Comments by "Mikko Rantalainen" (@MikkoRantalainen) on "Wes Roth" channel.

  1. The most important role for an AI would be to look through all the offered jobs on Upwork and weed out all the bad jobs (too low pay, unrealistic expectations, illegal tasks etc etc). In addition, Upwork requires you to pay for applying a job which makes it pretty risky to apply any job. The payment is not that much but higher than zero.
    5
  2. 7:55 The naming things as spelling mistakes is a result of trademark law: you cannot trademark words such as "partner" but you can trademark "partnr". An in case of US English, the "e" near the end of world is practically meaningless anyway so it doesn't matter for people.
    4
  3. Lawyers will ask GPT-4 and Q* for an opinion (or multiple) and then examine the AI outputs manually. Changes are pretty high that AI can generate interesting arguments but taking those to court without manually verifying every argument would be insane.
    2
  4. If that were true, why on earth are you watching this video?
    2
  5. Microsoft is talking a lot about "we're doing" when they refer to work done by OpenAI. The do have a big share of OpenAI but OpenAI was still a separate company I last checked.
    2
  6. It's the official "extra information to blind people" substream. Great if you are blind, distracting if you're not familiar with such additions.
    2
  7. There will be millions more AI jobs but the minimum skills required for those jobs are way above the skills of the unemployed people.
    2
  8. Indeed, the use of AI is already limited by the costs (e.g. GPT-4 Turbo with vision), not the capabilities of the AI. If the cost will go down enough, the usage will skyrocket.
    2
  9. Switching between languages does make sense for thinking AI system, too. My mother language is Finnish but I mostly read, think and write in English with it comes to programming. When I'm thinking about some complex algorithm in software, I'm thinking in English and when I'm thinking about my groceries, I'm thinking in Finnish. I don't see a reason why fully self-learned (RL) AI wouldn't do the same. The obvious difference being that the AI knows way more languages than I do. I have found it interesting to see how well different AI systems know Finnish. In my experience, none of the publicly available distilled models know Finnish before ~70B level for weights. I don't know if it's the result of distillation process or if Finnish is hard enough language to require that amount of weights to emit mostly flawless output. 7B and 14B models can understand Finnish to some degree but cannot emit a single sentence in Finnish without some mistakes.
    1
  10.  @GaryMillyz  5:24 gets you close. The difference for the V sign is if you show it palm side towards to viewer or towards yourself.
    1
  11. The big question is, if small number of people could do everything by increasing their throughput with AI, how is the rest of society going to pay for that effort? If the rest of the society is unemployed, they cannot provide valuable services to those few people that would produce all the daily used stuff. So why would those few keep working if there's no services to spend money when everybody else is unemployed?
    1
  12. 9:00 I think the physical animations are really good way to interact with humans but you should design the animations in a way where you don't slow down the user. For example, when I ask about the weather, animation to look towards the window while giving the response would be good UI. Doing full animation both ways before bothering to answer would be a bad UI. Looking at how iPhone works, I wouldn't expect Apple to go for fast response, though. They seem to be more interested in fluid animations first and functionality second.
    1
  13. 13:00 I think the "Time" column was actually total processing time of the script you executed (wallclock).
    1
  14. My take is that AI will slash the salaries in short run. AGI level AI will be expensive to run and if you're willing to work for less than the AGI would cost, you get to keep your job. Forget about ever getting a raise, though, because AI is only going to get cheaper. As I see it, the best strategy is going to be to learn effectively use AI in your work. AI is not going to take all the jobs, it will just make a small number of humans highly effective in their work and there will not be enough for for the same amount of people doing the same thing they used to do. If your AI skills are above average, you are not the first one to go out once AI improves throughput. And in long term, you don't need to be rich, you just need to be above average to have a nice life.
    1
  15. I think the biggest thing against intelligence explosion by AI is that the smarter we make the models, the slower they get for the throughput. For example, AlphaGeometry and AlphaProof did indeed get close to best result in the competition but only after using more time than the humans were allowed to use and this part seems be lost by many. So even if you need nearly supercomputer class system to run the greatest models, the throughput is still less than a skilled human. And I wrote "still" because it will get faster but I've yet to see signals about it getting exponentially faster which would be required for intelligence explosion.
