Posts
Comments
You're mistaken about lemon markets: the initial fraction of lemons does matter. The number of lemon cars is fixed, and it imposes a sort of tax on transactions, but if that tax is low enough, it's still worth selling good cars. There's a threshold effect, a point at which most of the good items are suddenly driven out.
Tubes for this would be more expensive but could be used for piping hydrogen. Pipes for moving hydrogen would be much smaller and at high pressure.
Gas pipelines have lower losses at high pressure, which is why natural gas pipelines are typically >40 bar.
Those Mach numbers are for the relevant speed in air. I would have written that differently, but that's how the cited paper worded things.
Mostly-sealing against part of the tube before cutting it is less problematic than dealing with a large pressure difference.
Aerodynamic support and propulsion in hydrogen is less expensive than magnetic propulsion and support in a vacuum-filled tube. Building an unpressurized tube is cheaper than a tube that doesn't buckle under compressive forces. And so on.
Downside is it doesn't work. There are several reasons it doesn't work, but one is: What happens to the vaporized rock? Were you thinking it just goes up a 20km deep hole without condensing on the walls?
Using sliding electrical contacts for power is fine for current high-speed trains, but it doesn't work as well above 200 m/s.
It can't use "air" around it for engines because what's around it isn't "air".
Oxygen is much heavier than the fuel it's used with, and you'd either need liquid oxygen (which increases costs) or pressurized tanks (which would perhaps double that mass). That's still lighter than batteries, yes, but engines are also needed. Piston engines are inefficient and/or heavy, and gas turbines are somewhat expensive.
It's not that difficult to separate water and hydrogen, that's true, but processing that much gas is still rather impractical when batteries have enough specific energy. Simply condensing it in the tube is...possible, but would increase drag, especially considering density variation issues, and you'd have to deal with getting it out of a long sealed tube without leaking hydrogen.
Also, if batteries are good enough, the cost of replacing the hydrogen alone probably makes batteries better than burning the hydrogen.
For a recent example of embezzlement, in Vietnam, 1 person embezzled about 3% of the country's annual GDP.
How about burgers at restaurants, then? They're more expensive in America too.
In 2019, the average Japanese employee worked 1,644 hours, lower than workers in Spain, Canada, and Italy. By comparison, the average American worker worked 1,779 hours in 2019.
The real overtime cultures are China, India, Mexico, and South Korea.
It's not like they were the only major investor in WeWork. There was Insight Partners, for example. But I could find some other examples if you want...
To be clear, I was talking about the period up to 2019 and COVID. The cultural attitudes about investing in stocks vs money in bank accounts came from Nikkei performance before the last few years.
Since 2019, it's up 2x, but the yen went down vs the dollar and there's been some inflation. In terms of what a unit of the Nikkei can buy in America, the valuation is flat since 2019. Which might actually be a hint about what's driving the relative valuation...?
I could easily write a similar list about American companies. Let's see...
- Investors keep funding dumb money-losing ventures like WeWork and Juicero!
- Employees aren't kept long enough to justify training them!
- Switching jobs frequently and smooth talk is valued much more than a solid track record!
- Developers keep switching to the latest JS framework for no good reason!
Maybe one culture is better economically than the other, but I can't see it explaining a 2x wage difference. But if you posit that Japanese management culture is universally really bad, then you have to explain things like:
- lack of foreign firms coming in and outcompeting
- Japanese companies succeeding making cars in America
Maybe you're misreading that. US average wealth is ~5x the US median wealth.
The actual required work by the winning teams was a relatively small amount of software and computation, guided (I think) by a relatively small amount of manual examination. This seems more of a triumph of GPU progress and pretrained vision transformers making difficult problems suddenly easy than a generally applicable system. Given that progress, just releasing the data and letting people work on it without the monetary prize or competition restrictions would've worked too - given the same amount of publicity, at least, but I don't think the argument here is "prizes are good because they can be an efficient way to get media attention".
There's been a significant amount of research on the relative merits of prizes, patents, and research funding, and the general conclusion has been that prizes aren't usually that great.
It has 3 parts: the Wuhan market cases, genetic analysis, and talking about Bayesian statistics. I searched for "government" and "selection bias" and "CCP" and didn't get any results in the article. What specifically do you want me to talk about here that I didn't already say?
Overall, I think the post sucks, but some of the comments are decent, like this and this. Peter's arguments wrt biased data seem worthless.
Both.
I haven't watched much of that debate, but my understanding is that the main points were:
- wet market cases
- genetic sequence analysis
(1) is stupid because the data collection and reporting was biased, with an institution (the Chinese gov) that's known to lie and cover up data regularly having its finger on the scales. Yes, if you only report test data from one area, you only get cases from that area, amazing.
