I like this, but I do think it misses the social stuff. Just a few neurotic/unethical people can completely ruin a team and make all the best people quit. I find this to be extra true on diverse teams (all types of diversity).
934 karma · joined November 13, 2019
- Are they smart? (understands quickly + communicates effectively)
- Are they cool? (won't put poison in the keurig + pleasant to be around)
- Are they high energy? (initiative + action bias)
- Do they have the experience needed to be successful in the role?
Those are pretty strong priors for success. If you find someone with all the above, you've got an ~80% chance of a hit. No need to over-complicate.
In my experience, adding more boxes tends to index towards box-checkers who grew up wealthy. That's how you miss the hyper-smart/diligent state school kids who happened to spend their summers working instead of doing model UN.
Edison was famous for caring about money more than other inventors. One of his most famous quotes was “Anything that won't sell, I don't want to invent. Its sale is proof of utility, and utility is success.” The guy helped found some of the largest companies of his era and died as one of the richest men in the world.
A) opt-out of participating to absolve yourself of future sins or
B) create the systems yourself, assuring you will have a say in the ethical rules engineered into the weapons
If you actually give a shit about ethics and safety (as opposed to the appearance thereof) the only logical choice is B.
The unconscious mind, dream interpretation, importance of early childhood, psychosexual development, talk therapy as we know it today, transference, etc.
What do we remember him for? Oedipus complex. Poor guy. It's like if all we remembered about Newton was alchemy.
This is about whether or not the FBI needs a warrant to see information that was already collected legally by another part of the gov't.
This is not about whether or not the gov't can collect this specific data in the first place, everyone involved seems to agree they do.
Is it that the American ascent to becoming the richest economy on earth wasn’t perfectly smooth and therefore not “fine”?
Most gut aversion to deflation comes from great depression PTSD.
PE is popular for one reason and one reason only: taxes. PE generally makes money on a trade called a leveraged buyout (LBO), where they take out a massive loan to buy a company. Because interest on debt is tax-deductible, going debt-heavy increases the take-home profits of the company (this is called a "tax shield"). Because the profits are higher, the value of the company is higher, and the PE firm makes money on their trade.
What this means in practice is that if you run your company sustainably (low debt, lots of assets). You become a target for a PE firm to attempt a hostile takeover of the company, all while claiming (defensibly, actually) to be doing whats in the best interest of the shareholders. So good companies will try to ward off these attacks by taking on lots of debt and going asset light to minimize the value gain a PE firm might have.
In short, both PE ownership and the brittle, debt-heavy nature of the American economy today can be traced to the tax advantaged nature of debt. For reasons I can't quite understand, nobody seems to be advocating for revoking this tax deduction. I can only surmise this is because everyone hates taxes.
Thank you for coming to my TED talk, your take home exam is a short essay on what you think the mortgage interest tax deduction (started in 1913) did to household debt.
If anyone is curious, there was another paper [0] that came out a few days ago that made a related observation in Vision Transformers. Transformer models appear to pick tokens to store global information in - they need tokens to "think". You can eek some performance improvements (and cool explanation images) by providing the model with specific tokens for this purpose.
Its pretty obvious that virtual TV/Computer screens at arbitrary sizes/distances in arbitrary environments will simply be a better UX than physical monitors.
Not sure if Quest 3 is there yet, but they're iterating towards it.
The database software behind the ANN algo is probably a little more important in practice than the ANN algo itself, unless you're operating at such scale and speed that its an actual issue (e.g. you're google).
Differences between algorithms are a little more interesting when they let you do something totally different, like, for example, minimize the speed hit from doing searches on disk (SPTAG, DiskANN).
- Probability math is confusing and difficult, and a base understanding is required to use PPLs in a way that is not true of other ML/DL. Most CS PhDs will not be required to take enough of it to find PPLs intuitive, so to be familiar they will have had to opt into those classes. This is to say nothing of BS/MS practitioners, so the user base is naturally limited to the subset of people who studied Math/Stats is a rigorous way AND opted into the right classes or taught themselves later.
- Probabilistic models are often unique to the application. They require lots of bespoke code, modeling, and understanding. Contrast this with DL, where you throw your data in a blender and receive outputs.
