Um, it is, isn’t it?
bokov
Karma: 460
I agree. My reason for posting the link here is as reality check—LW seems to be full of people firmly convinced that brain-uploading is the only only viable path to preserving consciousness, as if the implementation “details” were an almost-solved problem.
Finally, someone with a clue about biology tells it like it is about brain uploading
http://mathbabe.org/2015/10/20/guest-post-dirty-rant-about-the-human-brain-project/
In reading this, suggest being on guard against own impulse to find excuses to dismiss the arguments presented because they call into question some beliefs that seem to be deeply held by many in this community.
It depends. Writing a paper is not a realtime activity. Answering a free-response question can be. Proving a complex theorem is not a realtime activity, solving a basic math problem can be. It’s a matter of calibrating the question difficulty so that is can be answered within the (soft) time-limits of an interview. Part of that calibration is letting the applicant “choose their weapon”. Another part of it is letting them use the internet to look up anything they need to.
Our lead dev has passed this test, as has my summer grad student. There are two applicants being called back for second interviews (but the position is still open and it is not too late) who passed during their first interviews. Just to make sure, I first gave it to my 14 year old son and he nailed it in under half an hour.
Correct, this is a staff programmer posting. Not faculty or post-doc (though when/if we do open a post-doc position, we’ll be doing coding tests for that also, due to recent experiences).
it’s not strictly an AI problem—any sufficiently rapid optimization process bears the risk of irretrievably converging on an optimum nobody likes before anybody can intervene with an updated optimization target.
individual and property rights are not rigorously specified enough to be a sufficient safeguard against bad outcomes even in an economy moving at human speeds
in other words the science of getting what we ask for advances faster than the science of figuring out what to ask for
(Note that transforming a sufficiently well specified statistical model into a lossless data compressor is a solved problem, and the solution is called arithmetic encoding—I can give you my implementation, or you can find one on the web.
The unsolved problems are the ones hiding behind the token “sufficiently well specified statistical model”.
That said, thanks for the pointer to arithmetic encoding, that may be useful in the future.
The point isn’t understanding Bayes theorem. The point is methods that use Bayes theorem. My own statistics prof said that a lot of medical people don’t use Bayes because it usually leads to more complicated math.
To me, the biggest problem with Bayes theorem or any other fundamental statistical concept, frequentist or not, is adapting it to specific, complex, real-life problems and finding ways to test its validity under real-world constraints. This tends to require a thorough understanding of both statistics and the problem domain.
That’s not the skill that’s taught in a statistics degree.
Not explicitly, no. My only evidence is anecdotal. The statisticians and programmers I’ve talked to appear to overall be more rigorous in their thinking than biologists. Or at least better able to rigorously articulate their ideas (the Achilles heel of statisticians and programmers is that they systematically underestimate the complexity of biological systems, but that’s a different topic). I found that my own thinking became more organized and thorough over the course of my statistical training.
Also, I’m not sure if this is your intention, but it seems to me that the goal of spending 20 years to slow or prevent aging is a recipe for wasting time. It’s such an ambitious goal that so many people are already working on, any one researcher is unlikely to put a measurable dent in it.
In the last five years the NIH (National Institutes of Health) has never spent more than 2% of its budget on aging research. To a first approximation, the availability of grant support is proportional to the number of academic researchers, or at least to the amount of academic research effort being put into a problem. This is evidence against aging already getting enough attention. Especially considering that age is a major risk factor for just about every disease. It’s as if we tried to treat AIDS by spending 2% on HIV research and 98% on all the hundreds of opportunistic infections that are the proximal causes any individual AIDS patient’s death. I would think that curing several hundred proximal problems is more ambitious than trying to understand and intervene in a few underlying causes.
I have no illusions of single-handedly curing aging in the next two decades. I will be as satisfied as any other stiff in the cryofacility if I manage remove one or more major road-blocks to a practical anti-aging intervention or at least a well-defined and valid mechanistic model of aging.
Secondly, you probably shouldn’t worry about pursuing a project in which your already-collected data is useless, especially if that data or similar is also available to most other researchers in your field (if not, it would be very useful for you to try to make that data available to others who could do something with it). You’re probably more likely to make progress with interesting new data than interesting old data.
This is ‘new’ data in the sense that it is only now becoming available for research purposes, and if I have my way, it is going to be in a very flexible and analysis-friendly format. It is the core mission of my team to make the data available to researchers (insofar as permitted by law, patients’ right to privacy, and contractual obligations to the owners of the data).
If I ran “academia”, tool and method development would take at least as much priority as traditional hypothesis-driven research. I think a major take-home message of LW is that hypotheses are a dime a dozen—what we need are practical ways to rank them and update their rankings on new data. A good tool that lets you crank through thousands of hypotheses is worth a lot more than any individual hypothesis. I have all kinds of fun ideas for tools.
But for the purposes of this post, I’m assuming that I’m stuck with the academia we have, I have access to a large anonymized clinical dataset, and I want to make the best possible use of it (I’ll address your points about aging as a choice of topic in a separate reply).
The academia we’re stuck with (at least in the biomedical field) effectively requires faculty to have a research plan describable by “Determine whether FOO is true or false” rather than “Create a FOO that does BAR”.
