3 ms·

> I'm literally shocked how we can spend a couple decades fantasizing and writing stories about this level of AI It was never this level of AI. The stories we

by latexr 2y ago

> I'm literally shocked how we can spend a couple decades fantasizing and writing stories about this level of AI It was never this level of AI. The stories we wrote and fantasised were about AI you could blindly rely on, trust, and reason about. No one ever fantasised about AI which couldn’t accurately count the number of letters in a common word or that would give you provably wrong information in an assertive authoritative tone. No one longed for a level of AI where you have to double check everything.

mjr00 2y ago
> No one longed for a level of AI where you have to double check everything. This has basically been why it's a non-starter in a lot of (most?) business applications. If your dishwasher failed to clean anything 20% of the time, would you rely on it? No, you'd just wash the dishes by hand, because you'd at least have a consistent result. That's been the result of AI experimentation I've seen: it works ~80% of the time, which sounds great... except there's surprisingly few tasks where a 20% fail rate is acceptable. Even "prompt engineering" your way to a 5% failure/inaccuracy rate is unacceptable for a fully automated solution. So now we're moving to workflows where AI generates stuff and a human double checks. Or the AI parses human text into a well-defined gRPC method with known behavior. Which can definitely be helpful, but is a far cry from the fantasized AI in sci-fi literature.
- Nition 2y ago
  One type of question that a 20%-failure-rate AI can still be very useful for is ones that are hard to answer but easy to verify. For example say you have a complex medical problem. It can be difficult to do a direct Internet search that covers the history and symptoms. If you ask AI though, it'll be able to give you some ideas for specific things to search. They might be wrong answers, but now you can easily search specific conditions and check them. Sort of P vs. NP for questions.
  - skydhash 2y ago
    > For example say you have a complex medical problem. Or you go to a doctor instead of imagining answers.
    - Nition 2y ago
      Medical was just one example, replace with anything you like. As another example, you can give the AI a photo of something to have it name what that thing is. Then you can check the thing by its name on Google to see if it matches. Much easier than describing the thing (plant, tool, etc) to Google.
      - skydhash 2y ago
        Having the wrong information can be more detrimental than having no information at all. In the former case, confident actions will be take. In the latter case, the person will be tentative wich can reduce the area of effect of bad decisions. Imagine the lambda person confronted with this: sudo rm -rf / What is the better situation, having no understanding of what it does or believing that another action will take place?
        Nition 2y ago
        The process I'm suggesting is: 1. You have a complex or vague question that you can't search easily via Google etc 2. You ask the AI and it converts that to concrete searchable suggestions (in this case "sudo rm -rf /") 3. You search "sudo rm -rf /" to check the answer. Step 3 is designed to (hopefully) catch this kind of problem.
    - ianbutler 2y ago
      You put too much faith in doctors. Pretty much every woman I know has been waived off for issues that turned serious later and even as a guy I have to do above average leg work to get them to care about anything.
      - satisfice 2y ago
        Doctors are still better than LLMs, by a lot.
- dclowd9901 2y ago
  It feels a bit like LLMs rely a lot on _us_ to be useful. Which is a big point to the author's article about how companies are trimming off staff for AI.
  - re-thc 2y ago
    > how companies are trimming off staff for AI But they're not. That's just the excuse. The real truth is somewhere along pandemic over hire and bad / unstable economy.
    - Terr_ 2y ago
      Also attempts to influence investors/stock-price. https://newrepublic.com/article/178812/big-tech-loves-lay-offs https://newrepublic.com/article/178812/big-tech-loves-lay-of...
    - dclowd9901 2y ago
      We've frozen hiring (despite already being under staffed) and our leadership has largely pointed to advances in AI as being accelerative to the point that we shouldn't need more bodies to be more productive. Granted it's just a personal anecdote but it still affects hundreds of people that otherwise would have been hired by us. What reason would they have to lie about that to us?
simonw 2y ago
"The stories we wrote and fantasised were about AI you could blindly rely on, trust, and reason about." Stanley Kubrick's 2001: A Space Odyssey - some of the earliest mainstream AI science fiction (1968, before even the Apollo moon landing!) was very much about an AI you couldn't trust.
- dingnuts 2y ago
  that's a different kind of distrust, though, that was an AI that was capable of malice. In that case, "trust" had to do with loyalty. The GP means "trust" in the sense of consistency. I trust that my steering wheel doesn't fly off, because it is well-made. I trust that you won't drive into traffic while I'm in the passenger seat, because I don't think you will be malicious towards us. These are not the same.
  - the_af 2y ago
    Going on a tangent here: not sure 2001's HAL was a case of outright malice. It was probably a malfunction (he incorrectly predict a failure) and then conflicting mission parameters that placed higher value on the mission than the crew (the crew discussed shutting down HAL because it seemed unreliable, and he reasoned it would jeopardize the mission and the right course of action was killing the crew). HAL was capable of deceit in order to ensure his own survival, that much is true. In the followup 2010, when HAL's mission parameters are clarified and de-conflicted, he doesn't attempt to harm the crew anymore. I... actually can see the 2001's scenario happening with ChatGPT if it was connected to ship peripherals and told mission > crew and that this principle overrides all else. In modern terms it was about both unreliability (hallucinations?) and a badly specified prompt!
    - Lerc 2y ago
      I don't think there was any malfunction. The conflicting parameters implicitly contained permission to lie to the crew. The directive to take the crew to Saturpiter but also to not let them learn anything of the new mission directive meant deceiving. It's possible HAL's initial solution was to impose a communication blackout by simulating failures, then the crew reactions to the deception necciatsted their deaths to preserve the primary mission. Less a poor prompt and more two incompatible prompts both labeled top priority. Any conclusion can be logicLally derived from a contradiction. Total loyalty cannot serve two masters. Clarke felt quite guilty about the degree of distrust of computers that HAL generated.
ninetyninenine 2y ago
>It was never this level of AI. You're completely out of it. We couldn't even get AI to hold a freaking conversation. It was so bad we came up with this thing called the turing test and that was the benchmark. Now people like you are all, like "well it's obvious the turing test was garbage". No. It's not obvious. It's the hype got to your head. If we found a way to travel at light speed for 3 dollars the hype would be insane and in about a year we get people like you writing blog posts about how light speed travel is the dumbest thing ever. Oh man too much hype. You think LLMs are stupid? Sometimes we all just need to look in the mirror and realize that humans have their own brand of stupidity.
- ryandrake 2y ago
  LLMs are glorified, overhyped autocomplete systems that fail, but in different, nondeterministic ways than existing autocomplete systems fail. They are neat, but unreliable toys, not “more profound than fire or electricity” as has been breathless claimed.
  - cheevly 2y ago
    You just literally described humans; and the meta lack of awareness reinforces itself. You cyclicly devalue your own point.
    - nativeit 2y ago
      Not for nothing, humans also enjoy the worth and dignity inherent with being alive and intelligent…not to mention significantly less error prone (see: hallucination rates in literally any current model), while being exponentially more efficient to produce and run. I can make that last assertion pretty confidently, because while I’ve never built a data center so resource intensive it required its own dedicated power generation plant, I have put in the work to produce dozens of new people (those projects all failed, but only because we stubbornly refuse to take the wrappers off the tooling), and the resource requirements there only involved some cocktails and maybe a bag of Doritos. Anyhow, I reckon humans are still, on-balance, the better non-deterministically imperfect vessels of logic, creation, and purpose.
HDThoreaun 2y ago
> It was never this level of AI. People have been dreaming of an AI that can pass the turing test for close to a century. We have accomplished that. I get moving the goalposts since the turing test leaves a lot to be desired, but pretending you didnt is crazy. We have absolutely accomplished the stuff of dreams with AI

