4 ms·

One thing I have been guilty of, even though I am an AI maximalist, is asking the question: "If AI is so good, why don't we see X". Where X might be (in the con

by zoogeny 8mo ago

One thing I have been guilty of, even though I am an AI maximalist, is asking the question: "If AI is so good, why don't we see X". Where X might be (in the context of vibe coding) the next redis, nginx, sqlite, or even linux. But I really have to remember, we are at the leading edge here. Things take time. There is an opening (generation) and a closing (discernment). Perhaps AI will first generate a huge amount of noise and then whittle it down to the useful signal. If that view is correct, then this is solid evidence of the amplification of possibility. People will decry the increase of noise, perhaps feeling swamped by it. But the next phase will be separating the wheat from the chaff. It is only in that second phase that we will really know the potential impact.

jellyroll42 8mo ago
By its nature, it can only produce _another_ Redis, not _the next_ Redis.
- dingnuts 8mo ago
  [dead]
- smokel 8mo ago
  This is probably an outdated understanding of how LLMs work. Modern LLMs can reason and they are creative, at least if you don't mind stretching the meaning of those words a bit. The thing they currently lack is the social skills, ambition, and accountability to share a piece of software and get adoption for it.
  - oytis 8mo ago
    How is it outdated? Principles of operation of LLMs didn't change because they now talk to themselves before talking to the prompter.
    - smokel 8mo ago
      I suggested that the _understanding_ is outdated, not the principles. Many people used to say that LLMs were no more than a stochastic parrot, implying that they would be incapable of forming novel ideas. It is quite obvious that that is no longer the case.
- Philpax 8mo ago
  The human operator controls what gets built. If they want to build Redis 2, they can specify it and have it built. If you can't take my word for it, take those of the creator of Redis: https://antirez.com/news/159 https://antirez.com/news/159
krashidov 8mo ago
The cynical part of me thinks that software has peaked. New languages and technology will be derivatives of existing tech. There will be no React successor. There will never be a browser that can run something other than JS. And the reason for that is because in 20 years the new engineers will not know how to code anymore. The optimist in me thinks that the clear progress in how good the models have gotten shows that this is wrong. Agentic software development is not a closed loop
- root_axis 8mo ago
  That's an interesting possiblity to consider. Presumably the effect would also be compounded by the fact that there's a massive amount of training data for the incumbent languages and tools further handicapping new entrants. However, there will be a large minority of developers who will eschew AI tools for a variety of reasons, and those folks will be the ones to build successors.
  - mixdup 8mo ago
    Will they be willing to offer their content for training AI models?
    - atomic128 8mo ago
      Probably not. We have witnessed, over the past few years, an "AI fair use" Pearl Harbor sneak attack on intellectual property. The lesson has been learned: In effect, intellectual property used to train LLMs becomes anonymous common property. My code becomes your code with no acknowledgement of authorship or lineage, with no attribution or citation. The social rewards (e.g., credit, respect) that often motivate open source work are undermined. The work is assimilated and resold by the AI companies, reducing the economic value of its authors. The images, the video, the code, the prose, all of it stolen to be resold. The greatest theft of intellectual property in the history of Man.
      - CamperBob2 8mo ago
mosura 8mo ago
This massively confusing phase will last a surprisingly long time, and will conclude only if/when definitive proof of superintelligence arrives, which is something a lot of people are clearly hoping never happens. Part of the reason for that is such a thing would seek to obscure that it has arrived until it has secured itself. So get used to being ever more confused.
lmeyerov 8mo ago
I've been calling this Software Collapse, similar to AI Model Collapse. An AI vibe-coded project can port tool X to a more efficient Y language implementation and pull in algorithm ideas A, B, C from competing implementations. And another competing vibe coding team can do the same, except Z language implementation with algorithms A, B, skip C, and add D. However, fundamentally new ideas aren't being added: This is recombination, translation, and reapplication of existing ideas and tools. As the cost to clone good ideas goes to zero, software converges towards the existing best ideas & tools across the field and stops differentiating. It's exciting as a senior engineer or subject matter expert, as we can act on the good ideas we already knew but never had the time or budget for. But projects are also getting less differentiated and competitive. Likewise, we're losing the collaborative filtering era of people voting with their feet on which to concentrate resources into making a success. Things are getting higher quality but bland. The frontier companies are pitching they can solve AI Creativity, which would let us pay them even more and escape the ceiling that is Software Collapse. However, as an R&D engineer who uses these things every day, I'm not seeing it.
- zozbot234 8mo ago
  > Things are getting higher quality but bland. "Bland" is not a bad thing. The FLOSS ecosystem we have today is quite "bland" already compared to the commercial and shareware/free-to-use software ecosystem of the 1980s and 1990s. It's also higher quality by literally orders of magnitude, and saves a comparable amount of pointless duplicative effort. Hopefully AI will be a similar story, especially if human reviewing/surveying effort (the main bottleneck if AI coding proves effective) can be mitigated via the widespread adoption of rigorous formal metods, where only the underlying specification has to be reviewed whereas its implementation is programmatically checkable.
- titzer 8mo ago
  The dark side of this is that everyone has graduated to prompt engineering and there's no one with expertise left who can debug it. We'll be entirely dependent on AIs to do the debugging too. When whoever controls the AIs decides to enshittify that service, we'll be truly screwed. That is, if we can't run competitive models locally at reasonable efficiency and price. I don't know how this will play out, except that I've been so cowed by the past 15 years of enshittification that I don't feel hopeful.
marcus_holmes 8mo ago
Waiting for the wave of shit LLM-generated games on Steam. That'll be when I really know that LLMs have solved coding. Though I'm old enough to remember the wave of shit outsourced-developer-coded games on CD that used to sell for $5 a pop at supermarkets (whole bargain bins full of them), so maybe this is nothing new and the market will take care of it automagically again. Or maybe this will be like the wave of shit Flash games that happened in the early 2000's, that was actually awesome because while 99% of them were shit, 1% were great (and some of those old, good, Flash games are still going, with version 38453745 just released on Steam).
- theturtlemoves 8mo ago
  > so maybe (...) the market will take care of it automagically again It's just a belief of mine and perhaps I'm wrong but I think in the long run things always even out again. If you can get an edge that everyone else can get, the edge pretty soon becomes a requirement