So the nobrainer approach would be for me to take the tool I most want to develop, slap some age-related disease onto it as a motivating use-case, and make that my grant. But, this optimizes for the wrong thing—I don’t want to find excuses for engaging in fascinating intellectual exercises. I want to find the problems with the greatest potential to advance human longevity, and then bring my assets to bear on those problems even if the work turns out to be uglier and more tedious than my ideal informatics project.
The reason I’m asking for the LW community’s perspective on what’s on the critical path to human longevity is that I spent too much time around excuse-driven^H^H^H hypothesis-driven research to put too much faith in my own intuitions about what problems need to be solved.
Great idea! Here’s how I can convert your prospective experiment into retrospective ones:
Comparing hazard functions for individuals with diagnoses of infertility versus individuals who originally enter the clinic record system due to a routine checkup.
Thanks for reminding me about SENS and de Grey, I should email him. I should reach out to all the smart people in the research community I know well enough to randomly pester and collect their opinions on this.
People gain skills by working on hard problems, so it doesn’t seem necessary for you to take additional time to explicitly hone your skill set before starting on any project(s) that you want to work on.
The embarrassing truth is I spent so much time cramming stuff into my brain while trying to survive in academia that until now I haven’t really had time to think about the big picture. I just vectored toward what at any given point seemed like the direction that would give me the most options for tackling the aging problem. Now I’m finally as close to an optimal starting point as I can reasonably expect and the time has come to confront the question: “now what”?
So, for a retrospective approach with existing data, I could try to find a constellation of proxy variables in the ICD9 V-codes and maybe some lab values suggestive of basically healthy patients who consume a lower-than-typical amount of calories. Not in a very health-conscious part of the country though, so unlikely that a large number of patients would do this on purpose, let alone one specific fasting strategy.
Now, something I could do is team up with a local dietician or endocrinologist and recruit patients to try calorie restriction.
I should clarify something: the types of problems I can most efficiently tackle are retrospective analysis of already-collected data.
Prospective clinical and animal studies are not out of the question, but given the investment in infrastructure and regulatory compliance they would need, these would have to be collaborations with researchers already pursuing such studies. This is on the table, but does not leverage the clinical data I already have (unless, in the case of clinical researchers, they are already at my institution or an affiliated one).
My idea at the moment is to fit a hidden Markov model and derive a state model for human aging. But this pile of clinical data I have has got to be useful for all kinds of other aging-related questions...
If we are in an environment of open conversation and I say something that brings up an emotional trauma in another person and that person doesn’t have the self-awareness to know why he’s feeling unwell, that’s not a good time to leave him alone.
?! Depends. If you don’t understand that person intimately or aren’t experienced at helping less self-aware (aka neurotypical) people process emotional trauma, it’s probably a very good time to leave him alone. Politely.
I was tempted to vote “makes no sense at all”. I did not because I’ve had far too many experiences where I dismiss a colleague’s idea as being the product of muddled thinking only to later realize that a) the idea makes sense, they just didn’t know how to express it clearly or b) the idea makes practical sense but my profession chooses to sweep it under the rug because it’s too inconvenient. On Stackoverflow and LW I see the same tendency to mistake hard/tedious problems for meaningless problems and “solve” the problem by prematurely claiming to have dissolved the question or substituting in a different question the respondent finds more convenient.
Some questions really are meaningless or misguided. But experience has taught me to usually give questions the benefit of a doubt until I have enough background information to be more sure. So, I played along and gave the technically correct answer of “I’m parts both”.
Come to think of it, “Red/Blue makes no sense at all” is not even a valid answer to the question. The question did not ask whether it made sense. Such a meta-question should really be a checkbox orthogonal to the main poll question.
This has taught me that I find it more intuitive to think in terms of conditional probabilities than marginal probabilities.
Here is what you can do to make your post better:
At the top put a very short, concise TLDR with NO IMAGES.
More data. It sounds like you did a pretty rigorous deep-dive into this stuff. Instead of making assertions like “These projects usually take one of a few forms …” or “There appears to be almost nothing in this general pattern before January 2025″ show the raw data! I get that you need to protect the privacy of the posters, but you could at least have a scrubbed table with date, anonymized user IDs, name of subreddit, and maybe tags corresponding to various features you described in your piece. Or at least show the summary statistics and the code you used to calculate them. Social media can very much be analyzed in a replicable manner.
Fewer anecdotes. The images you embed disrupt the flow of your writing. Since you’re anonymizing them anyway, why not go ahead and quote them as text? It’s not like an image is somehow more authentic than quoted text. Also, as per above, maybe move them to an appendix at the bottom. The focus should be on the scope and the scale of this phenomenon. Then, if a reader is interested enough to pursue further they can choose to read the semi incomprehensible AI co-authored stuff in the appendix.
Without independently verifiable evidence, I expect there to be a low probability of this being a widespread trend at this time. However, it does point to something we should probably prepare for—mystically inclined people who don’t understand AI building cults around it and possibly creating a counter-movement to the AI-alignment movement as if that work wasn’t already hard enough.
So how do we nip this shit in the bud, people?