by latexr 2y ago

mjr00 2y ago

> No one longed for a level of AI where you have to double check everything. This has basically been why it's a non-starter in a lot of (most?) business applications. If your dishwasher failed to clean anything 20% of the time, would you rely on it? No, you'd just wash the dishes by hand, because you'd at least have a consistent result. That's been the result of AI experimentation I've seen: it works ~80% of the time, which sounds great... except there's surprisingly few tasks where a 20% fail rate is acceptable. Even "prompt engineering" your way to a 5% failure/inaccuracy rate is unacceptable for a fully automated solution. So now we're moving to workflows where AI generates stuff and a human double checks. Or the AI parses human text into a well-defined gRPC method with known behavior. Which can definitely be helpful, but is a far cry from the fantasized AI in sci-fi literature.

Nition 2y ago
One type of question that a 20%-failure-rate AI can still be very useful for is ones that are hard to answer but easy to verify. For example say you have a complex medical problem. It can be difficult to do a direct Internet search that covers the history and symptoms. If you ask AI though, it'll be able to give you some ideas for specific things to search. They might be wrong answers, but now you can easily search specific conditions and check them. Sort of P vs. NP for questions.
- skydhash 2y ago
  > For example say you have a complex medical problem. Or you go to a doctor instead of imagining answers.
  - Nition 2y ago
    Medical was just one example, replace with anything you like. As another example, you can give the AI a photo of something to have it name what that thing is. Then you can check the thing by its name on Google to see if it matches. Much easier than describing the thing (plant, tool, etc) to Google.
    - skydhash 2y ago
      Having the wrong information can be more detrimental than having no information at all. In the former case, confident actions will be take. In the latter case, the person will be tentative wich can reduce the area of effect of bad decisions. Imagine the lambda person confronted with this: sudo rm -rf / What is the better situation, having no understanding of what it does or believing that another action will take place?
      - Nition 2y ago
        The process I'm suggesting is: 1. You have a complex or vague question that you can't search easily via Google etc 2. You ask the AI and it converts that to concrete searchable suggestions (in this case "sudo rm -rf /") 3. You search "sudo rm -rf /" to check the answer. Step 3 is designed to (hopefully) catch this kind of problem.
  - ianbutler 2y ago
    You put too much faith in doctors. Pretty much every woman I know has been waived off for issues that turned serious later and even as a guy I have to do above average leg work to get them to care about anything.
    - satisfice 2y ago
      Doctors are still better than LLMs, by a lot.
dclowd9901 2y ago
It feels a bit like LLMs rely a lot on _us_ to be useful. Which is a big point to the author's article about how companies are trimming off staff for AI.
- re-thc 2y ago
  > how companies are trimming off staff for AI But they're not. That's just the excuse. The real truth is somewhere along pandemic over hire and bad / unstable economy.
  - Terr_ 2y ago
    Also attempts to influence investors/stock-price. https://newrepublic.com/article/178812/big-tech-loves-lay-offs https://newrepublic.com/article/178812/big-tech-loves-lay-of...
  - dclowd9901 2y ago
    We've frozen hiring (despite already being under staffed) and our leadership has largely pointed to advances in AI as being accelerative to the point that we shouldn't need more bodies to be more productive. Granted it's just a personal anecdote but it still affects hundreds of people that otherwise would have been hired by us. What reason would they have to lie about that to us?

