Comments by "Mikko Rantalainen" (@MikkoRantalainen) on "Lex Fridman" channel.

  1.  @itdepends604  Some say that there's no math category in Nobel prizes because a mathematician married the girl that Alfred Nobel fancied after. I guess Alfred was a bit bitter about that and excluded mathematics from the prize.
    19
  2. 1:10:45 I love that at least somebody (Aella in this case) actually checks the scientific references instead of blindly trusting them!
    9
  3. Hospitals are not free even here in Finland. Staying in hospital in Finland costs 54.60 EUR per day regardless of treatment so it's basically same as living in cheap hotel room except you get all the treatments you need included in the same fee.
    7
  4. 17:10 Not a single question about these unbased claims? Shame on you, Lex.
    6
  5. 8:20 Basically "Trust me bro, I'll make a deal. But I will not tell about how it would work, you just have to trust me." from the "leader" that has trouble finishing any sentence without at least a single lie.
    4
  6. 30:10 I currently think that the major selling point of all religions is to not to have to say "we don't know". For some reason introducing supernatural as the "explanation" is less frightening for many people than saying "we don't know".
    4
  7. This video reminds me of Tucker Carlson's interview with Vladimir Putin. Unfortunately. I think the AMA part would have been deserved to be in another video because it was great!
    4
  8. As a software developer/architect, I'm looking forward for the time where I can make only architectural decisions and AI can play the role of code monkey that implements the functions/methods for my design. I'd definitely do code review for every single line that the AI emits but I have no need to be the one that writes every line of code as long as the AI can emit code of high enough quality. I'm already trusting my collegues to write code and it rarely ends up the same I would have written but as long as it's readable, understandable and executes with high enough runtime performance, I'll accept the code. That said, I do complete code review for the code of my human collegues, too. And they verify my code, too.
    4
  9. 2:00:00 Great points about the vaccination. The way I interpret the current results is that vaccinating children probably reduces risk for Long Covid (LC) which would be bigger risk for the development of the children than the risk caused by the vaccination. The problematic part is that we don't have reliable statistics for LC and even less reliable statistics for the effectiveness of reducing the risk of LC via vaccination.
    3
  10. I'm a software developer (44m) that used to think about 5 years ago that we'll have AGI around year 2050 but I'm currently thinking we'll have human level AGI (IQ 100) around year 2030. I'm currently thinking that around year 2050 we'll have superhuman AGI that can do everything better than any human being. That is, a single AGI system can do philosophy, mathematics, compose music, create movies, design new robots, write new software and control ships, airplanes or fighter jets better than the top of the top people of each respective field. And as a corollary, it will create better improvements to itself than any group of humans could do. I truly believe that humans are a boot routine for a superhuman AI. And I think it's a good thing as a whole because I see only science, culture and art as the final output of humankind. If superhuman AI is better for each those fields, we should build the superhuman AI to advance those fields.
    3
  11. Great interview as usual. I think this kind of interview would deserve a lot more information in the description, though. For example, list all the books (titles and authors) mentioned in the interview. The same should apply to all social media accounts of other people mentioned by Grimes. As long as you clearly label those links in the description, it should be clear which is content related to your channel and which is content related to this interview.
    3
  12. Great interview! As an European, Bernie Sanders is probably the only US politician that I would support. He's not perfect but sensible enough to be worth supporting.
    2
  13.  @rafaelmoura947  Finns are pretty well known for overly humble wording for their estimates even when then the estimate is actually accurate to about 100 ns.
    2
  14. 1:06:00 Imagine future when you will need to scan players for e.g. Neuralink implants in their skull to make sure they don't communicate with computers. I guess we cannot ask people to shave bald and then search for scars in the skull either. Maybe the war has already been lost and we shouldn't be playing full knowledge games for money because cheating will be so lucrative if we keep doing that and the cheating methods will be harder and harder to catch. Or maybe full body MRI scans before entering the room and broadcast is delayed 1 or 2 turns.
    2
  15.  @toddhansen3131  Even if the voters did research before voting, money would still get more visibility for a candidate. You cannot expect everybody to do proper research about every candidate so publicity is required to first seem sensible enough option to warrant research.
