4 ms·

This is an amazing explanation, but I sense that it will become dated in some years, like, it will describe how computing worked in the old times, when people a

This is an amazing explanation, but I sense that it will become dated in some years, like, it will describe how computing worked in the old times, when people actually coded by hand. Because, > Flexibility and the ability to make inferences are skills that children learn as they grow up. But computers are like Peter Pan: they never grow up. LLMs will change all of this. Maybe in some decades we will see programming environments that don't require people to understand the solution in order to create a program. For example, it may be possible for people to create spellcheckers without understanding the rules of spelling. Unfortunately this will create a whole new class of bugs due to AIs hallucinating code, and I'm afraid that the average software quality will take a hit.

benj111 4y ago
The programmer has to understand the problem. If the person behind the keyboard offloads that to AI, the AI is the programmer. The person behind the keyboard is basically just a manager.
- scarface74 4y ago
  The gap between “understanding the problem” and “coming up with a solution” has become a lot narrower over the years. 1. To do anything interesting and at any speed on my Apple //e I had to know assembly. 2. Then I got a Mac in 1992 and the original Mac Toolbox routines could do a lot from a higher level language. 3. Then Visual Basic made graphical programs easier and later WinForms, C# etc. 4. Then AWS made provisioning infrastructure a matter of just writing a bunch of yaml 5. Now that LLMs exist and ChatGPT is well trained on AWS APIs in various languages, I can use natural language and it spits out code that works 95% of the time with only slight tweaks if any.
  - benj111 4y ago
    I'd state it more as, you don't have to reinvent to foundations to solve a new problem. If your problem is writing a link list, you still have to understand linked lists. If your problem just happens to be solved by linked lists, you can rely on a library. Re point 5. If debugging is twice as hard as writing code. How long before AI becomes half as intelligent as the average programmer making said programs undebuggable by the average programmer? Is it a good thing to have black boxes writing black boxes?
    - owl57 4y ago
      > How long before AI becomes half as intelligent as the average programmer making said programs undebuggable by the average programmer? It doesn't work this way. A programmer with a bit of experience knows to curb the desire for cleverness and write code they — and their teammates, probably having comparable intelligence and even less knowledge about specific code — can debug. Unless they are on a coffee trip and hacking on a cool weekend project, of course. Now that's another question, what in the hell will AI's weekend projects be.
      - benj111 4y ago
        >A programmer with a bit of experience knows to curb the desire for cleverness Are you suggesting AI will know to curb it's cleverness so that mere humans can understand it???
        owl57 4y ago
        Setting aside all that debate about future AI agency and focusing on currently known mode of operation, why not. I haven't tested, but GPT-4 may already know enough to act on "this part of my/your code is too clever, please rewrite it simpler" and "please add this feature to that program, and don't be too clever".
        scarface74 4y ago
        GPT3 could do that.
        benj111 4y ago
        You're suggesting that "too clever" is a 'known unknown'. If it's an unknown unknown how can you spot the too clever thing to request that it's simplified? Edit: further, that assumes an AIs definition of simple matches that of a human programmer. If there's anything we've learnt recently, AI doesn't think like humans.
  - marginalia_nu 4y ago
    Aren't the two always the same thing? I'll even go so far as to say you don't understand a problem if its solution does not appear childishly trivial.
