 days ago, we still don't know the answer to that question. Unless you have a source I'm unaware of?

https://www.cdc.gov/coronavirus/2019-ncov/vaccines/fully-vac... <- that's the advice from public health authorities. I myself am not a public health expert, and so I don't pretend to know more than public health authorities, though maybe that is not the case for you?

> Heart disease is a large risk for an even more significant portion of the population, yet McDonalds is still open and people can still choose to eat there?

Right, because you going to mcdonald's can't give someone else heart disease.

> Once the vaccine is widely available (and it already is for the at-risk groups), it's up to those who are at-risk to get themselves vaccinated. If they choose not to, well, that's their prerogative to assume extra risk and I no longer have any obligation to protect them (at least, not any more than I protect my obese friends from heart disease).

It is already widely available to people with high risk, yet it is still almost impossible to find appointments in some states (including my own). It may very well open up to the majority of the population before everyone at high risk is able to get vaccinated. Also, people with compromised immune systems cannot get vaccinated.

If someone chooses not to get the vaccine when they otherwise should and have no health conditions that preclude them getting it, sure I agree they are taking on that risk themselves, but there are plenty of people that should get the vaccine that aren't able to for whatever reason.

Also....protecting your obese friends from heart disease is absolutely not the same thing and is not a good analogy.

astrophysician··on Yes, experts will lie to you sometimes
I mean I see the point, I think their comments made a lot more sense in the context of where things were back then, and they didn't consider how they would be viewed after the initial panic phase was over.

> the result is eroded trust that health advice being given is in your immediate best interest.

I would argue that they are making health advice that minimizes the population-wide risk; I don't expect them to tell me the optimal thing for me to do to protect myself if that advice means that the population-wide payoff is worse (i.e. "hey everybody the best thing you can do is to grab an N95 mask" might be optimal for everyone who is able to do that but that is not the optimal thing to do for the population). I kind of understand people's frustration but I just don't personally feel that this was that controversial.

astrophysician··on Yes, experts will lie to you sometimes
You should wear masks in public when you are vaccinated because while you're protected from severe symptoms, you can still carry the virus and transmit it to others.

> The ultra risk aversion is what many people are souring on, especially considering the risks are low to begin with for the vast majority.

Yea but...they are not low for a significant portion of the population, which should matter? And those people are affected when people don't wear masks because many of them are not able to completely isolate. I don't really understand this line of reasoning -- wearing a mask helps people, so why not do it? Is it really that debilitating? I don't like wearing a mask either but if wearing one helps, why not wear it?

astrophysician··on Yes, experts will lie to you sometimes
Yea but....he's right? What is the controversy here? There was a huge concern about supply shortages and people hoarding PPE. Masks are the best thing the public can do at this point but they aren't perfect protection and the medical community are the ones with the greatest need for PPE...Not to mention, early on in the pandemic it wasn't clear how much protection masks would really give, and people thinking "oh i just need to wear a mask, then I can just do what I normally do" was something they were trying to fight against.
astrophysician··on What does the night sky look like on Mars?
Simply not true, starlink will have a real impact on research (see e.g. https://www.lsst.org/content/lsst-statement-regarding-increa... for the impact on LSST) and will make some science impossible to do. Ironically, amateur long exposure shots will be fine because amateur astronomy doesn’t rely on pixel-level brightnesses and you can easily subtract out streaks in a way that’s pleasing to the eye. Whether or not you think it’s justified, it’s not true to say this will not impact astronomy. Some science cases will be fine, others will be significantly impacted and still others will be SOL and impossible to do.
astrophysician··on Predicting Hard Drive Failure with Machine Learning
I agree and have been doing some similar analysis on the backblaze dataset. I suppose you can use this for prediction but I'm personally just interested in post-hoc analysis and (1) getting better AFR estimates when failure counts are low + (2) exploring time-dependence of hazard functions with different priors (GP priors, etc.). This post and your comment have motivated me to make a post this weekend! Thanks!
astrophysician··on Show HN: Compare prices that US hospitals charge patients, insurance companies
Unfortunately not my experience -- went to nearest ER for sudden and agonizing abdominal pain, got some basic pain meds, got a CAT scan to diagnose pneumonia, was sent home within a couple of hours (lots of waiting).

