6 ms·

I don’t understand how you u can’t worry about this. Like there’s a trend line of progress right? Ok so the thing isn’t effective now. But there’s a decade of

I don’t understand how you u can’t worry about this. Like there’s a trend line of progress right? Ok so the thing isn’t effective now. But there’s a decade of upward progress and that projection line point to a future where a better AI exists. Trend-lines don’t point to an exact future just a most probable future. It is unwise to discount the most probable future.

miraks 1y ago
If you gave me a model released two years ago and today and let me do some programming with both, I would have no problems telling you which one was released two years ago; progress on this front is very noticeable. But if you let me chat with each one for an hour, I'm honestly not sure I would be able to tell the difference.
- ninetyninenine 1y ago
  So 2 years of progress is enough to form a trend-line? Do you remember life during the time when AI at this level didn’t exist? Follow the 10 year trend-line. That’s the thing that points to the future. But either way there’s progress on both fronts. Talking to it has improved we just can’t measure it quantitatively imo.
  - ozgrakkurt 1y ago
    Can’t draw any line without being aware of underlying technicality of AI, how it works? What previous research enabled LLMs? etc.
    - ninetyninenine 1y ago
      Research into machine learning, transformers, there's tons of technicalities that enables LLMs. ANd there's more and more "technicalities" being thrown on top of LLMs and we're having modifications to LLMs as well. LLMs are the tip of a spear of a trendline that didn't involve LLMs. Prior to that we had AI generating art and music through diffusion algorithms. We had AI doing image recognition and doing mind reading. The trendline is clear to anyone but those who think the current state of LLMs and the problems we have with it are completely static in nature.
- kenjackson 1y ago
  You could definitely tell the difference. The persistence and context windows make a world of difference in just casual usage.
  - svachalek 1y ago
    Also the massive inflation in sycophancy.
wizzwizz4 1y ago
But the current systems are about as good as they'll get: we can make them a bit better in fields where we can cheaply generate synthetic data, but human communication is not one of them. (And even where you can generate synthetic data, your efforts would usually be better spent assembling a purpose-built system.) Sure, I probably could make something more effective, using non-LLM technologies (given a large enough budget), but… why would I, or anyone else, do that when it'd be obviously harmful, with no benefit?
- ninetyninenine 1y ago
  The current system we have are as good as they get? You’re just making this statement out of thin air? Did you not notice a trendline of technological improvement of AI? > Sure, I probably could make something more effective, using non-LLM technologies (given a large enough budget), but… why would I, or anyone else, do that when it'd be obviously harmful, with no benefit? Technology will improve. The likelihood of you being part of that progress is nearly zero. So what you say here is categorically wrong. You are not able to make anything better. Humanity collectively will make something better and we don’t know who will be the one to do it. People are willing to pay for companionship so there’s huge profitability in this area. Profit and self interest often at the expense of everything else is what drives progress.
  - techpineapple 1y ago
    > Did you not notice a trendline of technological improvement of AI? No, in fact I noticed a series of AI winters. In all things, progress is famously _not_ a straight line. Also I find it interesting that your argument seems to boil down to “I’m smart because line goes up, you’re dumb because you think line goes down.” Everyone Clearly can see what would happen if line went up, I just; looking at the broad history and totality of factors(that I’m aware of) don’t think it’s inevitable. “You can’t stop progress” We literally stop progress all the time, every time we choose not to invest in something, crypto progress slowed from its height, Vr progress, green energy, I’d argue it’s relatively few technologies that progress forever.
    - 1y ago
techpineapple 1y ago
Progress is famously not a straight line.
- ninetyninenine 1y ago
  Of course not. And neither is machine learning data. I'm talking about the trendline. If you deploy that machine learning knowledge and draw a best fit line. That line has a slope that is upward.
  - ryandrake 1y ago
    There were only 66 years between the Wright brothers' first successful powered flight in 1903 and the Apollo 11 moon landing in 1969. The slope of the following 66 years of aviation/flight history has been distinctly less steep. If you were in 1969 and extrapolated the trendline, we would have colonized the inner solar system by now.
    - ninetyninenine 1y ago
      Right so a single example proves your point and destroys the entire concept of trendlines. We should stop the use of trendlines immediately. Why don’t you take a more realistic look at it. Trendlines predict things. They don’t guarantee things. You want to make a prediction based off of bullshit? Or if you want to make a prediction based off of a trendline. I pick the later. You pick the former because of one example.
      - ryandrake 1y ago
        There are many, many real world examples where past performance is not indicative of future results.
        ninetyninenine 1y ago
        Of course there are. Because nothing goes up the line forever. Eventually any real world line actually ends. Nothing in the real world actually obeys the equation of a line to infinite. This happens to every fucking line in existence. So why do trend-lines even exist? Why don’t we flex some of those common sense muscles. If we know that locally we are on a line, and we know that the trend obeyed that of a line or upward trend -x units into the past. The most probable bet is to say that the line will obey the trend +x units into the future. The more data and the less noise we have on the past trend makes predicted future trend more likely. This is The data driven logic, no way around this other then using your gut to say random contrary shit. My claim is given all the information we have the best most probable bet is AI will follow the trend it followed in the past. And your bet is that it won’t and you didn’t even supply any qualitative evidence for why not. You just cited common trope quotes from investing without nuance. You realize that value investing on the index sp500 is a trendline based bet?

