4 ms·

>Anyone who agrees with any of the 3 sentiments, what is the justification backing up your belief? Do you feel the general cynicism/dismissiveness is justifiabl

by robswc 3y ago

>Anyone who agrees with any of the 3 sentiments, what is the justification backing up your belief? Do you feel the general cynicism/dismissiveness is justifiable? >While most junior level dev work is easily automatable, mid-senior level dev work cannot be replaced and in 5 years time will be no sooner able to be replaced Because companies were not set up with the idea of being able to contextualize and feed all their internal information into an AI. Maybe going forward, this will change... _maybe_ in 10-20 years (rather arbitrary I guess) one or two "managers" could use AI to interface with customers but for the foreseeable future, I just don't see how it could work. Assume you had an LLM that would never hallucinate or give a "wrong" answer. Someone still has to hook that up and constantly feed it context or its rather useless.

atleastoptimal 3y ago
This is something I understand as a criticism and my assumptions for why it won't hold have to do with my assumption (based on product releases and papers I've read) that the following capabilities will be rolled out in the coming years: 1. Fine tuning LLM's with a company's internal information (already on the "coming soon" part of OpenAI ChatGPT enterprise) 2. More agentic models able to use tools to retrieve, process, and record data, then access its own work as a sort of memory (lots of open source tools are working on these problems now) 3. Much larger context windows with lower tradeoff, and "networks" of AI models controlled by higher level agents with rotating/overlapping context windows
- verdverm 3y ago
  We will need a completely new set of models, LLMs are not sufficiently complex enough to deal with what we deal with. You cannot remove hallucinations so will always need humans to double check what the AI outputs. Most of the time, it's faster and cheaper for the human to do it the first time. Make the AI the junior, not the human We first have to move beyond text token streams, but to do so requires insanely large models. It will probably be 5 years before we can even make enough GPUs to support every company having an OpenAI scale model they can build & run Large context windows come with their own problems and there are other papers that show the agent pyramids are fools errand (the "liar" checks the "liar" issue) Generally, I think you are holding onto the positive hype and not looking at the limitation, negative results, and real world, where these programming AI become more of a burden than a boon. So no, jobs are not going to magically disappear in the next 5 years. There is plenty of economic history and research to more accurately estimate the job changes that will come with a new technology
  - atleastoptimal 3y ago
    > You cannot remove hallucinations What makes you sure that will hold for the next 5 years as tech improves? > It will probably be 5 years before we can even make enough GPUs to support every company having an OpenAI scale model they can build & run Not every company needs its own monster model, the point is OpenAI type companies will build the base model and other companies will fork their own versions >Large context windows come with their own problems and there are other papers that show the agent pyramids are fools errand (the "liar" checks the "liar" issue) I looked up agent pyramids and "liar" checks the "liar" issue and found nothing. Could you link those papers you're referencing? I don't imagine jobs are just going to "disappear", what I do imagine is that gradually senior devs will offload more and more of their work to internal AI models. Big companies like Google will have the best, so they'll signal how things will change for the rest. As the models improve, the chance of any task a model is set to has a higher and higher chance of success and more human readable documentation, to the point where senior devs no longer have to check that they did their job correctly, and other AIs will begin to read/use the code and documentation other AIs wrote. This process will inch up the ladder piece by piece until senior devs begin assigning AI tasks that other senior devs would have done, notice that the AI does it better, and the total efficacy of the senior staff erodes. At this point companies will emerge entirely autonomously run, which is the final signal for other companies that human labor is no longer needed the moment something is automatable.
    - verdverm 3y ago
      1. I said we will need new models, not that it wasn't possible 2. It's more of a total through out problem, even if OpenAI ran all the models, you still need the hardware capacity. We are already two years behind in supply vs demand vs manufacturing capacity. A lot of people cannot even try to build, this happened to me over the weekend 3. I'm not going to do research for you. Certainly doing simple searches is not enough to understand where things are going. You have to stay tuned in to the progress and have your finger on the pulse. There are plenty of prominent accounts on X to follow and be aware of what I'm pointing out here. You claims about what is to come are likely exaggerated because you are only understanding the surface at this point. Keep learning and researching!

by robswc 3y ago

atleastoptimal 3y ago

This is something I understand as a criticism and my assumptions for why it won't hold have to do with my assumption (based on product releases and papers I've read) that the following capabilities will be rolled out in the coming years: 1. Fine tuning LLM's with a company's internal information (already on the "coming soon" part of OpenAI ChatGPT enterprise) 2. More agentic models able to use tools to retrieve, process, and record data, then access its own work as a sort of memory (lots of open source tools are working on these problems now) 3. Much larger context windows with lower tradeoff, and "networks" of AI models controlled by higher level agents with rotating/overlapping context windows

verdverm 3y ago
We will need a completely new set of models, LLMs are not sufficiently complex enough to deal with what we deal with. You cannot remove hallucinations so will always need humans to double check what the AI outputs. Most of the time, it's faster and cheaper for the human to do it the first time. Make the AI the junior, not the human We first have to move beyond text token streams, but to do so requires insanely large models. It will probably be 5 years before we can even make enough GPUs to support every company having an OpenAI scale model they can build & run Large context windows come with their own problems and there are other papers that show the agent pyramids are fools errand (the "liar" checks the "liar" issue) Generally, I think you are holding onto the positive hype and not looking at the limitation, negative results, and real world, where these programming AI become more of a burden than a boon. So no, jobs are not going to magically disappear in the next 5 years. There is plenty of economic history and research to more accurately estimate the job changes that will come with a new technology
- atleastoptimal 3y ago
  > You cannot remove hallucinations What makes you sure that will hold for the next 5 years as tech improves? > It will probably be 5 years before we can even make enough GPUs to support every company having an OpenAI scale model they can build & run Not every company needs its own monster model, the point is OpenAI type companies will build the base model and other companies will fork their own versions >Large context windows come with their own problems and there are other papers that show the agent pyramids are fools errand (the "liar" checks the "liar" issue) I looked up agent pyramids and "liar" checks the "liar" issue and found nothing. Could you link those papers you're referencing? I don't imagine jobs are just going to "disappear", what I do imagine is that gradually senior devs will offload more and more of their work to internal AI models. Big companies like Google will have the best, so they'll signal how things will change for the rest. As the models improve, the chance of any task a model is set to has a higher and higher chance of success and more human readable documentation, to the point where senior devs no longer have to check that they did their job correctly, and other AIs will begin to read/use the code and documentation other AIs wrote. This process will inch up the ladder piece by piece until senior devs begin assigning AI tasks that other senior devs would have done, notice that the AI does it better, and the total efficacy of the senior staff erodes. At this point companies will emerge entirely autonomously run, which is the final signal for other companies that human labor is no longer needed the moment something is automatable.
  - verdverm 3y ago
    1. I said we will need new models, not that it wasn't possible 2. It's more of a total through out problem, even if OpenAI ran all the models, you still need the hardware capacity. We are already two years behind in supply vs demand vs manufacturing capacity. A lot of people cannot even try to build, this happened to me over the weekend 3. I'm not going to do research for you. Certainly doing simple searches is not enough to understand where things are going. You have to stay tuned in to the progress and have your finger on the pulse. There are plenty of prominent accounts on X to follow and be aware of what I'm pointing out here. You claims about what is to come are likely exaggerated because you are only understanding the surface at this point. Keep learning and researching!