Turned out the hospital was not in network, they charged me $15,000 (including $10,000 for a "level 5" ER room stay, which is twice the most expensive price listed on the OP's website and is 10x the local average price for that service, despite the fact that my experience could not possibly have been Level 5 service). Eventually they just sold the debt to a collections agency. I tried to fight but they didn't seem to care, and why would they? They already sold the debt!

astrophysician··on We don't need data scientists, we need data engineers
I'm guessing just venting personal frustrations due to their own experiences, plus maybe poor hiring and guidance of data scientists in their own teams? I have definitely seen DS people in my experience that fit some of the descriptions here, but I think it's a mistake to trivialize the DS position itself. A good DS is a valuable asset, but depending on your company/data, maybe not worth the cost. Plus there is no "single" DS candidate or role, a lot of these roles (data engineer, DS, analyst, swe) blend together at times, and it's about finding the right balance of skills.

Sometimes I think a company (not having the DS experience themselves) mistakenly over-hire DS roles in today's hype of "AI" when their data is mostly run-of-the-mill and only requires simple linear models that can be architected an understood by a stats/math-savvy engineer. Even then, a good DS is still useful (even linear models can be complex: e.g. what priors do you want to use? Do you want a multi-task solution? etc.), but maybe not worth the cost.

astrophysician··on We don't need data scientists, we need data engineers
Just want to say that while the data science profession definitely includes a wide range of people and skillsets, a good data scientist should be practical and able to work with the available data in whatever state it's in.

No good data scientist should ever expect data to be pristine. And a good data scientist, even if they don't have quite the engineering chops necessary to build a production-quality ETL, should know enough about the process to help guide it. If they aren't a part of that process, they're not being a good DS. They can't expect someone not involved with their problem to know what tradeoffs to make, and if they don't know exactly how their data went from raw form to the ETL-ed form, they're probably going to make bad assumptions, and those assumptions may very well make their architected solution a complete pile of garbage. Not to mention, how can a DS offer suggestions for solutions if they aren't deeply familiar with the raw data that's available?

To me, a good data scientist should, at bare minimum, have several skills.

* They should first and foremost (but not solely) be an in house expert in statistics and machine learning to know what can be done with data, and what can't be done with data. They should arrive with that knowledge. Engineers I think have a tendency to trivialize this, but true expertise in this domain comes only with years of experience.

* They should strive to find modeling solutions that are right for a particular business problem. If they seem to be only applying the hottest research regardless of the tradeoffs for the particular business problem, that's a red flag.

* Their focus should be on integrating themselves with the product/business as much as possible, and with the engineering team as much as possible. If they're expecting to be handed directives, that's a recipe for a ton of wasted time.

DS should never, ever be siloed into their own little DS world. They will be useless without a deeply intimate knowledge of the business goals, the needs of product, and the capabilities of the engineering team.

As they progress, they should become more and more "full-stack", otherwise they are stagnating.

astrophysician··on Superintelligence cannot be contained: Lessons from Computability Theory
> I think you're missing the central point, which is that computation is observer relative. Anything can be interpreted as a computational process.

I see what you're saying, and maybe I am misunderstanding your point, but to me it seems like you've gotten yourself bogged down in wordplay when there is something much simpler going on: say I have a human named Bob from Des Moines, and next to him is a machine constructed to approximate Bob to arbitrary accuracy (this is possible because Bob is a made up of a finite number of particles/wavefunctions). Are you arguing that there's something special about Human Bob? If so, what is your argument for that? The two are "indistinguishable" and by that I mean whatever threshold you have for two things to be "indistinguishable" (practically speaking), you can technically make a reproduction of Bob that satisfies that threshold.

> That's circular reasoning. The point is that qualia are not something which, in principle, can be the subject of computation. There is no way to represent the fullness of sensation itself, like the redness of red or the softness of silk, as information. So how can our brains be "computing" it?

I would argue this is circular reasoning. "There is no way to represent the fullness of sensation itself" -- yes I would argue there is: whatever time-dependent set of physical states make up this "realization" in your brain.

astrophysician··on Superintelligence cannot be contained: Lessons from Computability Theory
> I don't think you can say that life is built on computational processes unless you use a definition of "computation" that is so vague and all-encompassing that it becomes effectively meaningless.

I mean, there's nothing physically stopping us from simulating a brain, right? It's a finite object with a finite amount of physical ingredients, and therefore with a finite amount of computing power we can simulate what it does. To me personally, that's a computational process. Maybe that's an overly broad definition of computation, but I think these debates tend to be about whether there is something fundamentally different about "life" (by which I assume you include consciousness). But maybe that's not what you're saying.