(2) is stupid because none of the people involved had the technical understanding required to even interpret papers on the topic.
Thus, I don't have much respect for that debate.
I think you're comparing apples to oranges here. When people get into topics where they have no personal contact with reality, you should be comparing to fields where experiments aren't feasible and people are mostly pontificating. When people read books written by people that are crazy or lying but popular, you should be comparing to fields where professors are faking data to get publications - while also trading citations, of course. When you compare individuals to something like experimental physics or civil engineering, you should be comparing that to something like cooking food.
The thing is, I've noticed that a lot of the curriculum in schools is blind imitation of high-status science-y people, which can end up as cargo culting. Textbooks and classes have students memorize terms for certain things because "smart people know those words, and we should make kids be more like the smart people" - but they use those words because they understand underlying concepts, and without that the words can be useless.
This reminds me of Feynman talking about textbooks.
Now that we have the internet, you can simply see what kinds of technical language experts use and learn about those terms yourself. The distinctions "autodidact" is supposed to imply might have been weakened by the internet, and for that matter by libraries.
Of course, you'd also have to pick the experts you're going to imitate on your own, but then, you also have to pick professors or schools or textbooks. Following a societal consensus about competence should work about equally well either way.
Autodidacts are easily recognized by their unknown unknowns. They may know a lot, but what they don't know, they usually don't even know that it exists.
I remember talking with a director of a Max Planck institute. Ion transport was relevant to the conversation, and I said something about how "of course, while lithium ions are light, that doesn't mean they diffuse quickly because they're strongly bound to their solute complex, more than sodium ions". He said, "Aha, I see you're well-educated". I thought that was funny because that wasn't something I ever took any classes on or got mentoring about. The other funny thing about that conversation was that his pet project was this energy storage idea that had no chance of working because he didn't understand ion solvation well enough.
These days, professional exams have been largely subsumed by university degrees, but there used to be people who'd just study on their own for a bar exam, highway engineer exam, actuarial exam, etc, and then go work in that field. How would that fit into your classification?
I too like Sam Altman, his writing and the way he often communicates. He strives for what he believes is good.
I'm really not sure why you (or other people) say that. I saw, for example, his interview with Lex, and my impression was that he doesn't care about or understand AI safety but he memorized some phrases to appeal to people who don't think too hard about them. Also, that he's a good actor who can make it seem like he cares deeply; it reminded me of watching characters like Sam Carter in SG1 talk about physics.
I would add I am a big fan of his medical and fusion efforts.
I'm not, but that's because I understand the technical details and know those approaches can't work. In any case, my view is that his investments in those were driven largely by appealing to people like you.
Anyway, I'm wondering why people don't exploit this?
I think people usually do this with art instead of stocks.
Yes, I'm defining "market failures" as things that more markets don't solve, but macroeconomics and sociology aren't fields where such things are mathematically proven. We have to go on empirical evidence, and my interpretation of the empirical evidence is that such market failures exist and are significant.
Market failures that the market doesn't subsequently correct aren't a big enough problem that something "needs to be done" about them, especially given the risk that government intervention will make things worse.
So, we're back to "giving up on improving" market failures, because you don't think they're actually a significant problem?
Anyway, most people disagree with you here. In the real world, you're outvoted and will lose elections, and I guarantee the above argument won't change most people's minds.
How many horses were there?
Also, development of those 3 technologies wasn't limited by available power.
New technology could also create new market failures, eg by enabling new kinds of collusion or creating new externalities. You could see the RealPage suit or PFOA pollution for examples.
Is your proposal to selectively direct technological progress towards techs that solve market failures and away from techs that create them? How would you do this besides...government involvement?
Non-governmental actions are either markets, which won't solve market failures, or they're non-market solutions, which are either governmental or pseudo-governmental with the same potential problems as government solutions. Or can you give an example of a solution to a true market failure that avoids the problems that government action has?
a general rule of 'the government shall not attempt to do anything about this category of problem'
Is that not "giving up on improving" market failures?
The argument of this post seems to be:
"There are problems with the results of markets, you say? Sure, but government probably won't make things better, so we should just give up on improving the situation."
That is dumb. Government actions change depending on the pressures on government, and they can be improved by popular awareness of issues. Saying "stop thinking about issue X" is never going to be a good strategy - not rhetorically, practically, or philosophically. What you can reasonably say is "problem Y is currently more important than X, so we should focus on pressure against Y for now". That approach involves clearly elucidating problem Y, its causes, and what the effects of potential mitigations would be.
