Comments by "Mikko Rantalainen" (@MikkoRantalainen) on "Sabine Hossenfelder" channel.

  1. In the meetings I have to attend, I'm intentionally going with declaring my opinions pretty quickly with a prefix roughly along the lines "My first feeling is that ..." – I've found that it helps other people to express their non-final opinions, too. And I think it pretty clearly conveys that I'm willing to change my opinion during the discussion.
    129
  2. The only thing worse than H2 energy storage is not having any energy storage. Here in Finland, we have way too much wind energy in the market and when we get a Winter day without enough wind, the energy prices skyrocket instantly because during windy days the energy prices are low enough to push older technologies off the market. Daily fluctation can be price increases around 50x or 5000%. That's a bit more than acceptable inflation! And I fully agree that about 10–30% of end-to-end effiency of H2 as energy storage is close to receiving "worst battery ever" award. Li-ion batteries have end-to-end efficiency close to 90% which makes them clearly superior technology even if pretty expensive method.
    55
  3. I've been doing web development professionally for 20 years and it seems obvious to me that AGI will happen sooner than semantic web where the web site owners at large supposedly start adding correct metadata to the content to make it easier to understand for computers. As long as the content is understandable for humans, it will be understandable for AGI, too, and no extra effort is needed to add semantic markup.
    47
  4. If you want to make it totally "fair", you have to first decide what you're actually trying to measure in the competition. Best genes? Maybe disallow all training because that can obviously give you unfair advantage over the performance of another competitor with fully natural state body based on genes only. If you want to measure who has improved their performance the most (best training or most work), then you would need to compare their performance to their natural untrained genetic level. In that case the winner could be the competitor that appeared to finish the 100 meter run as 6th. I think the part about entertainment value around 11:30 is the most important one. If sports events were not broadcasted at all, would they still be considered important overall? I would say no. For most people, the importance of sports events would then be about the same as the winner of local Checkers Club tournament. As a result, the winner is defined by whatever method that's generally considered the most entertaining. And that's subject to discussion as time goes by so we get constant rule changes.
    46
  5. 2:50 If you use encryption algorithms that support perfect forward secrecy, the full symmetric encryption key is never transmitted over the network so future adversary cannot decode the encryption without guessing the correct symmetric key. Algorithms that support perfect forward secrecy cannot be used offline, though. And it currently seems that not even quantum computers can help breaking symmetric ciphers such as AES-256.
    22
  6. I sure hope Helion can get their technology to work for real because ability to generate electricity directly from the fusion without using high-tech steam engine sounds really good.
    14
  7. Except we know that fusion reactors are at least possible to create because nature has lots of examples. We still don't know if a real practical quantum computer can be ever done. A few qubits for a few microseconds is a cool toy, not a practical machine.
    10
  8.  @briseboy  Most people I've ever met are really bad in statistical probabilities so that cannot scale. The problem is that people try to apply common sense and think that it's same as statistics.
    8
  9.  @BillyViBritannia  Good points. I would like to add that I'm not in manager position and I don't want to be. My position is more like specialist/expert. I agree that if the manager starts by stating his or her own preference, going against that would be hard for many people.
    6
  10. It reminds me of video titled "Someone improved my code by 40,832,277,770%" where Matt Parker created and executed a program that took about a month to compute the results for a problem Matt had. And the some nerds optimized the code a bit and it can now run in 300 µs instead of 30 days.
    5
  11. What I find interesting is that flat earthers are happy to accept that they can trust smartphones and GPS navigation but somehow the same science that allowed creating those is then assumed incorrect when it comes to the shape of the Earth. If they went full on Amish on everything, it would seem more logical and the explanation that they just want to invent everything by themselves would seem applicable.
    4
  12. Another great video, great work! I'll be sharing links to this video for sure. One suggestion for the video production: it sounds like your mic setup is prone to have sibilance issues so I would recommend adding de-esser filter to your audio post-production setup.
    4
  13. I thought the Amazon delivery time is already unpredictable enough to demonstrate it must be running on quantum physics.
    4
  14.  @marco21274  Nuclear is expensive because politicians want to enforce stupid decisions. For example, Finland built the new OL3 reactor because politicians wanted to allow exactly one reactor and to get the required output they had to double the reactor size. The project was 18 years overdue because doubling the reactor size turned out to be a bit harder task than expected. The OL1 and OL2 reactors at the same location were built in 1978 and 1980 and the combined output of those reactors matches OL3. The old units too 4–5 years each during 1970s so it should have been safe to have similar schedule for new duplicate reactors but that option was turned down by politicians. The current plan for OL1 and OL2 is to keep both in full production until year 2038 so those reactors are still considered safe.
    4
  15. I think even a huge battery system built from recycled car lead batteries connected together would beat end-to-end efficiency of hydrogen storage. I agree with Sabine that this is more about procrastination than anything else. I think the future will be flow batteries but for short term solution, li-ion batteries are the least worst option.