    2
  16. I'm not sure if I'd call Elon a genius but he's definitely very smart nonetheless. The most interesting thing for me (as a software developer) is the amount of details he knows about Tesla FSD implementation details and he still uses lots of his time on SpaceX. If Tesla FSD were the only thing Elon were doing it would be much more expected to know this kind of details.
    2
  17. 1:07:11 "I wanted to make a bold claim that was precise and testable and correctable not a wooly another woolly information in biology argument information touring machine blah blah blah blah blah... A concrete series of statements that can be falsified and explored and either the theory could be destroyed or built upon." Respect! This is gettting too rare in modern science.
    2
  18.  @cyclamen831  The cost per night here in Finland is set so that rare homeless people don't get into hospital instead of getting a home to save money. Finland also has Housing First program to make sure nobody needs to be homeless.
    1
  19. 51:00 Interesting discussion. Maybe each world championship match should be best of 7 games and time control would be 1 hour per player, no increments? That would give theoretical upper limit of 7 hours per player in total and you could lose one or two games (and three in theory) and still win the player pairing. And 1 hour should be enough to allow thinking for important situations but wouldn't allow calculating so much for every move. And the rating change (win, lose, draw) would be computed for the whole match (up to 7 games), not as a result of a single game.
    1
  20. You should definitely play Fallout 4 the same way.
    1
  21. 1:35:00 Listening to this discussion is a bit nostalgic because my first proper computer programming langauge was Pascal (to be exact, the Turbo Pascal variant by Borland) and I wrote simple realtime 3D graphics using mode 13h graphics (named after the value you had to set for the AL register before executing int 10 software interrupt, the key point was that this graphics mode had linear memory layout and didn't need paging or any other tricks to make pixels to appear on screen after CPU wrote the bits). In that time I wrote most of the program in Pascal but it supported inline assembly and I wrote the inner loop in assembly. I had one physical paper book about assembly instruction timings (basically how many clock cycles each instruction took to execute) and tried to come up with a sequence that was fast enough. And yeah, I considered id software games marvellous because they seemed to be able to compute maybe 2-8 times more pixels per second than my code. The major problem that Pascal had was that it only supported real mode 16 bit programming and once I started to do texture mapping stuff, it quickly become clear that I had to switch to Watcom DOS4GW stuff to allow use of 32 bit linear memory mode to use megabytes of RAM without headache. Hence C programming. Watcom C compiler also allowed use of inline assembly which which was nice.
    1
  22. The self-domestication isn't that novel idea. Studies from year 2012 already discuss about self-domestication of Bonobos with very similar rationale as the AI emitted here.
    1
  23. 44:19 I think the most sensible suggestion to fix laws is to require that law book must be short enough to be readable by common people. I'd argue that it would put it in class of about 200 pages of typical pocket book format with typical font size. If you want to introduce a new law and the 200 limit is already full, you have to replace some existing law instead of simply adding yet another law. At least here in Finland, the ignorance of the law is not considered a valid defence in court. If laws simply keep getting more complex over time, there will be situation where nobody can follow the law because the whole law is too big to understand. And at the same time, ignorance is not good either because that will get you sued.
    1
  24. 45:30 Some say that everything we have today had to be imagined by somebody in the history before it existed in the first place. If you cannot imagine something, there's no way you can design or build it. Therefore, to have better future, we have to be able to imagine a better future. It's good to be realist and also imagine possible future gone wrong scenarios to understand how big or small the difference between the really bad and really good future is.
    1
  25. I think the solution you had was the best you could do with the data you had. I guess the best step forward is to have two cameras recording in parallel so you have fully redundant backup of the discussion and in worst case, you simply have to use the non-optimal backup camera angle. Cameras are cheap enough that the cost of the extra cameras shouldn't be a big problem. You can even opt to cheaper lenses for backup cameras because it should be pretty rare that you have to switch to backup. As an option, you could use intentionally different camera angles for the backup cameras and in case all cameras work as intended, you have more camera angles to use for the final production video.