Had a practical steam engine somehow failed to materialize during early industrialization, the first industrial revolution may very well have gone down in the annals of history as just another lost "golden age"
No, it would have been entirely possible to skip steam piston engines and go directly from water wheels to internal combustion engines, large steam turbines, and/or hydroelectric power.
The most promising approach is using genetic material from smart people to make children, but there are the obvious social issues with that. In general, I suspect that human intelligence is limited partly by increased value drift leading at some point to people having fewer kids.
The 2nd-most promising approach is genetic screening of embryos and embryo selection, but currently that's only understood well enough for smallish improvements, and if it was common enough I suspect it could lead to problems from reduced genetic diversity.
I was not impressed by the first linked post. Many relevant genes only affect early brain development. Genetic engineering of a fertilized egg cell is much easier, and significantly improving intelligence that way is still not a thing considered practical to do today. The author's understanding of biology is poor by my standards.
I'm also not optimistic about Neuralink:
- Implants for input are worse than vision and displays.
- Neuron development for high-bandwidth output requires direct feedback to neurons, which for muscle control comes from proprioceptors. That feedback is in the form of complex patterns of chemicals being released, which a brain implant wouldn't be able to do even if those patterns were fully understood, which they aren't.
On a timescale of 1-15 years, if you want smarter people, the best you can hope for is probably better education. I think that AI training has some insights for designing education, and I suppose understanding AI architectures and operation has made me a little bit smarter.
When I posted this, I actually thought I'd get stuff like
- "Do you think [startup]'s technology is practical?"
- "What do you think the best approach to [problem] is?"
- "Do you know of any notable promising inventions regarding [subfield]?"
because that's what I've gotten by DM elsewhere, but apparently a public post on LessWrong is different.
I linked to a thread on twitter, but what I was pointing at was not primarily the thread but the person. He also wrote replies to many arguments.
it shows that it was known since 2013 that coronaviruses could infect humans without host animals.
Some amount of binding to some human receptors doesn't mean a virus can infect humans. When something can infect humans, people say that instead.
Also, you're not using those words correctly. Perhaps you mean "without evolution in an intermediate host" - which was, of course, never found.
Roko's argument and the Rootclaim debate are both low-quality. For a stronger case for COVID coming from a lab, you can see Prof Richard Ebright.
...no?
- I'd expect an Altman-led AI chip maker to be worse than other companies and get outcompeted.
- If $7 trillion of investment isn't enough for oversupply or at least commoditization of chips for AI, then that's a world in which property rights of AI chips are no longer guaranteed.
In any case, this post isn't even evaluating investments by profitability, it's evaluating them by societal consumer surplus. Those are very different: 1 involves incremental value from current production, while the other involves marginal value from future production with the investment.
"This is about how to spend money on AI safety" isn't the point of the opening of the post. It's more:
- Here's some stuff I'd talk about anyway (like thinking of the economy in terms of energy flows) and a convenient way to frame it that was in the news!
- Wow, we could also spend money on not maximally accelerating AI!
- I see a lot of people saying that AI is urgently needed to solve [problem] like global warming, but here is how you solve some problems by solving the problems.
AI is something I've thought about a lot, but I think I've already posted everything about that that I want to, and people didn't seem to appreciate this that much.
Do people ever try to pitch you on projects
Not exactly. I've evaluated project pitches for other people, and sometimes people pitch me personally on ideas. To the extent that I've been trusted on technical things by people directing large amounts of money, it's been a narrow/limited sort of trust, not "let's do whatever this guy says".
Where are the investors that you expected?
Relative to VCs, I think (private equity that quietly does large acquisitions) and (internal project evaluations at engineering companies (eg BASF)) do better technical due diligence.
But internal project evaluation at big companies is conservative yet typically requires 15%+ ROI for approval - which is a result of managers' job cycles. As for private equity, people seem more interested in acquisitions to establish enough of a monopoly to have pricing power than new technology. Which is smart.
The actual answer to that question, tho, is that they don't exist to the extent it seems like they should, that cultural factors prevent an economic equilibrium. Elon Musk is dumb enough to think the Hyperloop is genius, and there's free money on the sidewalk for people with that level of technical knowledge who also have either enormous wealth or credibility with and connections to investors. But the selection processes involved reject them at earlier points.
Is this because of that thing about multi-stage processes and Keynsian beauty contests?
Yes, that's a big factor in how the current situation got established. I thought of that independently and have mentioned it before.