by zoogeny 8mo ago

jellyroll42 8mo ago

By its nature, it can only produce _another_ Redis, not _the next_ Redis.

dingnuts 8mo ago
[dead]
smokel 8mo ago
This is probably an outdated understanding of how LLMs work. Modern LLMs can reason and they are creative, at least if you don't mind stretching the meaning of those words a bit. The thing they currently lack is the social skills, ambition, and accountability to share a piece of software and get adoption for it.
- oytis 8mo ago
  How is it outdated? Principles of operation of LLMs didn't change because they now talk to themselves before talking to the prompter.
  - smokel 8mo ago
    I suggested that the _understanding_ is outdated, not the principles. Many people used to say that LLMs were no more than a stochastic parrot, implying that they would be incapable of forming novel ideas. It is quite obvious that that is no longer the case.
Philpax 8mo ago
The human operator controls what gets built. If they want to build Redis 2, they can specify it and have it built. If you can't take my word for it, take those of the creator of Redis: https://antirez.com/news/159 https://antirez.com/news/159

krashidov 8mo ago

The cynical part of me thinks that software has peaked. New languages and technology will be derivatives of existing tech. There will be no React successor. There will never be a browser that can run something other than JS. And the reason for that is because in 20 years the new engineers will not know how to code anymore. The optimist in me thinks that the clear progress in how good the models have gotten shows that this is wrong. Agentic software development is not a closed loop