> He points out that "if we are to suppose that the brain is a digital computer, we are still faced with the question 'And who is the user?'"

What does that question even mean? I think it seems deep because we humans have a tendency to ascribe some sort of supernatural aura to our lived experience. Life is something incredible but that (at least to my knowledge) is not uncomputable...

> There is no computational process that will produce the sensations of colour or sound or touch.

Got one: the brain!

> At best you will have some representation that requires having actually experienced those sensations to understand it.

Why does a computer not "experience" something?

astrophysician··on Superintelligence cannot be contained: Lessons from Computability Theory
> The burden of proof is on those claiming minds and computers are the same thing.

I never understand this argument. IMHO the rest of your argument seems to obfuscate something rather simple... your brain is a physical object. The atoms that make up the neurons, glial cells, etc. are real and are governed by physics. Are you saying that what you think of as "you" is made up of more than these physical ingredients? To me that would be invoking something "magical" and would thus bear the burden of proof.

> You can have concrete green triangular objects in the world, but not triangularity or green as such...except in a mind that is able to abstract universals from particulars

Why can a sufficiently powerful computer not conceptualize these things? How is that fundamentally different than what goes on in your brain?

astrophysician··on London will be overwhelmed by Covid in a fortnight says leaked NHS briefing
because tolerating flu deaths doesn't completely overwhelm healthcare systems?
astrophysician··on Why I'm Lukewarm on Graph Neural Networks
In my opinion there aren't many deep learning applications that are fundamentally snake-oil, it's more that people who don't quite understand the limitations and advantages of particular methods that end up being (witting or unwitting) snake-oil salespeople.

Deep graph NN's are very useful for lots of data that are inherently graph-structured (another commenter above mentioned chemistry applications, and there are lots of other examples). Whether or not for the time-being they give SOTA results on common datasets, being able to work directly with graph-structured data is quite appealing in many cases.

astrophysician··on Stan is a state-of-the-art platform for statistical modeling
Just to clarify: prophet implements a particular model with Stan as a backend, not Stan models generally.
astrophysician··on Congress is set to ban most surprise medical bills
As the recipient of a $25,000 bill for an emergency visit to an out-of-network hospital that I have been fighting for two years (still ongoing), this is such an overdue blessing.

FYI -- many individual states ban "balance billing" i.e. hospitals sending you the leftover bill that insurance didn't cover (btw -- hospitals can charge a special "out-of-network" rate, in my case 10x the average cost around the area for the services I received). Problem is, if you are on a "self-insured" plan (your card should say this, these are plans you'll get at university or when working for a large company), the state law can't do anything, and there has been no federal law banning this practice. It's absolutely inhumane.

astrophysician··on I Regret Quitting Astrophysics
Unfortunately right now its `localhost` but once I finish a couple of posts (tomorrow? a year from now?) I'll put it online. Thanks for asking though :)
astrophysician··on I Regret Quitting Astrophysics
As another former astrophysicist-turned-datascientist (though with only 2 years after the transition instead of Marcel's 7), this hits close to home, but my experience in both worlds I think is a bit different.

If you are an astrophysicist curious about moving into the datascience profession, I would just like to point out that satisfaction with a move like that depends on a large number of factors, and that you should consider not only Marcel's experiences (which I think are important and shared by many) but your own personal circumstances. There are pro's and con's for every person.

Let me offer another datapoint:

* I left after my PhD, which ended absolutely terribly. I knew years before finishing that I was just not cut out for a life in academia, and positioned myself in the best way I knew how for a life in industry, but I did not anticipate just how truly painful the experience would be. I stayed to finish my PhD out of a nagging obligation to follow through with a lifelong dream of mine. I thought I would deeply regret leaving after receiving my masters degree. Whether or not that is true, I really don't (and will never) know. But I was not happy.

* My personal reasons for unhappiness are not uncommon but are definitely not universal. I did not excel at the entrepreneurial and social aspects of astrophysics that are crucial for success. Without strong social and professional ties to close collaborators within the community, and without an innate ability to thrive in an environment with very little structure, add to that a predisposition for mental health problems, I burnt out quickly. I found myself withdrawing from social settings as burnout set in, and the hill was steeply downward from there. When I don't socialize, my ideas stagnate, my work suffers, and my perspective warps until it's easy to submit to delusions like "I am doing great" when really I was (obviously, in retrospect) not doing great at all.