    4
  16.  @justaskin8523  Sounds pretty similar to my experiences! Though the percentage for "this is not high enough priority right now, postpone" is even higher for our team. The problem with that style is that in long run, you'll have a huge pile of postponed things simply because no couldn't make a decision "we'll not do that at all".
    4
  17.  @SabineHossenfelder  I think short videos for short updates are fine. Maybe add a link or two from the description to the older related videos for more context?
    3
  18. 6:10 I think these examples are not truthful. If you took 17 year old from year 1200 and put him or her into a modern car, it would take quite some time to teach them to drive. Similarly, a ten year old kid from year 1200 would have hard time learning to use a dishwasher oneshot. It's only because those 10 and 17 year old kids have seen a lot of examples during their lives they have pretty good idea what to do. As I see it, the LLM matches the human thinking part pretty well but the AI still needs improved vision system to feed suitable data to the thinking part and the AI also needs methods to interact with the world. A human baby around the age of a couple of weeks still seems to fail to understand that he or she has these things called limbs and may be failing to understand the concept of self. We don't yet know for sure. And as a result, we don't know how close to AGI we really are. Even if we had some commonly agreed on definition for AGI, which we don't have. In fact, humans still don't have even commonly agreed on definition for even human intelligence. We have IQ tests that actually measure g-factor and we sometimes pretend that it's the same thing as intellligence but we don't really believe it in the end. Considering we still don't have cost effective way for inference step in LLM (equal or better energy efficiency compared to human brain), the big question is who can afford to run AGI even if it were invented during 2025? Let's assume that OpenAI can scale ChatGPT o4 or o5 to true AGI level but it casts $15000 per task to run because the amount of compute needed, who is going to run it for anything?
    3
  19. Unless totally different rocket engines will be inventend, I don't think even the best rocket engines have high enough specific impulse to escape the gravity of Jupiter – even without any payload!
    3
  20.  @oneukum  Finland already built new OL3 reactor to Olkiluoto nuclear power plant. It was finished only 18 years overdue. If we start a new project that ends up with similar schedule in practice, we can expect to see electricity from the new reactor around year 2040. We need something before that.
    3
  21. 11:28 "Probably the only fat burning excersice that actually works" Nice burn!
    3
  22. Like a scientific version of Louis Rossmann.
    3
  23.  @timetraveler_0  That's true that we don't yet know when AGI will happen. My current guess is around year 2030. If AI research gets equally good funding as during last couple of years, I think it's going to happen. The only question is how much electricity you need to run human level AGI. If you need a supercomputer to run one instance of AGI, it will be used for AI research only because it will be much cheaper to employ humans. In long run AGI will be cheaper to run than human minds, of course. But we're nowhere close to that yet.
    3
  24.  @phoneisnotketchup  I'm sceptical, too, but I think humans still understand plasma physics so poorly that I wouldn't claim it's totally impossible yet.
    2
  25. Yes, but in the case of LLM based systems, the magic is in the training process (creating the really complex lookup tables that can be used to extrapolate the answers not directly included in the tables).
    2
  26. Great video! I agree that it would be really important to clearly set out the limits of our current understanding. I loved the prediction test: "What theories did this measurement invalidate? None?".
    2
  27. As a Finn, our house has only two heat pumps: air-to-air and exhaust air heat pump that heats the floors and water. And making house air tight is probably even more important than the amount of insulation. Unfortunately, fixing a poorly built house that has lots of airleaks is next to impossible when it's already built.
    2
  28. Great video as usual! I would assume li-ion batteries and flow batteries will win for the storage in near term future. As li-ion cell costs should drop / already dropped below $100/kWh, I would assume building li-ion storage units should get closer to $100/kWh. The efficiency of the storage is very important, too. If the efficiency is only 25% (e.g hydrogen), the cost of electricity from the storage is at minimum 4x more expensive than the actual energy production. The li-ion storage is highly efficient so it makes a lot of sense. Flow battery should allow for even lower cost to manufacture than li-ion and about the same efficiency. It still has technical issues (in the membrane if I've understood correctly) so I wouldn't call that technology production ready yet, but the theory seems solid and I don't see any reason why flow batteries couldn't be the long term solution. It sure looks more promising than generating hydrogen for energy storage. (The only exception is generating hydrogen to manufacture fully synthetic fuel for existing cars and airplanes – in that case the CO2 reduction may be more important than ultimate efficiency.)
    2
  29.  @Nill757  I guess the word typically was a poor choice there. I meant that's the best you could expect to find in reality and in many cases, storage will be even more expensive. I mean that "typically" when you read about some new product, that's the best case scenario. And yes, you're right that the prices have increased recently. According to BloombergNEF, year 2021 was the cheapest ever and they reported 108 USD/kWh for the cells and 141 USD/kWh for the battery pack. Since then, the cost of cells has been increasing.