    1
  26. 2:22:20 I definitely agree with iterative progress but I hope GPT-4 were much more open. I think publishing the source code would be the best way forward but I understand that it seems risky move because there're also bad actors in the world. However, I think good people make the majority and bad party will fail similarly how bitcoin succeeds because majority of the miners are honest.
    1
  27. Looking forward for the full interview. This short teaser clip appeared really good!
    1
  28. I think we should be using "g factor" for speed of information processing (similar to clock frequency in case of computers). In addition we should be measuring how complex problems you can solve without time limit – that's something I would call intelligence test. The g factor I would call processing performance instead of intelligence. True intelligence test should have harder and harder tests and the result is about how hard problems you can solve at all. Of course, this all boils down to the fact that people cannot even agree what "intelligence" means. And that's why people will also disagree when we have artificial intelligence.
    1
  29. I think the audio/video sync is even more important for some people. In my experience, my threshold seems to be somewhere between 15–25 ms. If the sync is further from the correct, the result starts to bother me. As a result, I also hate bluetooth audio for most uses because it has higher latency pretty much in all cases. And people whose brains accept much bigger de-syncronization between audio and video often fail to understand why I'm complaining about the sync when it seems totally okay for them.
    1
  30. The biggest problem with Pascal was poorly designed system for strings. Different compilers interpreted strings differently and there were serious limitations such as Turbo Pascal having maximum string length of 256 bytes.
    1
  31. 1:52:00 The supply chain collapse is interesting scenario indeed. Here in Europe, thanks to the war that Putin started, the electricity production has dropped by maybe 10% and prices have gone up 5-10x already. It's hard to imagine the chain reaction if even a single supply chain failed even 99% instead of 10%. Total loss of any specific supply chain without a decent alternative would be catastrophic indeed.
    1
  32. 7:20 I think one possible simulation argument could be that "The fraction of all posthuman civilizations running whole universe simulations is very close to zero". If future posthuman civilizations run simulations of evolutionary history, they might simulate one human mind at a time and simply generate all sensory information on the fly. This seems like a logical software optimization because the information you could extract from the brain behavior of ancestors should be possible to figure out by simulating just one brain and change what kind of information you allow it to have (that is, change the existing memories and sensory feedback). This would require only simulating full chemistry of one brain and then change the inputs it can receive during the simulation. Humans have very poor I/O bandwidth so it should be easy to simulate all possible inputs for a single human. And since you can fake the experienced time inside the simulation, you could pause the simulation when it's about to go out of known domain and use some superhuman processing to calculate sensory feedback that the simulated human should experience for that situation. That would easily explain why we cannot create any physics experiment to show that we're inside a simulation – any idea we have, the entity running the simulation could just decide to output the results that appear to demonstrate real universe to the simulated mind. And since it's probably only a selected historicians who would be running this kind of simulations, the resulting humans in simulation would be close to zero relative to the history of humans. (It seems that this was later discussed around 25:25.)
    1
  33.  @alemz_music  I don't believe AI would attack humans. I think we can agree that human intelligence is superior to bears, dogs, cats or even chimpanzee. Still we don't try to actively attack those entities. Humans have attacked some species and killed them to extinction for stupid reasons such as trophy hunting. Superhuman AI should be able to notice that symbiotic life with humans would be a better solution than declaring a war. And superhuman AI should be intelligent enough to not to start collecting trophy kills. Much bigger risk seems to be that superhuman AI can create such a great entertainment that humans will be entertained to extinction. Imagine totally real feeling VR system where everything is nicer and more gratifying that anything in real world. Would you see the possibility that humans would want to spend so much time there that they don't bother having physical sexual intercourse and actually raising children? Remember that full body VR system would allow you to enjoy all the best parts of sex and interacting with (virtual) children without any of the con-sides. It seems to me that people are selfish enough that many would entertain themselves in VR environment instead of doing the extra effort to do stuff in real world. (And I'm not talking about current VR experiences, something more akin to stuff in the movie Matrix, without the human battery stuff.) Note that superhuman AI doesn't really need to be afraid of humans attacking it. Are you afraid that chimpanzees or cats are suddenly going to take over and kill all the humans? Note that superhuman AI wouldn't even need to care about issues like global warming – that's a problem for humans only and if humans decided to take no action, superhuman AI doesn't need to enforce it either. Even if it were the best for humankind.