- Uncertainty quantification often is not the most important outcome for sexy ML use cases. That is more frequently things like "accuracy," "residual error," or "wow that picture looks really good".
- PPL package tooling and documentation are often very confusing and don't work similarly to one another. This isn't necessarily the developer's fault, this stuff is hard, and the people with the domain knowledge needed to actually understand this stuff often have spent fewer hours in the open-source trenches.
In particular:
- "COVID does seem to impact CD8+ T cells": the adjustment I'd make here is that evidence exists that COVID *can* influence T cells, not that it always does.
- "causing T cell exhaustion and reduced effector response.": a quick google suggests there is some conflicting evidence on this point, and most of the studies are done on t cell responses to vaccines delivered after infections, which is a very particular thing.
It seems like you're drawing very heavily from Anthony Leonardi, who, while definitely a scientist and likely right about some things, isn't exactly someone with median (or emotionless) opinions on long covid. The HIV comparison in particular is really out there and unnecessary.
Almost none of these predictions have come true, and while I suppose they may eventually, the posts/articles getting increasingly difficult to pay attention to.
Every time I see another of these my reaction is “god, another one?” I’ve been struggling to evaluate these articles on the merits, I’m sure I’m not alone.
If you want to write an article like this and avoid this problem I recommend that you take extreme care to at least pretend Elon failing wouldn’t give you profound schadenfreude.
1. Commercially license to one or more corporate entities
2. Spin up a start up and license it themselves
Your employer has a right to set their terms of employment and you have a right to tell them to pound sand if you don't want to abide by them.
A more cynical take might be that a "lib-left" orientation is the least likely to draw the ire of regulators/investors/employees/activists/the general public, so most companies drift towards it.
If there's a published theory on this somewhere I'd love to read it.
When a manic depressive with substance abuse issues and prior suicide attempts walks into a clinic nobody needs a "new cognitive science tool" to tell them if they're high suicide risk, and these cases are a larger proportion of suicides than most people think.
Also that sort of missed the point of my comment, which is perhaps my fault for not being clear enough. Preventing suicide is a two step process:
1) find high risk people
2) prevent their suicide attempts
My comment was simply relaying the POV of someone who does this professionally: as a society we are better than most people think at #1 and worse than most people think at #2.
- 239/657 suicides were people under psychiatric care (please note they used present tense, more of them might have had a history of care)
- 187/657 had a previous attempt (!!!)
- "About 22% of the deaths were accompanied by no known inciting event or identified life stressor" Phrased differently, 78% of suicides accompanied a life event or stressor.
Contrary to popular belief, risk prediction isn't usually rocket science. Suicide correlates are extremely well studied, and any experienced mental health professional can point high risk individuals out to you if they ever cross paths with one.
The most challenging problem often is: what do you do about it? Can you get them the help they need? Can you navigate them through a big bureaucracy like the VA? Do you have grounds to forcibly keep them in in-patient psychiatry? Etc, etc.
Source: Relative is a psychologist @ the VA.
With respect to subjectivity it really depends on the task - some tasks are quite amenable to objective classification. One common task for science language models is citation prediction: do these two papers share a citation link? Obviously that's a really simple accuracy metric to report.
Often things are not so simple. An example might be keyphrase extraction - standard practice there is to have grad students sit down with a highlighter and use the terms multiple students agree on (simplification, but not by much). From there it just gets messier. Are you reporting accuracy of all keywords identified or all sentences correctly processed? What about sentences with multiple keywords? What about sentences with no keywords? Very messy, appropriate metrics can be a real topic of debate.
I think a good way to think about scientific language models is that they're useful in exactly the same ways general language models are, but in a very narrow domain (stuff having to do with scientific papers & patents, for the most part).
Use-cases that are possible/useful today:
- Annotation of scientific texts: is this paper about computer science?
- Scientific search: please give me researchers or papers most similar to an input query.
- Helping PhD Students graduate (only kind of kidding)
Use-cases I think will be possible/useful in the forseeable future:
- Scientific question answering: e.g. ask the model to explain a chemical process
- Scientific advice or guidance: e.g. ask what method might be appropriate in a situation.
