4 ms·

This is almost exactly how I see GPT and code. I have seen smart people who don't write code toss GPT a request and get a working thing but going from "hey wri

This is almost exactly how I see GPT and code. I have seen smart people who don't write code toss GPT a request and get a working thing but going from "hey write me an auto presser for a key on a timer" to "The client wants you to update the repo to handle this business logic and do this functionality" is a huge leap. From my perspective, you still need a developer-minded person to do the job, AI just kind of makes their lives easier in the process. I agree with the sentiment that it's the same in the art world. It's easy to get a compelling image but to get specific meaningful images usually requires a lot of post processing that a layman wouldn't be capable of doing.

ilaksh 3y ago
True but I feel like people are assuming these limitations will be long term. But we have seen progress in leaps and bounds. LLM-based coding tools are getting better. LLMs are getting better. Context size is increasing. And the interest in LLMs is even motivating development of new approaches that will be more effective. Give it a few years, things like Lecun's JEPAs or whatever hybrid supercoder DeepMind is working on, or some open-source LLM, will blow GPT-4 out of the water for programming.
- chankstein38 3y ago
  I agree these things are going to move likely faster than we can imagine and become better than we can imagine. That said, last week I got a ticket that had no body but a title akin to "Add a property for this" with no specifications about the system. When I asked for clarity on the ticket the ask was very different than the title implied. When I started the work it started to look different than was described so I went back to the requestor and explained the discrepancy and we changed the direction of the change. I say this to say, one skillset I have as a developer is taking the vague requests product owners have and figuring out how to turn them into actionable code steps in a massive existing codebase with several repos. I don't say this to say I'm impossible to replace but to say that half the time people don't even know what they want or how to describe it. Then from there you have giant codebases that wouldn't fit in anything but the biggest (current) context windows. I agree the accuracy limitations will likely evaporate but these things aren't necessarily something an LLM can solve. I'm probably going to be proven wrong over time but I use GPT for code pretty regularly and right now I'm not too worried about my job.
  - ilaksh 3y ago
    I don't know for sure if an LLM could do it. But theoretically one could build a system that sends chat messages asking for clarification and also eventually with more context or something is able to translate vague or stupid "requirements" into ones that make sense. In five years or so the capabilities may be pretty amazing.
    - chankstein38 3y ago
      Definitely agreed! I'm both nervous and excited. I fear for my livelihood but also if we continue to make even a percentage of the progress we've made in the last year, the next 5 years are going to be wild!

ilaksh 3y ago

True but I feel like people are assuming these limitations will be long term. But we have seen progress in leaps and bounds. LLM-based coding tools are getting better. LLMs are getting better. Context size is increasing. And the interest in LLMs is even motivating development of new approaches that will be more effective. Give it a few years, things like Lecun's JEPAs or whatever hybrid supercoder DeepMind is working on, or some open-source LLM, will blow GPT-4 out of the water for programming.

chankstein38 3y ago
I agree these things are going to move likely faster than we can imagine and become better than we can imagine. That said, last week I got a ticket that had no body but a title akin to "Add a property for this" with no specifications about the system. When I asked for clarity on the ticket the ask was very different than the title implied. When I started the work it started to look different than was described so I went back to the requestor and explained the discrepancy and we changed the direction of the change. I say this to say, one skillset I have as a developer is taking the vague requests product owners have and figuring out how to turn them into actionable code steps in a massive existing codebase with several repos. I don't say this to say I'm impossible to replace but to say that half the time people don't even know what they want or how to describe it. Then from there you have giant codebases that wouldn't fit in anything but the biggest (current) context windows. I agree the accuracy limitations will likely evaporate but these things aren't necessarily something an LLM can solve. I'm probably going to be proven wrong over time but I use GPT for code pretty regularly and right now I'm not too worried about my job.
- ilaksh 3y ago
  I don't know for sure if an LLM could do it. But theoretically one could build a system that sends chat messages asking for clarification and also eventually with more context or something is able to translate vague or stupid "requirements" into ones that make sense. In five years or so the capabilities may be pretty amazing.
  - chankstein38 3y ago
    Definitely agreed! I'm both nervous and excited. I fear for my livelihood but also if we continue to make even a percentage of the progress we've made in the last year, the next 5 years are going to be wild!