    1
  34.  @graye2799  Did you understand what I was writing? The biggest threat is social problems – when (not if) AI can do all jobs better than humans, what do you think humans are going to do? That's a HUGE problem and it seems no country has done anything to deal with the issue. And that's going to happen within 10–30 years so we should be feeling hurry. I'm expecting to see lots of social issues because of AI taking care of all the stuff we often spend most of our days. I see Universal Basic Income (UBI) as the only practical long term solution but politicians haven't yet figured that out in any country.
    1
  35. 2:34:00 I totally agree that behind the scenes review process should be stopped. Just push the papers in public view and then look for reviews. The current system is designed by journals and it's used mainly to make more money for the journals. Even pretending that the first published version is complete is funny. If we switched to style where everybody publishes versioned papers and reviews are done for specific versions, you (as a reader) could see existing reviews and compare them to changes in the later versions, that would allow real dialogue between authors and reviewers. And doing this fully public would allow anybody to take part of any paper. Current system reduces the possible reviewers and prevents publishing corrections (new versions) of papers effectively. It appears to me that scientific publishing would need something similar to GitHub which is used for software. (Technically you could use GitHub for publishing papers in above sense but you would have even better results with a system optimized for this purpose.)
    1
  36. I remember writing stuff in Quake C after already knowing Pascal and some C. It was a bit weird but allowed to create bots to play in multiplayer game maps that I created for me and my brothers to play using two computers and null modem.
    1
  37. 1:03:00 There's nothing more permanent than a temporary solution that works. Alway assume that the program you're writing now is critical infrastructure of a bigger system in a decade.
    1
  38. I think the major problem is current publishing methods. Journals publish only "new" and "interesting" research and researchers get evaluated by the amount of papers the can get published. Veritasium made a great video about the issue with the title "Is Most Published Research Wrong?" – it nicely explains why things cannot get better unless we can introduce better measurement of researcher quality.
    1
  39. 1:39:00 What I would really want to see is a static analyzer for AI systems. I don't know if such a thing can exist.
    1
  40. 1:22:39 "Roughly, I don't know, one and a half milliseconds or so... maybe 1.6 milliseconds of latency..." Elon might as well be a Finnish engineer with such accurate estimates for things he's talking about.
    1
  41. 2:19:15 The DeepMind AlphaZero is really mindboggling when it comes to board games. In 2017, AlphaZero was trained from scratch within 24 hours to defeat Stockfish, elmo and AlphaGo Zero. And AlphaGo Zero was already the variant that won AlphaGo 100–0 in hundred game tournament. And AlphaGo won the best human player 4–1 in 2016 (or about one year earlier!). It's like if you would be going to attend international grand master level tournaments for both Chess and Go at the afternoon and reading the rule books for both those games at the morning of the very same day. Human grand masters typically have minimum of 20 years of serious training for those games and that gets you nowhere close the level of these AI systems. Since then, in 2022 DeepMind unveiled AlphaTensor which is based on the same technology and it learned mathematics and was able to improve the best known solution to optimal 4x4 matrix multiplication algorithm – previous record was Strassen algorithm from year 1969. So in some limited cases, AI has already beaten humans in mathematics, too.
    1
  42. 58:00 If you could "push a button" to make you feel like you've just eaten the best meal of your life and you're loved by all persons that are meaningful for you, would you use it? Human brain seems to be really eager to set a new baseline so even when your needs were fully fullfilled this way, it would appear boring really fast. Would we have to press "make me hurt" button every now and then to calibrate our brains?
    1
  43. Fridman and Vsauce in the same video. The only bad part was that I only found it today. YouTube algorithm is clearly not effective enough even when it should have had enough clues years ago that this video would keep me on the site.