- Text completion/editing/etc: e.g. help me write my paper. You could probably do more of this today if more $ was invested in science models, we're likely ~5 years behind whatever is going on in the "normal" language space.
As far as potential / shortcomings I'm really pessimistic. I don't think large language models for science are very useful outside of bespoke projects or ever will be for people doing serious science. The main issue is that these models are way too general - if you have a specific science problem you want to solve, its almost always going to be better to train a model to specifically address that problem. You would never, for example, ask a model like Galactica to do what AlphaFold does. Eventually you might be able to, but its never going to outperform a specific model, so if you're a researcher trying to get the best results why would you use it?
I should also add, scientists really care about precision. When summarizing a news story exact words might not be that big a deal, but if you're trying to summarize a scientific paper getting a word wrong can REALLY matter. The bar these models need to clear before scientists trust them with tasks where precision matters is likely much, much higher than in other domains.
I think the most likely outcome is that ~75% of LLM use for scientific text outside of academic research papers will be for search related products. That's definitely a place where they can make a big difference: help people find and understand cool papers that are relevant to their research.
- They benchmarked against general models like GPT-3 but not well-established specific models that have been trained for specific tasks like SPECTER[0] or SciBert[1]. Specter outperformed GPT-3 on tasks like citation prediction two years ago. Nobody seriously uses general LLMs on science tasks, so nobody who actually wants to use this cares about your benchmarks. I want to see task-specific models compared to your general model, otherwise whats going to happen is I either need to run my own benchmarks or, much more likely, I shelve your paper and never read it again. If you underperform some that's fine! If you don't compare to science-specific models all you're claiming is that training on science data gives better science results... thats not exactly an impressive finding. Fine-tuning is a separate thing, I get it, but pleeeeeease just give the people what they want.
- Not released on huggingface. No clue why not. On the back-end this appears to be based on OPT and huggingface compatible, so I'm really confused.
- Flashy website. Combine 1&2 with a well designed website talking about how great you are and most of my warning lights got set off. Not a fan.
@authors, if you're lurking, please release more relevant benchmarks for citation prediction etc. Thanks.
[0] - https://arxiv.org/abs/2004.07180 [1] - https://arxiv.org/abs/1903.10676
I'm sure I'm missing out on quality of life enhancements and whatnot, but really all of those are lower on my hierarchy of needs than staying on task.
[0] https://www.frbservices.org/financial-services/fednow/about....
I bet most of the people here agree 100% with the problems and structural shifts this memo identified, and that technological innovation in the housing space to facilitate both social cohesion and affordability would be a *very good thing*.
Sure, Adam had some highly publicized issues at WeWork. Also, a16z is on a lot of people's shit list right now. Still, why do so few people seem to be rooting for their success? If you're prone to cynicism I get it - the only way this can end is grift - but personally I would like to see more companies trying to address these issues and Adam seems like he genuinely cares about it. Maybe a little too much, even.
But if I must:
- Daly points out that GDP growth != welfare growth. In fact, growing GDP has negative externalities that might decrease aggregate welfare. By reducing GDP growth in wealthy countries he argues we might open up resources for developing economies to grow. This reduces to something like: Daly thinks GDP growth is worth it in poor places but thinks the costs associated with GDP growth in rich economies outweigh the benefits. This is my second point explicitly: he wants growth, but only the type he likes and not the type he doesn't.
- Daly talks about richer countries needing to make "ecological room" for poor countries to grow. What he means is that the growth of richer countries requires the use of natural resources that could otherwise be used to fill the more basic needs of poorer countries. Weird zero sum-ism aside, this is essentially my third point, that we can redistribute wealth/outputs (technically inputs, but let's not split hairs) to reach a state of stable global contentment.
I have no way of knowing if your statement was entirely rhetorical or not, but indeed, I have both benefitted significantly from our imperfect status quo and I wish for something like it to continue.
Over a billion humans have been lifted from extreme poverty since 1990 [0]. I like that. I want more of that. Thats not to say I would like for our system to remain totally unchanged, however. You and I probably agree on a lot.