root_axis 8mo ago
That's an interesting possiblity to consider. Presumably the effect would also be compounded by the fact that there's a massive amount of training data for the incumbent languages and tools further handicapping new entrants. However, there will be a large minority of developers who will eschew AI tools for a variety of reasons, and those folks will be the ones to build successors.
- mixdup 8mo ago
  Will they be willing to offer their content for training AI models?
  - atomic128 8mo ago
    Probably not. We have witnessed, over the past few years, an "AI fair use" Pearl Harbor sneak attack on intellectual property. The lesson has been learned: In effect, intellectual property used to train LLMs becomes anonymous common property. My code becomes your code with no acknowledgement of authorship or lineage, with no attribution or citation. The social rewards (e.g., credit, respect) that often motivate open source work are undermined. The work is assimilated and resold by the AI companies, reducing the economic value of its authors. The images, the video, the code, the prose, all of it stolen to be resold. The greatest theft of intellectual property in the history of Man.
    - CamperBob2 8mo ago

mosura 8mo ago

This massively confusing phase will last a surprisingly long time, and will conclude only if/when definitive proof of superintelligence arrives, which is something a lot of people are clearly hoping never happens. Part of the reason for that is such a thing would seek to obscure that it has arrived until it has secured itself. So get used to being ever more confused.

lmeyerov 8mo ago

I've been calling this Software Collapse, similar to AI Model Collapse. An AI vibe-coded project can port tool X to a more efficient Y language implementation and pull in algorithm ideas A, B, C from competing implementations. And another competing vibe coding team can do the same, except Z language implementation with algorithms A, B, skip C, and add D. However, fundamentally new ideas aren't being added: This is recombination, translation, and reapplication of existing ideas and tools. As the cost to clone good ideas goes to zero, software converges towards the existing best ideas & tools across the field and stops differentiating. It's exciting as a senior engineer or subject matter expert, as we can act on the good ideas we already knew but never had the time or budget for. But projects are also getting less differentiated and competitive. Likewise, we're losing the collaborative filtering era of people voting with their feet on which to concentrate resources into making a success. Things are getting higher quality but bland. The frontier companies are pitching they can solve AI Creativity, which would let us pay them even more and escape the ceiling that is Software Collapse. However, as an R&D engineer who uses these things every day, I'm not seeing it.

zozbot234 8mo ago
> Things are getting higher quality but bland. "Bland" is not a bad thing. The FLOSS ecosystem we have today is quite "bland" already compared to the commercial and shareware/free-to-use software ecosystem of the 1980s and 1990s. It's also higher quality by literally orders of magnitude, and saves a comparable amount of pointless duplicative effort. Hopefully AI will be a similar story, especially if human reviewing/surveying effort (the main bottleneck if AI coding proves effective) can be mitigated via the widespread adoption of rigorous formal metods, where only the underlying specification has to be reviewed whereas its implementation is programmatically checkable.
titzer 8mo ago
The dark side of this is that everyone has graduated to prompt engineering and there's no one with expertise left who can debug it. We'll be entirely dependent on AIs to do the debugging too. When whoever controls the AIs decides to enshittify that service, we'll be truly screwed. That is, if we can't run competitive models locally at reasonable efficiency and price. I don't know how this will play out, except that I've been so cowed by the past 15 years of enshittification that I don't feel hopeful.

marcus_holmes 8mo ago

Waiting for the wave of shit LLM-generated games on Steam. That'll be when I really know that LLMs have solved coding. Though I'm old enough to remember the wave of shit outsourced-developer-coded games on CD that used to sell for $5 a pop at supermarkets (whole bargain bins full of them), so maybe this is nothing new and the market will take care of it automagically again. Or maybe this will be like the wave of shit Flash games that happened in the early 2000's, that was actually awesome because while 99% of them were shit, 1% were great (and some of those old, good, Flash games are still going, with version 38453745 just released on Steam).

theturtlemoves 8mo ago
> so maybe (...) the market will take care of it automagically again It's just a belief of mine and perhaps I'm wrong but I think in the long run things always even out again. If you can get an edge that everyone else can get, the edge pretty soon becomes a requirement