    1
  44. 8:21:30 Considering the mouse cursor movement that's already possible, if you could come up with somewhat accurete triggers for start and end (logicall mouse down and mouse up), you could already get pretty fast input with swype-like text input already. However, it seems that ability to accurately click is still not there.
    1
  45. I think you should look for more information about "AAlib". Wikipedia has an article about it.
    1
  46. I think this would be valuable shortcut syntax for many languages. It's a bit limited because it must always use hardcoded "and" operator between comparisions but nothing else matches with typical mathematical markup so I guess that's fine. Considering that nearly all curly braces languages have opted to use angle brackets for generics, I guess allowing both generics and this syntax would result in hard to understand parser in many cases.
    1
  47. 3:03:30 One interesting way for future UIs to interact with AI is one that I've seen on video only (I don't know if it was fake but the idea seems solid): allow the user to interrupt the thinking process of R1 and modify the text on the fly and then allow the R1 to continue it's thinking. That way you can guide it's thinking if you don't like some specific part of its chain of thought or you think it made a specific mistake (for example, if it wonders if thing X is true and concludes that "yes, it's true") and you know it made a mistake, you can fix the mistake and then let the R1 to continue from there. This is obviously possible because the AI models are deterministic in sense that they only depend on model, system prompt, user input, PRNG seed results that far. If you modify the "history of generated output" and let the AI to continue from there, it will happily do so and take your correction into account correctly.
    1
  48. 1:45:00 This investment nationwide into "low tech" technologies has been important in solar panels, too. Majority of the solar panels are being manufactured in China so is it really a wonder that there's also majority of the people who can then try tweaking the processes to maybe manufacture improved parts? When US (or Europe) outsources stuff to China, they not only outsource the physical manufacturing of the given part, they only outsource the whole subvendor chain in long run. And getting all that back to US or Europe later is next to impossible.
    1
  49. 2:11:00 I think representing the AI as a creature might be good for the UI but it should be clearly explained to the users that its metaphorical truth: it's the repsentation that's easy to understand and works for practical purposes but it's factually false. Religions are a great example of people taking a metaphorical truth and mistaking it with facts and actual truths.
    1
  50. 8:30 As a software developer, the MIT licensing is the biggest thing about DeepSeek. All the other published weights, including Llama, are too restricted when it comes to licensing. If you cannot have totally free licensing, why would you spend your own resources to build something on top of the model? Do you really want to position yourself as a customer of a commercial in monopoly situation and still spend extra money to make yourself fully dependant on that position?
    1
  51. Failed training run in AI training is not that different from SpaceX Starship exploding before reaching orbit. This far failed AI training run has been cheaper but if we scale the models 10x more, the cost per try starts to be in the same range!
    1
  52. 54:48 "Money is a database for resource allocation" That's a great way to put it in theoretical sense. I've said for years that money is just a loan certificate and a promise to repay the loan in the future. It doesn't matter who took the loan first but the money is promise that the loan will be paid back to the owner of the loan certificate. Once you understand this, it's much more easy to understand how "banks can create money" – that's obvious because banks give new loans so it should be obvious that banks can create new loan certificates. (The history for cash was that Alice needed something from Bob and as Alice didn't have anything that Bob wanted as a return right away, the solution was to create a certificate where Alice promised to pay back the effort that Bob had done. This was basically a contract written on paper with a clause that whoever holds the contract is the target of the pay back. As such, the contract paper was valuable as itself and you could offer such a contract as a payment for whatever work was done for you. Fiat money is just a standardized format of such contracts. And Musk saying that it's just a database of resource allocation is obviously true, too. The more money you have, the more people are in dept towards you and you can get them to do work for you; the more money you have, the more potential future work will be allocated for you.)
    1
  53. What we really need is a distilled model of R1 where the distilled model is not based on Llama or Qwen. And that distilled model should be around 10B-16B parameters to be usable with real world gaming GPUs. Hopefully it would be MIT licensed, too.
