Comments by "Michael Nurse" (@michaelnurse9089) on "Internet of Bugs" channel.

  1. Branson went surfing with naked models. He was a CEO. His companies still operate.
    35
  2. They already did. How many layoffs in the last two years again?
    21
  3. Straw man alert. 4-o was never supposed to be better. It was always supposed to a trimmed down version that is faster and cheaper to use but roughly the same accuracy. 5 will be better, but much slower again, obviously.
    19
  4. Nasdaq 7.5% down after being 115% up over the last 5 years. I don't think you know what reset means, or maybe you read too many newspapers.
    8
  5. Yes. We get it. But so do your competition. When your competition switch to LLM boosted seniors with a 30-50% reduction in billings to clients suddenly you will be under pressure. Can all those juniors' collective growth outrace the productivity growth from LLMs? If so, you continue as before. If not, its time to panic. Which way it will go is not clear at this stage, but LLMs have a lot of muscle.
    7
  6. Coders are the only profession that earns high salaries after 3-4 years training. To get the same in medicine requires 10-15 years. The risk is that coder salaries return to normal.
    5
  7. New version allows for new opinion. I commented previously this is how it would play out. Some people just ignore the incoming future.
    4
  8. "As for AI .... I did a 3 month shroom microdosing regimen 😂. I don't care about AI taking my job anymore. " How many kids do you have?
    4
  9. Loss in the middle problem is a temporary issue. Some LLMs are already calling it solved.
    4
  10. If a rewrite is needed 3 years later then the AI was a success.
    3
  11. Corporations were never going to be the ones to use it like this - too risky. Take a smaller business with less to lose and a bigger focus on cost and you will find that code is being written with AI.
    3
  12. I worked in an industry - insurance - where the average age of a broker was 58. It creates terrible stagnation and inability to innovate.
    3
  13. Often just properly prompting or regenerating the response gets you to the right answer.
    3
  14. Most are getting this correct now. My 286 used to randomly crash in certain applications after I pressed the turbo button. Yes, computers used to come with a turbo button.
    3
  15. Memory: the AI can be trained to write and read to a db. Still would be very far from human level memory but it is possible to improve this area.
    2
  16. It just needs to be 'human' within the domain. For example, translation - editing: here is it good enough to be considered human. Programming may get its day in a couple years or maybe not - progress always tapers off, it is just a question of when.
    2
  17. Yet 50% of tokens used on LLMS are for coding. Further, surveys show heavy use of LLMs by software engineers. Watch this video "She makes 6 figures with AI" - ignore the clickbait title.
    2
  18. As an AI person I find this to be somewhat of a whitewashing attempt - two examples only: 1. Tesla may have faked a test in 2016 but it is 9 years later now and FSD is looking better and better each month. 2. Amazon Go was not faked - the India team existed to do labeling - a crucial part of every AI engineering effort - Amazon stopped it because it was going to be a hard problem - like driving a car with AI - and take 10 years to nail down with big investment and a long break even period.
    2
  19. An AI can never have a good working relationship with a person.
    1
  20. You can teach them simple things not in their training data - just explain how to do it in the prompt. If this can be scaled up 1^n times then guy who wrote book will be very wrong.
    1
  21. Or you could tell it to use Rust and forego all exploits.
    1
  22. It depends what sort of programming you do. Some programming work has very little communication - for example, I once paid an Indian company to code up a full stack website from a frozen design. They sat there coding for 6 months and almost never contacted me - everything was in the design document. There were only a small number of bugs at the end.
    1
  23. Strictly speaking: the algorithms have no mathematical limit to what they can learn. There is only the question of insufficient good data and enough compute. And there will be enough compute, I can assure you of that. Because programmers don't write down WHY they are writing or editing code(it would take a long time), and prefer to keep it in their head, it is unlikely there is enough data to make a good programmer AI. Of course, that all changes when an LLM company hires or contracts 10 000 good programmers to start giving them a full account of WHY they are doing the things they do. Then they will have the dataset. It will cost 10s or 100s of billions.
    1
  24. AI maths is an issue because of tokenization, which is a perf shortcut. This problem should mostly go away with the next generation of models. Or failing that the ones thereafter.
    1
  25. An additional point: Web has had many faked features as well, people writing fake reviews on Amazon or whatever - but the underlying technology is solid, so it endures.
    1
  26. If you ask it a common task or language it is too easy for it.
    1
  27. Surveys show they are speeding up devs by 30%, o1 will be 40% and GPT4 will be 50%. This means 50% unemployment is not impossible. Or a drop in salaries.
    1
  28. As an AI guy and a programmer I don't think we are done with big advances in the short term. Once OpenAI or whomever get around to building a LLM specifically to work alongside senior programmers - large code bases, negotiation with users, testing and so on, then there will be another leap forward. The thing is is that these startups will be in stealth mode, they will have capital and a couple hundred programmers on staff and will be looking to really move the envelope. Don't make the mistake of taking the skill of ChatGpt1o and extrapolating forward. Look fundamentally at what AI does, pattern recognition learned from data, and you will see it can carve even bigger chunks out of programming. It will never completely eliminate programmers but a firm with 20 seniors and 40 juniors could be reduced to 10 seniors and 4 juniors, leaving 70% of the industry unemployed. This happened to Data Scientists which I was training towards, one day we needed a hundred data scientists to hand craft algorithms for data business like Netflix, a few months later we needed zero, because neural nets clean the data themselves and there is only one algorithm.
    1