    - scarface74 4y ago
      You don’t see the gap between “I want a natural language chatbot” and being able to create ChatGPT?
      - marginalia_nu 4y ago
        Right, the gap is a whole lot deeper understanding of the problem than "1 chatgpt plx", probably starting with understanding just what chatgpt is.
  - wruza 4y ago
    Each of these steps made computation increasingly more expensive cycle- and power-wise. We counteracted that with Moore’s for a while, then had to go “scalable”. It worked because it was still cheaper than hiring an employee. I bet that LLMs will create another level of bloat industry where computation will be as expensive as making a human to do the same thing, because there will be no human-can-do-it constraint anymore. This promise of “AI will create cool software” is akin to “robots will fly in cars in vertical cities” of 19/20th century scifi. All we got is worse traffic and exponentially expensive housing.
    - scarface74 4y ago
      I bought my first laptop in 1996. It was definitely much less efficient power wise than my ARM 15 inch Mac. It could barely last 2 hours on battery. My current laptop can easily last two workdays (16 hours) without being plugged in doing normal work.
  - andrei_says_ 4y ago
    I continue to use heroku because I’ll never be able to “write a bunch of yaml” matching my needs. I don’t want to even enter that world of agony.
bmitc 4y ago
I strongly suspect that the impact of LLMs is overblown. Unless they suddenly are able to synthesize complex and ill-defined requirements that change over time, I don't expect their impact will be beyond novelty.
- mistercow 4y ago
  Their impact is already beyond novelty. I don’t know when we’ll get to the point of not needing to understand the underlying concepts, but that much isn’t necessary for the models to be useful. Their use right now is just as time savers.
  - bmitc 4y ago
    > Their impact is already beyond novelty. Do you have some examples?
    - mistercow 4y ago
      Copilot. For that matter, TabNine several years earlier, even when using a low parameter local GPT-2.
    - JoeyJoJoJr 4y ago
      One anecdotal example: I gave ChatGPT 4 some prompts to write a shader to generate a signed distance field based on specific requirements. Within half an hour I had a working result. It solved a specific problem I was having that would have otherwise required in-depth research or finding an expert in two niche areas.
      - timr 4y ago
        See, this is the danger zone. How do you know that the solution you were given is correct? Even if you're an expert in the space, if you didn't know enough to solve the problem yourself, you're by definition, incapable of evaluating the output of a mindless text-generator. Now, I grant you that a lot of people work this way today. They're in trouble. We're about to enter the golden age for competency, and the dark ages for incompetence.
schwartzworld 4y ago
> don't require people to understand the solution in order to create a program. For example, it may be possible for people to create spellcheckers without understanding the rules of spelling Lol, no. How would you validate the output of that program if you didn't understand the rules?
- nextaccountic 4y ago
  Ultimately I suspect it will be validated by an AI too. Yeah, the average quality will nosedive, but programming will finally be open to the masses
  - schwartzworld 4y ago
    It's open to the masses now. I taught myself using free resources online.
emrah 4y ago
> LLMs will change all of this AI ultimately means humans are obsolete. Machines made most physical labor obsolete, AI will make most mental work obsolete eventually.
andrei_says_ 4y ago
> LLMs will change all of this. LLMs have weighted predictive algorithms for auto completion which are often useful but do not and will not have the necessary understanding of context, ever. By design. Therefore the reliability of their contribution will always be limited by the complete absence of context. If my understanding is incorrect please help me understand why.