    1
  54. Modern mobile apps are really wasteful hardware-wise because the hardware is paid by the end user. As such, low performance code will not hurt the developer or the company as long as the app seems interesting enough for the end user. It turns out that end users are willing to get over $1000 mobile devices just for entertainment. If that's your expected user base, you don't need to care if a web page requires one or two gigabytes of RAM to run. 20 years ago computers used to have 32 MB RAM and web sites worked just fine.
    1
  55. Getting money out of politics would be nice but do you have any idea how to actually do that in practice? Remember that money is nothing more but standardized form of loan certificates.
    1
  56.  @Jcewazhere  As long as billionaires have legally acquired their wealth on their own merits, I think it would be stupid not to hear what they suggest because they are obviously above average when it comes to handling money. Of course, you should always keep in mind that their suggestion might be simply about acquiring even more personal benefit so do not blindly implement their suggestions or recommendations. But getting them out of politics without hearing their ideas would be just dumb.
    1
  57. I think the most important part starts at 43:08.
    1
  58. Great interview! The only question I would have loved to see would have been as follows: If both Rust and C++ existed in their current state with no existing software written either language, why pick C++ over Rust today? I understand that when you have hundreds of millions of lines of existing C++ code, comparing just the language is not the only consideration you should do. However, we should be asking, which language is the best to teach to the next generation and the generation after that? For me personally, even though I know C++ better, Rust seems like a better language in long run thanks to its memory safety and especially its data race free promise. Multithreaded programming is so hard when you mix in shared memory and allocating and freeing resources in multiple threads that it's rare that people can get that correct without a lot of support from the compiler. And Rust seems to be the only language that even tries to fully do this. And I'm mostly interested in languages that have good enough performance. That basically rules out all garbage collection languages such as Java and C#. You only need to check the implementation of those languages to come to that conclusion: both JVM and CLR are written in C++. If Java and C# were actually generic languages, surely their own runtime systems would have been written in those languages, right? In reality, Java and C# performance is poor enough require writing the runtime in C++ (or C or Rust, but C++ was selected for historical or practical reasons).
    1
  59. 59:30 I think the new export rules are just stupid. If you want to limit the performance of AI supercomputers, you have to limit the interconnect technology, not the computing power per card. This is because you can always run more GPUs in parallel to get the same amount of GFLOPS to the restricted GPU. However, if you limit the interconnects, you can only effectively run the GFLOPS that a single GPU can do because you cannot effectively combine the GPUs to create big and fast clusters. If you allow fast interconnect technology, the workaround is "just buy many small parts and use interconnects to put those together to create one big system". That is, unless the plan was to just make it more expensive to China to use Nvidia GPUs. When you allow high performance interconnect but reduce max GFLOPS to 4x slower than could be implemented otherwise, you can instantly sell 4x more GPUs for the same use cases!
    1
  60. 42:40 Imagine the amount of permanently locked supposedly inherited property if we actually had smart contracts. Normal people lose their passwords all the time. How do you think they could handle cryptographic secrets where losing the secret causes losing all the money they have in bank, ownership of the house or a car?
    1
  61. Thanks for sharing this and I guess it make a lot of sense for you career-wise but it was dissappointing to find out that you had zero hard questions. If all the questions were pre-screened and you couldn't ask anything more, it should have been disclosed at the start of the video.
    1
  62. 5:20:00 So basically you have 1024 channel oscilloscope with 20 kHz sample rate per channel and you have to have low enough power usage to be able to run this for the whole day using a tiny battery inside the implant.
    1
  63.  @f7y8u9ugf  I think I'm (still) better communicating with non-developers to figure out what's actually needed. I've communicated enough with non-developers to know that "normal people" cannot formulate their needs or ideas clearly and figuring that out and filling in blanks from the context (real world) still seems too hard to do for an AI. I do believe that software architects will be replaced by AI in the long run but I would assume that's nearly the last occupations that AI can successfully do. Human psychology is a field in general that will be hard to tackle with the current AI designs. However, the rules for psychology are less strict than for writing software so "good enough" may be easier to achieve there. There are still too many unknowns to say anything for sure, though.
    1
  64. Great discussion. I feel that every discussion should do at least some steelmanning to stay real.
    1