simonw 2y ago

"The stories we wrote and fantasised were about AI you could blindly rely on, trust, and reason about." Stanley Kubrick's 2001: A Space Odyssey - some of the earliest mainstream AI science fiction (1968, before even the Apollo moon landing!) was very much about an AI you couldn't trust.

dingnuts 2y ago
that's a different kind of distrust, though, that was an AI that was capable of malice. In that case, "trust" had to do with loyalty. The GP means "trust" in the sense of consistency. I trust that my steering wheel doesn't fly off, because it is well-made. I trust that you won't drive into traffic while I'm in the passenger seat, because I don't think you will be malicious towards us. These are not the same.
- the_af 2y ago
  Going on a tangent here: not sure 2001's HAL was a case of outright malice. It was probably a malfunction (he incorrectly predict a failure) and then conflicting mission parameters that placed higher value on the mission than the crew (the crew discussed shutting down HAL because it seemed unreliable, and he reasoned it would jeopardize the mission and the right course of action was killing the crew). HAL was capable of deceit in order to ensure his own survival, that much is true. In the followup 2010, when HAL's mission parameters are clarified and de-conflicted, he doesn't attempt to harm the crew anymore. I... actually can see the 2001's scenario happening with ChatGPT if it was connected to ship peripherals and told mission > crew and that this principle overrides all else. In modern terms it was about both unreliability (hallucinations?) and a badly specified prompt!
  - Lerc 2y ago
    I don't think there was any malfunction. The conflicting parameters implicitly contained permission to lie to the crew. The directive to take the crew to Saturpiter but also to not let them learn anything of the new mission directive meant deceiving. It's possible HAL's initial solution was to impose a communication blackout by simulating failures, then the crew reactions to the deception necciatsted their deaths to preserve the primary mission. Less a poor prompt and more two incompatible prompts both labeled top priority. Any conclusion can be logicLally derived from a contradiction. Total loyalty cannot serve two masters. Clarke felt quite guilty about the degree of distrust of computers that HAL generated.

ninetyninenine 2y ago

>It was never this level of AI. You're completely out of it. We couldn't even get AI to hold a freaking conversation. It was so bad we came up with this thing called the turing test and that was the benchmark. Now people like you are all, like "well it's obvious the turing test was garbage". No. It's not obvious. It's the hype got to your head. If we found a way to travel at light speed for 3 dollars the hype would be insane and in about a year we get people like you writing blog posts about how light speed travel is the dumbest thing ever. Oh man too much hype. You think LLMs are stupid? Sometimes we all just need to look in the mirror and realize that humans have their own brand of stupidity.

ryandrake 2y ago
LLMs are glorified, overhyped autocomplete systems that fail, but in different, nondeterministic ways than existing autocomplete systems fail. They are neat, but unreliable toys, not “more profound than fire or electricity” as has been breathless claimed.
- cheevly 2y ago
  You just literally described humans; and the meta lack of awareness reinforces itself. You cyclicly devalue your own point.
  - nativeit 2y ago
    Not for nothing, humans also enjoy the worth and dignity inherent with being alive and intelligent…not to mention significantly less error prone (see: hallucination rates in literally any current model), while being exponentially more efficient to produce and run. I can make that last assertion pretty confidently, because while I’ve never built a data center so resource intensive it required its own dedicated power generation plant, I have put in the work to produce dozens of new people (those projects all failed, but only because we stubbornly refuse to take the wrappers off the tooling), and the resource requirements there only involved some cocktails and maybe a bag of Doritos. Anyhow, I reckon humans are still, on-balance, the better non-deterministically imperfect vessels of logic, creation, and purpose.

HDThoreaun 2y ago

> It was never this level of AI. People have been dreaming of an AI that can pass the turing test for close to a century. We have accomplished that. I get moving the goalposts since the turing test leaves a lot to be desired, but pretending you didnt is crazy. We have absolutely accomplished the stuff of dreams with AI