miraks 1y ago

If you gave me a model released two years ago and today and let me do some programming with both, I would have no problems telling you which one was released two years ago; progress on this front is very noticeable. But if you let me chat with each one for an hour, I'm honestly not sure I would be able to tell the difference.

ninetyninenine 1y ago
So 2 years of progress is enough to form a trend-line? Do you remember life during the time when AI at this level didn’t exist? Follow the 10 year trend-line. That’s the thing that points to the future. But either way there’s progress on both fronts. Talking to it has improved we just can’t measure it quantitatively imo.
- ozgrakkurt 1y ago
  Can’t draw any line without being aware of underlying technicality of AI, how it works? What previous research enabled LLMs? etc.
  - ninetyninenine 1y ago
    Research into machine learning, transformers, there's tons of technicalities that enables LLMs. ANd there's more and more "technicalities" being thrown on top of LLMs and we're having modifications to LLMs as well. LLMs are the tip of a spear of a trendline that didn't involve LLMs. Prior to that we had AI generating art and music through diffusion algorithms. We had AI doing image recognition and doing mind reading. The trendline is clear to anyone but those who think the current state of LLMs and the problems we have with it are completely static in nature.
kenjackson 1y ago
You could definitely tell the difference. The persistence and context windows make a world of difference in just casual usage.
- svachalek 1y ago
  Also the massive inflation in sycophancy.

wizzwizz4 1y ago

But the current systems are about as good as they'll get: we can make them a bit better in fields where we can cheaply generate synthetic data, but human communication is not one of them. (And even where you can generate synthetic data, your efforts would usually be better spent assembling a purpose-built system.) Sure, I probably could make something more effective, using non-LLM technologies (given a large enough budget), but… why would I, or anyone else, do that when it'd be obviously harmful, with no benefit?

ninetyninenine 1y ago
The current system we have are as good as they get? You’re just making this statement out of thin air? Did you not notice a trendline of technological improvement of AI? > Sure, I probably could make something more effective, using non-LLM technologies (given a large enough budget), but… why would I, or anyone else, do that when it'd be obviously harmful, with no benefit? Technology will improve. The likelihood of you being part of that progress is nearly zero. So what you say here is categorically wrong. You are not able to make anything better. Humanity collectively will make something better and we don’t know who will be the one to do it. People are willing to pay for companionship so there’s huge profitability in this area. Profit and self interest often at the expense of everything else is what drives progress.
- techpineapple 1y ago
  > Did you not notice a trendline of technological improvement of AI? No, in fact I noticed a series of AI winters. In all things, progress is famously _not_ a straight line. Also I find it interesting that your argument seems to boil down to “I’m smart because line goes up, you’re dumb because you think line goes down.” Everyone Clearly can see what would happen if line went up, I just; looking at the broad history and totality of factors(that I’m aware of) don’t think it’s inevitable. “You can’t stop progress” We literally stop progress all the time, every time we choose not to invest in something, crypto progress slowed from its height, Vr progress, green energy, I’d argue it’s relatively few technologies that progress forever.
  - 1y ago

techpineapple 1y ago

Progress is famously not a straight line.

ninetyninenine 1y ago
Of course not. And neither is machine learning data. I'm talking about the trendline. If you deploy that machine learning knowledge and draw a best fit line. That line has a slope that is upward.
- ryandrake 1y ago
  There were only 66 years between the Wright brothers' first successful powered flight in 1903 and the Apollo 11 moon landing in 1969. The slope of the following 66 years of aviation/flight history has been distinctly less steep. If you were in 1969 and extrapolated the trendline, we would have colonized the inner solar system by now.
  - ninetyninenine 1y ago
    Right so a single example proves your point and destroys the entire concept of trendlines. We should stop the use of trendlines immediately. Why don’t you take a more realistic look at it. Trendlines predict things. They don’t guarantee things. You want to make a prediction based off of bullshit? Or if you want to make a prediction based off of a trendline. I pick the later. You pick the former because of one example.
    - ryandrake 1y ago
      There are many, many real world examples where past performance is not indicative of future results.
      - ninetyninenine 1y ago
        Of course there are. Because nothing goes up the line forever. Eventually any real world line actually ends. Nothing in the real world actually obeys the equation of a line to infinite. This happens to every fucking line in existence. So why do trend-lines even exist? Why don’t we flex some of those common sense muscles. If we know that locally we are on a line, and we know that the trend obeyed that of a line or upward trend -x units into the past. The most probable bet is to say that the line will obey the trend +x units into the future. The more data and the less noise we have on the past trend makes predicted future trend more likely. This is The data driven logic, no way around this other then using your gut to say random contrary shit. My claim is given all the information we have the best most probable bet is AI will follow the trend it followed in the past. And your bet is that it won’t and you didn’t even supply any qualitative evidence for why not. You just cited common trope quotes from investing without nuance. You realize that value investing on the index sp500 is a trendline based bet?