    1
  16. I tried to ask it to write basically the same code that took me half a day last Tuesday and it got nearly perfect code after thinking for 20 seconds! The emitted code had a couple of sequences where I wasn't happy with the results but it wouldn't have taken long to fix those parts. The results were already close enough to complete solution that it would have been easier for me to fix the output by myself instead of trying to prompt the AI to fix the issues. The Gemini generated code had totally different writing style but the result was still readable enough to tell without running the code if the emitted code was good or not. Of course, I only could write perfect prompt because I had already solved the original problem already but this is definitely much much smarter than Copilot or similar products where I cannot get the AI to emit good response even if I prompt it very accurately.
    1
  17.  @WhiteArtsMagic  Market over reacting to news like this is a great way to make money if you're brave enough. Lots of stock was on sale.
    1
  18. It's easy to see that the AI companies already have enough data. Just think about how much data humans can take during their whole lifetime and humans can still show pretty intelligent behavior. As a result, the AI can definitely get much much smarter simply by thinking the existing data more. The "old" LLMs basically played the game "read all the books in the world and try to figure out gut feeling which word is the best for your next sentence without thinking anything, given that you have already said 'The solution'". The DeepSeek changes this to "think about about you were about to say with gut feeling and consider if it might have some errors when you re-read the whole output before returning your full response" which I still consider more like elementary school level thinking for humans. Once the AI thinking gets to level "rewrite your answer 10 times, each time improving it as much as possible and look for references linking to published research for every claim you make before emitting the output" it will beat all humans already. And if it runs on supercomputer with enough to storage to contain all the published content in the world, it will be able to check every reference, too. However, all that is going to be pretty slow on modern hardware. That would result in highly intelligent answers but poor throughput. Maybe it will be used to train the next model that can give similar responses by gut feeling?
    1
  19. 6:10 Whoever did this infographics wrote Helen Toner's last name without a capital letter. Accident or intentional signaling?
    1
  20. Running full DeepSeek R1 model on your hardware is definitely not cheap unless you're willing to run it really slowly. Remember that the R1 is 671B model which means about 700 GB of VRAM to run even the Q8 quantized model. Double that for the original open sourced FP16 model. You can run the model on any Linux computer if you have a couple of terabytes disk space by simply enabling enough swap you then you can expect AI output speed around 2 words per minute.
    1
  21.  @iupring  QwQ is indeed interesting but I just tested it with tasks that require understanding both Finnish and English and it was definitely worse than Llama 70b distilled version of R1. And even Llama 70b model requires pretty powerful machine to run fast.
    1
  22. I think the only thing you need for RL would be to ask the AI to find contradictions in the internet and then try to figure out which claim is the correct one. Every complex human generated work contains more or less errors which often go unnoticed by humans because humans are either too stupid to notice the problem or the error is so obvious that smarter inviduals simply ignore it as meaningless clerical error. And given the amount of all the publicly assessible data on internet, you could spend your whole life combining fragments of it in different pairs and checking for contradictions. The more you compare the fragments, the better you can evaluate the trustworthiness of any given fragment and then you can train on the results and recheck the whole thing again using your improved understanding of the reality as a whole. The key is to understand that there's always unknown amount of noise in any data and the only way to figure it out is to compare it to all the other data you have. Modern AI systems with billions of weights already have much much more data than any human being ever. So they don't lack data, only thinking about it.
    1
  23. 28:30 By definition, a median skilled adult human has IQ 100. I wouldn't trust my taxes or other official documents to such AI as long as I'm responsible for those documents. Would I let the AI to help me by prefilling all forms that I have to fill? Sure. Even "Expert" level AI being at least 90th percentile of skilled adults would probably have IQ close to 115 which is far from genius level. Even "Virtuoso" level wouldn't be enough to get Mensa membership but pretty close to it. Unless I'm mistaken, typical office worker has median IQ around 105–110.
    1
  24. 4:43 Modern nuclear bombs are fusion bombs. Commonly called thermonuclear bombs.
    1
  25. During the next 10–20 years, I would guess the whole world must switch to UBI (Universal Basic Income) or we'll see a lot of miserable life. I'm afraid that with current political systems, we'll end up with a lot of people with miserable life. I'd love to be proven wrong.
    1