Great post. I particularly like the 4 column pr view idea. I’ve built something similar to review speckit specs. The scaffolding to ensure that code actually works makes sense too.
There’s one thing that’s bothering me though — all of these have us consuming what LLMs produce. I learned and internalized how things worked by building systems. At least within the constraints of the abstractions I was using. This makes me think of TikTok and shorts. Everybody is just watching. Maybe that’s not charitable though. Maybe the abstraction has now changed from an “os/gui/event loop” to “llm” and the rest is just accidental complexity.
>> A more serious irony is that the automatic control system has been put in place because it can do the job better than the operator, yet the operator is being asked to monitor that it is working effectively.
Automatic control is good for standard case, and the expected errors. For weird cases, for now humans are better to detect them and press the big red button.
Except experience shows they aren't. Experience shows that when humans are just pressing "ok" 99% of the time they'll press it 100% of the time and completely miss the other cases.
"Yup, looks good to me!" said the human reviewing thousands of lines of generated code everyday.
This has become common now, and I've seen it ruin large projects until nobody understands the codebase full of verbose nonsensical code and comments. It becomes impossible to further develop the project without AI. Contributors not using AI for any reason are all leaving because no one enjoys working with mindless robots (and humans) as colleagues. It's demeaning and devaluing their labor and humanity. They're not getting "left behind", they are leaving the madness of crowds behind.
> ..the possibility of runaway dynamics: once a critical threshold is crossed, small increases in adoption can trigger rapid population-level shifts toward *persistent dependence, with abrupt losses in cognitive competence*.
We can all see it happening. As this technology permeates society, the challenge to maintain control, coherence, and quality will become greater and ever more serious - for example, AI in healthcare, military, management and governance.
Even if the human's judgement is correct in every LGTM, there's a bias towards merging. Each merge that adds a feature but degrades understandability 1% will be merged. It won't be held back until it improves it by 5% instead. After 100 merges you have degraded understandability 100% even though nobody made an error.
To be fair, models are more accurate than how they were earlier this year. In march I was still frustrated and was writing my own event loops because the model simply wasn't built for it. By January I was still hand checking statistical outputs to make sure it wasn't making any errors, and it was. After august it was pretty much over. There were no stats errors, code generated was shit but it could do what is asked. Instruction following is still a problem, and I sometimes catch silly things in the scrolling diffs, but it has been much easier since then.
I am often thinking recently of the littel loop diagrams from the "The Capability Trap" paper. We need to keep a loop reinvesting into our own Human capability in anything we're doing.
This is copium that is gone with the wind when the models perform up to the level of standard that the user requires. I'll give the author 6 months before they feel like its okay to just send it.
Karpathy had the same trajectory. Primeagean had the same trajectory. I think in the end everyone except for Jonathan Blow will have the same trajectory
There’s one thing that’s bothering me though — all of these have us consuming what LLMs produce. I learned and internalized how things worked by building systems. At least within the constraints of the abstractions I was using. This makes me think of TikTok and shorts. Everybody is just watching. Maybe that’s not charitable though. Maybe the abstraction has now changed from an “os/gui/event loop” to “llm” and the rest is just accidental complexity.
Automatic control is good for standard case, and the expected errors. For weird cases, for now humans are better to detect them and press the big red button.
This has become common now, and I've seen it ruin large projects until nobody understands the codebase full of verbose nonsensical code and comments. It becomes impossible to further develop the project without AI. Contributors not using AI for any reason are all leaving because no one enjoys working with mindless robots (and humans) as colleagues. It's demeaning and devaluing their labor and humanity. They're not getting "left behind", they are leaving the madness of crowds behind.
> ..the possibility of runaway dynamics: once a critical threshold is crossed, small increases in adoption can trigger rapid population-level shifts toward *persistent dependence, with abrupt losses in cognitive competence*.
We can all see it happening. As this technology permeates society, the challenge to maintain control, coherence, and quality will become greater and ever more serious - for example, AI in healthcare, military, management and governance.
Karpathy had the same trajectory. Primeagean had the same trajectory. I think in the end everyone except for Jonathan Blow will have the same trajectory
I guess you could ask the model. "Yes well done, you're a visionary software engineer".
The lawyers and judges who trusted an ML system to be up to snuff are currently pounding sand.