There are certainly benefits to llms. Just the more casual interaction is one. This to me very much mirrors graphics as compared to terminal. many many things done on computers do not require graphics and can be done at the terminal but usage by ordinary folks increased with the advent of graphics systems and guis. graphics require more power but are a more user friendly interface. Similarly llms take more power than search queries but have a more user friendly interface. In addition it has generative capability allowing folks without significant talent to do some things they could not before. There are benefits but right now its a matter of if the benefits outweigh the costs. Seriously if this was graphics vs terminal the graphics basically double power usage when doing nothing in particular that need it. You start throwing a bunch of monitors up and it will bring it up significantly and actually using graphics like gaming and video go way more and it keeps increasing as we increase standar resolutions and what is considered normal performance. We would use way less power globally if folks only used computers at terminals and went back to listening to the radio.
A stark difference is that GUIs are designed with the same feature set as the CLI. Unlike an LLM, the GUI is not going to start suddenly making up data, offering nonexistent features, or straight up lying. The GUI is a more easily accessible interface that more or less behaves like the CLI. It’s like an orbital sander versus manual sanding - generally easier but you can also really fuck shut up if you don’t know what you’re doing.
Furthermore, coding is not merely clicking buttons or typing commands. You need to understand your language, your compiler, the intended runtime environment, etc., as well as best practices and then also have the purely human ability to create a new way of doing something. The LLM simply regurgitates what it’s been trained on without regard to the quality of that data - it can’t make value judgements on best approaches or even syntax, it’s just monkey-see, monkey-do. Hell, it can’t even sometimes differentiate between the correct syntaxes for different languages. In the end it gets you nothing but the illusion of convenience and the very real problem of non-functional, unmaintainable, and/or unoptimized code. It’s a fool’s bargain.
I think you are looking at llms in to large a context. Your issues you have with it are the same as search and if used as a further abstraction of search it is going to carry forward weaknesses. LLM’s trained more narrowly for specific purposes are going to do better if their narrow training data is of high quality rather than throwing everything at it.
I believe you are mistaken - I’m speaking about LLMs within the context of FOSS and coding as that’s the subject of OP’s post and the article they linked to. You may have thoughts about LLMs in general, but you don’t seem to be addressing their use (and abuse) when it comes to coding or giving them carte blanche when it comes access to FOSS repos. Because if you, as the article suggests, just bend over, you’re giving access to all LLMs, not just some hypothetical specially trained one.
Yeah my response was directly for your statement that there are no real and tangible benefits of llms. I think how useful llms may or may not be is immaterial when it comes to the repos and they will be used in coding. Wikipedia went a good direction and worked out a deal for them to contribute back and linus was sensible in his response to usage for coding.
“AI coding assistants feel productive because they give instant feedback. You type a prompt and code drops in right away. That loop feels like progress, the same reward you get from closing a ticket or fixing a failing test. The problem is that dopamine rewards activity in the editor, not working code in production.”
regardless of if the feedback is from the ai or a blog example its up to the programmer to evaluate the relative value of what they are looking at. No one codes from nothing for things they have not done before and even then they are likely to start with code they wrote before and start editing. Also ai assitance for coding is not limited to coding. I found ai commenting to be useful as I am more likely to correct an ai’s comment as I go than comment as I code. Its very easy to be like. I will get the commenting in later but when you see something you don’t like, at least for me, then I don’t like to leave it there.
There are certainly benefits to llms. Just the more casual interaction is one. This to me very much mirrors graphics as compared to terminal. many many things done on computers do not require graphics and can be done at the terminal but usage by ordinary folks increased with the advent of graphics systems and guis. graphics require more power but are a more user friendly interface. Similarly llms take more power than search queries but have a more user friendly interface. In addition it has generative capability allowing folks without significant talent to do some things they could not before. There are benefits but right now its a matter of if the benefits outweigh the costs. Seriously if this was graphics vs terminal the graphics basically double power usage when doing nothing in particular that need it. You start throwing a bunch of monitors up and it will bring it up significantly and actually using graphics like gaming and video go way more and it keeps increasing as we increase standar resolutions and what is considered normal performance. We would use way less power globally if folks only used computers at terminals and went back to listening to the radio.
A stark difference is that GUIs are designed with the same feature set as the CLI. Unlike an LLM, the GUI is not going to start suddenly making up data, offering nonexistent features, or straight up lying. The GUI is a more easily accessible interface that more or less behaves like the CLI. It’s like an orbital sander versus manual sanding - generally easier but you can also really fuck shut up if you don’t know what you’re doing.
Furthermore, coding is not merely clicking buttons or typing commands. You need to understand your language, your compiler, the intended runtime environment, etc., as well as best practices and then also have the purely human ability to create a new way of doing something. The LLM simply regurgitates what it’s been trained on without regard to the quality of that data - it can’t make value judgements on best approaches or even syntax, it’s just monkey-see, monkey-do. Hell, it can’t even sometimes differentiate between the correct syntaxes for different languages. In the end it gets you nothing but the illusion of convenience and the very real problem of non-functional, unmaintainable, and/or unoptimized code. It’s a fool’s bargain.
I think you are looking at llms in to large a context. Your issues you have with it are the same as search and if used as a further abstraction of search it is going to carry forward weaknesses. LLM’s trained more narrowly for specific purposes are going to do better if their narrow training data is of high quality rather than throwing everything at it.
I believe you are mistaken - I’m speaking about LLMs within the context of FOSS and coding as that’s the subject of OP’s post and the article they linked to. You may have thoughts about LLMs in general, but you don’t seem to be addressing their use (and abuse) when it comes to coding or giving them carte blanche when it comes access to FOSS repos. Because if you, as the article suggests, just bend over, you’re giving access to all LLMs, not just some hypothetical specially trained one.
Yeah my response was directly for your statement that there are no real and tangible benefits of llms. I think how useful llms may or may not be is immaterial when it comes to the repos and they will be used in coding. Wikipedia went a good direction and worked out a deal for them to contribute back and linus was sensible in his response to usage for coding.
“AI coding assistants feel productive because they give instant feedback. You type a prompt and code drops in right away. That loop feels like progress, the same reward you get from closing a ticket or fixing a failing test. The problem is that dopamine rewards activity in the editor, not working code in production.”
https://www.cerbos.dev/blog/productivity-paradox-of-ai-coding-assistants
regardless of if the feedback is from the ai or a blog example its up to the programmer to evaluate the relative value of what they are looking at. No one codes from nothing for things they have not done before and even then they are likely to start with code they wrote before and start editing. Also ai assitance for coding is not limited to coding. I found ai commenting to be useful as I am more likely to correct an ai’s comment as I go than comment as I code. Its very easy to be like. I will get the commenting in later but when you see something you don’t like, at least for me, then I don’t like to leave it there.