Is your idea that if some entrepreneurial LW users swallow this particular red pill, they will be less held back by their maladaptive honesty and be more competitive in raising money, and that this will result in more rational entrepreneurs?
Partly that. Partly it's an example of irrational thinking that people might see in their own approach. Partly other things.
Have you read The Scout Mindset? entrepreneurs who honestly and publicly gave low odds of success
I haven't.
Bezos gave some of his investors a 70% chance they'd lose their whole investment. Those investors...were his parents.
Elon Musk was hooked up to the Palpal Mafia social network.
Anyway, a lot of stories like that are misleading. My understanding is that those examples are mostly just their after-the-fact disclosures of their private thinking, not what they told investors at the time?
A snarky reply to that would be: "OK, now put that on a powerpoint slide with 3 bullet points."
Part of the point of my post is that the (reasoning follows) can't actually be a real analysis in a pitch like this; to be appealing, it has to be a narrative that fits people's preconceptions, a story that's the kind of thing the executive-types being pitched to like. The part with technical details seems to just fill the same role as technobabble in Star Trek. Sometimes you'll see sales guys who do the pitch listen to the engineering stuff and just pick out a couple keywords to use in the story they'll make up.
Considering saturation fields, maybe 2x the field strength of Nd magnets would be possible, meaning 4x the energy density. Obviously superconductors can be much more magnetic than that.
If it was 100% efficient, then air resistance would be the main loss. But the losses from something like: (motion -> hydraulic cylinder -> hydraulic motor -> electric motor -> power electronics -> energy storage -> power electronics -> electric motor -> hydraulic motor -> hydraulic cylinder) are pretty substantial.
Springs are more efficient, and animals use protein elasticity to store some energy while walking/running.
Yes, most of those clever mechanical mechanisms were invented 150 to 60 years ago, but there are a few newer ones, eg the Thompson Coupling from 1999. Also, the specific torque of planetary gears has increased from improvements in materials, lubrication, and modeling. There's also been increased usage of designs that aren't technically new but weren't practical in the past, like roller pinions.
The actuators I mentioned were developed by a private community I was invited to because of my blog. I don't have permission to post the designs, but I can list some relevant properties if you're interested.
Thanks.
Yes, in some cases it's possible for a robot to have proportionately lighter limbs, but fiberglass or carbon fiber would be better than metal tubes, and greater material strength is offset by increased height. Maybe you're underestimating bone; it's less dense than steel and its specific strength isn't always worse. It's possible to get higher speeds with parallel robots like delta robots, or run long cables, but there are real tradeoffs between series and parallel kinematic chains that often justify putting drive systems out on limbs.
Or maybe you were thinking of decreasing overall mass evenly. It's certainly possible for humans to be relatively skinny for a given height. Similarly, it's possible to drive hydraulic cylinders very slowly, but the goal here is specifically a human-like robot, with similar gaits and payload capability. Note that air resistance is also an issue if mass is much lower relative to height.
My math for pendulum swinging was just calculating step cadence, not walking speed with 2 legs - which is doubled that way, yes. I was just making a point about the relation of pendulum frequency to walking dynamics.
That's just how pendulum frequency scales with size. You have a sqrt(X) period with X distance, so sqrt(X) speed, so X energy over sqrt(X) time.
You can also consider a jumping robot, which accelerates at a constant rate for a fixed % of its height. The final speed and time taken are both sqrt(X).
That approach just adds a loss from a GAN-type discriminator to the distillation loss when training the distilled sampler. Coalescer models should be able to do generation in at least as few steps as consistency models, so, if you want, you could do the same thing with coalescer models: initial training as normal, then more training that adds in a loss from a discriminator.
I see, that's what you meant.
Tenuous far-mode thinking? At this very moment, I see people using ChatGPT to emails and reports that have wrong technical statements but are written in a way that appeals to people who don't know they're wrong. Even medical students are doing this. This is even increased intentionally: commercial LLMs have a RLHF process that in some ways makes them more useful but in some ways makes them less accurate.
You seem to be misunderstanding the point of my post. The title isn't about AI being better at politics than politicians, it's about AI being ineffective/incompetent at accomplishing desired goals but chosen by selection systems, like how some current politicians are. It's about whether AIs would have similar types of flaws to current politicians.
You don't ask "an AI" to "make a technology". You choose a possible AI to make that technology for you, based on some criteria. Saying "the problem is just alignment here" implies that all of those potential AI selections are capable of making the technology for you. But what's actually being selected for is getting selected. Why should we expect an AI's output to be a real answer instead of some BS that appeals to people more than a real answer and is easier to output?