benj111 4y ago

The programmer has to understand the problem. If the person behind the keyboard offloads that to AI, the AI is the programmer. The person behind the keyboard is basically just a manager.

scarface74 4y ago
The gap between “understanding the problem” and “coming up with a solution” has become a lot narrower over the years. 1. To do anything interesting and at any speed on my Apple //e I had to know assembly. 2. Then I got a Mac in 1992 and the original Mac Toolbox routines could do a lot from a higher level language. 3. Then Visual Basic made graphical programs easier and later WinForms, C# etc. 4. Then AWS made provisioning infrastructure a matter of just writing a bunch of yaml 5. Now that LLMs exist and ChatGPT is well trained on AWS APIs in various languages, I can use natural language and it spits out code that works 95% of the time with only slight tweaks if any.
- benj111 4y ago
  I'd state it more as, you don't have to reinvent to foundations to solve a new problem. If your problem is writing a link list, you still have to understand linked lists. If your problem just happens to be solved by linked lists, you can rely on a library. Re point 5. If debugging is twice as hard as writing code. How long before AI becomes half as intelligent as the average programmer making said programs undebuggable by the average programmer? Is it a good thing to have black boxes writing black boxes?
  - owl57 4y ago
    > How long before AI becomes half as intelligent as the average programmer making said programs undebuggable by the average programmer? It doesn't work this way. A programmer with a bit of experience knows to curb the desire for cleverness and write code they — and their teammates, probably having comparable intelligence and even less knowledge about specific code — can debug. Unless they are on a coffee trip and hacking on a cool weekend project, of course. Now that's another question, what in the hell will AI's weekend projects be.
    - benj111 4y ago
      >A programmer with a bit of experience knows to curb the desire for cleverness Are you suggesting AI will know to curb it's cleverness so that mere humans can understand it???
      - owl57 4y ago
        Setting aside all that debate about future AI agency and focusing on currently known mode of operation, why not. I haven't tested, but GPT-4 may already know enough to act on "this part of my/your code is too clever, please rewrite it simpler" and "please add this feature to that program, and don't be too clever".
        scarface74 4y ago
        GPT3 could do that.
        benj111 4y ago
        You're suggesting that "too clever" is a 'known unknown'. If it's an unknown unknown how can you spot the too clever thing to request that it's simplified? Edit: further, that assumes an AIs definition of simple matches that of a human programmer. If there's anything we've learnt recently, AI doesn't think like humans.
- marginalia_nu 4y ago
  Aren't the two always the same thing? I'll even go so far as to say you don't understand a problem if its solution does not appear childishly trivial.
  - scarface74 4y ago
    You don’t see the gap between “I want a natural language chatbot” and being able to create ChatGPT?
    - marginalia_nu 4y ago
      Right, the gap is a whole lot deeper understanding of the problem than "1 chatgpt plx", probably starting with understanding just what chatgpt is.
- wruza 4y ago
  Each of these steps made computation increasingly more expensive cycle- and power-wise. We counteracted that with Moore’s for a while, then had to go “scalable”. It worked because it was still cheaper than hiring an employee. I bet that LLMs will create another level of bloat industry where computation will be as expensive as making a human to do the same thing, because there will be no human-can-do-it constraint anymore. This promise of “AI will create cool software” is akin to “robots will fly in cars in vertical cities” of 19/20th century scifi. All we got is worse traffic and exponentially expensive housing.
  - scarface74 4y ago
    I bought my first laptop in 1996. It was definitely much less efficient power wise than my ARM 15 inch Mac. It could barely last 2 hours on battery. My current laptop can easily last two workdays (16 hours) without being plugged in doing normal work.
- andrei_says_ 4y ago
  I continue to use heroku because I’ll never be able to “write a bunch of yaml” matching my needs. I don’t want to even enter that world of agony.

bmitc 4y ago

I strongly suspect that the impact of LLMs is overblown. Unless they suddenly are able to synthesize complex and ill-defined requirements that change over time, I don't expect their impact will be beyond novelty.

mistercow 4y ago
Their impact is already beyond novelty. I don’t know when we’ll get to the point of not needing to understand the underlying concepts, but that much isn’t necessary for the models to be useful. Their use right now is just as time savers.
- bmitc 4y ago
  > Their impact is already beyond novelty. Do you have some examples?
  - mistercow 4y ago
    Copilot. For that matter, TabNine several years earlier, even when using a low parameter local GPT-2.
  - JoeyJoJoJr 4y ago
    One anecdotal example: I gave ChatGPT 4 some prompts to write a shader to generate a signed distance field based on specific requirements. Within half an hour I had a working result. It solved a specific problem I was having that would have otherwise required in-depth research or finding an expert in two niche areas.
    - timr 4y ago
      See, this is the danger zone. How do you know that the solution you were given is correct? Even if you're an expert in the space, if you didn't know enough to solve the problem yourself, you're by definition, incapable of evaluating the output of a mindless text-generator. Now, I grant you that a lot of people work this way today. They're in trouble. We're about to enter the golden age for competency, and the dark ages for incompetence.

schwartzworld 4y ago

> don't require people to understand the solution in order to create a program. For example, it may be possible for people to create spellcheckers without understanding the rules of spelling Lol, no. How would you validate the output of that program if you didn't understand the rules?

nextaccountic 4y ago
Ultimately I suspect it will be validated by an AI too. Yeah, the average quality will nosedive, but programming will finally be open to the masses
- schwartzworld 4y ago
  It's open to the masses now. I taught myself using free resources online.

emrah 4y ago

> LLMs will change all of this AI ultimately means humans are obsolete. Machines made most physical labor obsolete, AI will make most mental work obsolete eventually.

andrei_says_ 4y ago

> LLMs will change all of this. LLMs have weighted predictive algorithms for auto completion which are often useful but do not and will not have the necessary understanding of context, ever. By design. Therefore the reliability of their contribution will always be limited by the complete absence of context. If my understanding is incorrect please help me understand why.