    2
  30.  @Nill757  I agree that inverter efficiency differs maybe 15% at max between cheap and expensive units. However, a better inverter also means all your devices using electricity actually work correctly. You don't want the whole house grid voltages to jump around when the compressor of your freezer kicks in or the hot water heater. Generating stable AC voltage when the currents change all over the place is hard and cannot be compensated with lots of cheap extra battery capacity. See specs for a good UPS: you should expect to see prices in multiple thousands of dollars for units that can output 4 kW and typical time until battery is empty is less than half an hour.
    2
  31. Even 100 µA would be interesting because you can literally take full power at all times. Use it to charge a 3V supercapacitor and run more power hungry parts in short bursts. Of course, that's not good for smartphones that you use for hours per day but could be enough for taking a phone once an hour for hundreds of years.
    2
  32. 2:05 I still fail to understand how flat earthers explain sunset or sunraise. It's pretty obvious that it's not a spotlight above the ground during the sunset/sunrise, yet pretty much every flat earther related video uses the spotlight explanation.
    2
  33.  @MrHuddelchen  But you can easily see that half the Sun is falling/raising behind the horizon. That cannot be explained by refraction in any way refraction normally works. All the arguments for flat earth seem logically identical to creationists that try to claim that they have scientific "evidence" for their beliefs.
    2
  34. On possible interpretation is that female sports are a series for disabled people and you just have to specify more accurate rules for the accepted disabilities instead of simply looking at legal gender of the contestant.
    2
  35.  @MrLightZenith  Finland doesn't have any mountains and the green party is totally opposed to any new hydroplants, no matter how small. Politicians directly caused the current situation with heavy monetary aids to allow building wind turbines without market demand. However, those same politicians totally failed to consider what happens when there's no wind, probably because wind turbine lobbyists were claiming that "it always winds somewhere in Finland so we only need to build turbines everywhere". In reality it turns out that sometimes it winds nowhere in Finland. Finland has theoretical wind turbine capacity around 7000 MW and if you're unlucky with the wind, you get only about 1% or 70 MW total output. You have to somehow generate the missing 6930 MW using other means and in practice Finland cannot do that reliably which causes electricty prices to skyrocket until somebody stops using electricity because of high costs. That's not a solution, it's a workaround.
    2
  36.  @iam5085  That's right. Or to be accurate, the max power depends on the overall grid workload and how close to max power the international powerline connections are. The idea is that if OL3 suddenly drops out of the grid (e.g. main transformer catches fire), the rest of the grid must be able to balance the loads without causing blackouts or voltage drops. And OL3 at max power is such a huge power source that it's really hard to balance the whole grid if it suddenly drops out of the system. This is yet another example why it was pretty stupid decision by Finnish politicians to require building one huge reactor instead of multiple smaller ones.
    2
  37. Automatic verification of responses is definitely what ChatGPT is missing. ChatGPT is very good at imagining stuff but it's missing verification process to check if the response is actually based on known facts or trusted references.
    2
  38. I keep saying that I don't trust science. I trust scientific method. Scientific method is self-correcting but sometimes it takes a lot of time.
    2
  39. Great summary of the problems! The only think I would have liked to see is some kind of ballbark estimate for the total cost and weight for the hydrogen technology vs EV battery tech. As I see it, hydrogen car would need the tank which is both heavy and expensive, fuel cell which is highly expensive and a small battery bank and/or super capacitor bank to be usable in practice. This is because fuell cells that can output 100–800 kW peak power to be competitive with EV cars are not in the price range that can be used in the cars in practice. Hydrogen car needs exactly the same tech as battery EV cars because hydrogen + fuel cell (+ smallish battery/capacitor) is simply a replacement to the battery in battery operaterad EV car. One additional problem with fuel cells is that the output voltage is DC and pretty low voltage. To get it to about 400–800 V range for the motors to be competitive with battery operated EV cars, you also need very exotic inverter or an additional high power DC-DC converter. I think wind turbines will be used to generate synthetic fuel (diesel, gasoline, jet fuel) but pure hydrogen is actually pretty poor energy storage format and as you correctly pointed out in the video, hydrogen is only a storage format, not an energy source.
    2
  40. I think Helion Energy gets funding because the idea sounds plausible and ability to skip all the spinning generators and instead get electricity directly from fusion sounds really good if they can actually get it working. I would guess most people funding the tech do not actually expect working fusion but instead some other related stuff that can be patented and licensed for big enough amount of money to make the investment worthwhile.
    1
  41. Great video and I'll sure link to it in future when I see misunderstanding between total power output vs plasma output. I was already aware of this difference but I was never able to express the difference as clearly as you did.