* The transition to datascience was on one hand fantastic. I have a structure that (1) forces me to socialize daily, (2) forces me to confront my productivity levels frequently. Even when completely remote these past 9-10 months, I find my productivity is far more consistent than it was in graduate school, and my life is far more sane. I have a true work-life balance that I could only fantasize about before.

* The stress is much lower (in my current role). My work no longer defines who I am as a person. Doing what you love sounds great on paper but really what it means are higher highs and lower lows with each small success and failure. While in the beginning of my graduate school experience, I was reasonably well-rounded and rational, as I dived deeper into my own thesis work, I started noticing that I became far less receptive to criticism. My thesis work was so integral to my identity that it became harder and harder to be objective about it.

* Financially I am far more secure, and that is incredibly important to me. I'm not earning anywhere near as much as a $500k AI/ML researcher at FAANG, but I'm also not earning a $30k grad student salary or a $50-60k postdoc salary. I can live extremely comfortably while still building a savings (and am no longer in credit card debt). I also can find a job in the location I want. I don't need to write 20 detailed applications (all during "application season") for roles around the world, wait months to hear replies, and then pack my bags to go wherever I'm accepted. That's a hell of a perk.

* Like Marcel, however, I do find myself missing many aspects of the world of astrophysics. I miss a community completely made up of people at the top of their professional game, making groundbreaking contributions on existentially important things, where everyone seems to share the same "datascience" mindset. I miss symposiums and lunch talks and people with deep passion for what they do. However I have found it incredibly satisfying to work alongside people with other strengths; programmers, product, designers, and managers. These people may or may not have a deep passion for what they do (for some its a job, for others its a passion), but they all have incredibly valuable expertise to share and perspectives to offer. So in that sense, it's not clear to me which world is better.

* As time wears on, I'm finding myself losing some of the knowledge I once had, and that's a bummer. Datascience allows me to keep the statistics channels of my brain somewhat active, but I'm no longer doing things like graphical models and MCMC. I'm trying (slowly) to build a blog where I can do these things as a hobby, but it's difficult to motivate a lot of times; I find that a lot of times it takes me back to the stressful mindset of grad school. I still have three unpublished papers from my graduate school days but the longer time goes, the less of a chance I think these will ever get published, but they stay in their current state because the last time they were touched I was almost losing my mind from stress and I just avoid teleporting myself back to that terrible time.

I apologize for the rant but my point is: your mileage may vary; Marcel makes some very good points but consider all the pros and cons for yourself. As someone else commented, a sufficiently motivated data scientist can definitely get back in the academic game (though it will be more challenging).

astrophysician··on US internet speeds 91% faster in 2020 according to user speed tests
For the uninitiated:

https://www.nngroup.com/articles/law-of-bandwidth/

astrophysician··on Studies Point to Big Drop in Covid-19 Death Rates
hmm..I see your point but I'm not sure I agree, saying 2/3 of people that otherwise would have died are no longer is tailored to the problem. E.g. if mortality rate from all covid infections went from 0.1% -> 0.001%, thats incredibly impressive and important since the mortality is now 1% of what it was, meaning 99% of deaths are being avoided but "the mortality rate went down by 0.099%" sounds small and insignificant.
astrophysician··on When we lose weight, where does it go?
> The most common misconception by far, was that fat is converted to energy. The problem with this theory is that it violates the law of conservation of matter, which all chemical reactions obey.

I mean what they're getting at must be true, we are obviously not converting 1kg of rest mass into energy but that sentence is very much not true.

astrophysician··on California Bill Would Mandate 'Crime Prediction' Algorithms Instead of Cash Bail
As a data scientist, this is terrifying. A questionnaire is asking things like "have you only held low-wage jobs" to determine if someone should be locked up? The model isn't completely open to the public? (how that is justified is beyond me, this model literally decides whether or not you, as a citizen, are locked in a box).

If it's not obvious: these methods easily reinforce all sorts of terrible biases in the data. Even with a perfectly transparent model, it's unclear what sort of data biases you are reinforcing, because getting unbiased data is extremely difficult if not impossible.

astrophysician··on Covid vaccine makers commit to not seek approval until complete Phase III trials
> This isn't even about stopping the trial early - this is about approving it conditionally before the trial is done and then waiting for the trial to end to really approve it

I see -- yea totally agree, I thought they were committing to a rigid timetable or something but that's not what's happening here as you say.

Page 1 of 4Next →