    1
  42. 5:02 PSA of the year: "So please do not eat used nuclear fuel rods. It’s not healthy and also probably not all that tasty."
    1
  43. I think there's no need to have more data to train AGI. You can just take the existing data and run a big cluster of near-AGI AI systems discussing about things together. This works for human researchers so it seems perfectly plausible that a group of near-AGI AIs could try to figure out new ideas and the rest of the group could cross-check the ideas with existing knowledge to weed out hallucinations. This is no different from MuZero playing against copies of itself to learn more. All you need is insane amounts of computing power. And as long as high end AI systems can output a couple of tokens of output per second with computers costing close to $1M, I don't see intelligence explosion happening too fast. However, if we can run multiple copies of near-AGI in a cluster, they can communicate between each other much faster than humans which does speed up things somewhat. If you try to fix the speed issue by increasing the amount of hardware you run, you just made the resulting AI less competitive because it just got a lot more expensive.
    1
  44. Agentic AIs will succced because most people are lazy. And by succeed, I mean make lots of money for AI companies. If you use "deep research" agentic AI of modern AI services you get nice results but you have to read long output. And if you're too lazy to do that, you just tell the AI to figure it out and "execute the best option" without even telling you what it did. Hopefully something good!
    1
  45. The problem with quantum computers is that if you think creating custom software for traditional computers is expensive, you've seen nothing yet because creating correctly working programs for quantum computers is a lot harder.
    1
  46. I find it funny that people are trying to explain what causes consciousness even though they cannot even clearly define consciousness. How could you correctly explain something that you cannot even define? Most commonly used "definition" seems to be "awareness of internal and external existence". I think even ChatGPT 3 exceeded that level so it cannot possibly be the definition most people actually think when they use the word.
    1
  47. IMO, I think this can be battled the same way AI should be trained even today: start with a small amount of trusted data (hand selected articles and books from various fields verified by experts). Then for every new piece of content, make the AI estimate the quality of the data (can it be deduced from the trusted data?) and skip training if the content is not (yet?) considered high enough quality. Note that the quality of the data is about it being well supported by the existing knowledge. Do this multiple times to create chains of high quality data (that is, the potential content was not deemed trusted earlier but now that the AI has learned more, it will estimate the quality of the same data differently). Keep track of the estimated quality of a given piece of data and recompute the estimated quality again for all documents every now and then. If the quality estimate of the AI is good enough, it should increase the estimated quality of the content over time (because more chains allows accepting more new data) and cases where previously trusted content turns on untrusted later would point out problems in the system. Also run the estimation against known good high quality data not included in the training set every now and then. These should be considered high quality data but the AI may fail to identify the data correctly, which would demonstrate lack of general understanding by the AI. Once you demonstrate that the estimated quality matches well enough with the expert evaluations of the same content, you can start to train the AI to understand misunderstandings of humans, too. Train low quality content as examples of humans failing to think / research correctly. In the end, you should have an AI that can successfully estimate quality of any new content and automatically use it to either extend its knowledge (chains of known good content) or to automatically learn it as an example of low quality content that the AI should avoid but the AI should be made aware of. If the AI doesn't have negative feedback from failed human content, it cannot understand failures in tasks given to said AI.
    1
  48.  @levelwithz3779  🤡
    1
  49. One thing that would deserve more public knowledge is difference between currently used light water reactors that cannot use natural uranium and cause lots of nuclear waste. Fast breeder reactors and heavy water reactors are not similarly picky and can use natural uranium and produce less waste. And when built correctly, you could burn the "nuclear waste" from light water nuclear reactors (remember, those things that we currently use) as fuel in heavy water reactors. And the reason heavy water reactors are not used? Mostly political issues in form of nuclear proliferation. The thinking seems to be that because heavy water reactors can be used in theory to manufacture weapons grade uranium, too, we should not build any. Not even if those could burn the "nuclear waste" we have today. Instead we're planning to bury all the waste into the ground! Update: I see you discussed some of this near the end of the video.
    1
  50. I predict that AI will replace humans in long run. Not because AI goes for killing spree but humans opt out not to reproduce when humans figure out they are inferior on all metrics compared to superhuman AI and robots. Once superhuman AI can do better research, better poetry, better movies, better vehicles, better music, better medicine and have any kind of robotic body without need for food or sleeping, why would any human feel anything but inferior? I still think we'll see superhuman AI around year 2030 but it will not result in intelligence explosion because if you look at ChatGPT it's far from instant in its "thinking". Superhuman AI will improve research rate of then current AI but it will not be instantanous but will take months or years to improve to even higher levels. I would expect superhuman AI to first figure out next gen architecture for the superior AI system, then design an improve lithography process needed to manufacture the parts and then somebody (ASML?) needs to create the new process before you can even start building the new parts. Without new improved parts, running ever bigger AI systems requires so much electricity that it's already becoming the limiting factor.
    1