Call me old school, but I find the workflow of "divide and conquer" to be as helpful when working with LLMs, as without them. Although what is needed to be considered a "large scale task" varies by LLMs and implementation. Some models/implementations (seemingly Copilot) struggles with even the smallest change, while others breeze through them. Lots of trial and error is needed to find that line for each model/implementation :/
And at the same time, absurdly slow? ChatGPT is almost 3 years old and pretty much AI has still no positive economic impact.
So eg., one line of code which needed to handle dozens of hard-constraints on the system (eg., using a specific class, method, with a specific device, specific memory management, etc.) will very rarely be output correctly by an LLM.
Likewise "blank-page, vibe coding" can be very fast if "make me X" has only functional/soft-constraints on the code itself.
"Gigawatt LLMs" have brute-forced there way to having a statistical system capable of usefully, if not universally, adhreading to one or two hard constraints. I'd imagine the dozen or so common in any existing application is well beyond a Terawatt range of training and inference cost.
Now look at the past year specifically, and only at the models themselves, and you'll quickly realize that there's been very little real progress recently. Claude 3.5 Sonnet was released 11 months ago and the current SOTA models are only marginally better in terms of pure performance in real world tasks.
The tooling around them has clearly improved a lot, and neat tricks such as reasoning have been introduced to help models tackle more complex problems, but the underlying transformer architecture is already being pushed to its limits and it shows.
Unless some new revolutionary architecture shows up out of nowhere and sets a new standard, I firmly believe that we'll be stuck at the current junior level for a while, regardless of how much Altman & co. insist that AGI is just two more weeks away.
Even if it could perform at a similar level to an intern at a programming task, it lacks a great deal of the other attributes that a human brings to the table, including how they integrate into a team of other agents (human or otherwise). I won't bother listing them, as we are all humans.
I think the hype is missing the forest for the trees, and I think exactly this multi-agent dynamic might be where the trees start to fall down in front of us. That and the as currently insurmountable issues of context and coherence over long time horizons.
When you look at it from afar, it looks potentially good, but as you start looking into it for real, you start realizing none of it makes any sense. Then you make simple suggestions, it does something that looks like what you asked, yet completely missing the point.
An intern, no matter how bad it is, could only waste so much time and energy.
This makes wasting time and introducing mind-bogglingly stupid bugs infinitely scalable.
It will take some time for whatever reality is to actually show truthfully in the financials. When VC money stops subsidising datacentre costs, and businesses have to weigh the full price against real value provided, that is when we will see the reality of the situation.
I am content to be wrong either way, but my personal prediction is if model competence slows down around now, businesses will not be replacing humans en-mass, and the value provided will be notable but not world changing like expected.
Don’t get me wrong: the current models are already powerful and useful. However, there is still a lot of reason to remain skeptical of an imminent explosion in intelligence from these models.
I can't fire half my dev org tomorrow with that approach, I can't really fire anyone, so I guess it would be a big letdown for a lot of execs. Meanwhile though we just keep incrementally shipping more stuff faster at higher quality so I'm happy...
This works because it treats the LLM like what it actually is: an exceptionally good if slightly random text transformer.
For some reason my pessimism meter goes off when I see single sentence arguments “change has been slow”. Thanks for brining the conversation back.
-Being a parent to a small child and the associated sleep deprivation.
-His reluctance to read documentation.
-There being a language barrier between him the project owners. Emphasis here, as the LLM acts like someone who speaks through a particularly good translation service, but otherwise doesn't understand the language spoken.
Software today is written to accommodate every possible need of every possible user, and then a bunch of unneeded selling point features on top of that. These massive sprawling code bases made to deliver one-size fits all utility.
I don't need 3 million LOC Excel 365 to keep track of who is working on the floor on what day this week. Gemini 2.5 can write an applet that does that perfectly in 10 minutes.
LLMs are like bumpers on bowling lanes. Pro bowlers don't get much utility from them. Total noobs are getting more and more strikes as these "smart" bumpers get better and better at guiding their ball.
Nobody seems to consider that LLMs are democratizing programming, and allowing regular people to build programs that make their work more efficient. I can tell you that at my old school manufacturing company, where we have no programmers and no tech workers, LLMs have been a boon for creating automation to bridge gaps and even to forgo paid software solutions.
This is where the change LLMs will bring will come from. Not from helping an expert dev write boilerplate 30% faster.
I do like the idea of smaller programs fitting smaller needs being easy to access for everyone, and in my post history you would see me advocate for bringing software wages down so that even small businesses can have software capabilities in house. Software has so much to give to society outside of big VC flips and tech monoliths. Maybe AI is how we get there in the end.
But I think that supplanting humans with an AI workforce in the very near future might be stretching the projection of its capabilities too far. LLMs will be augmenting how businesses operate from now and into the future, but I am seeing clear roadblocks that make an autonomous AI agent unviable, and it seems to be fundamental limitations of LLMs, eg continuity and context. Advances recently seem to be from supplemental systems that try to patch those limitations. That suggests those limits are tricky, and until a new approach shows up, that is what drives my lack of faith in an AI agent revolution.
But it is clear to me that I could be wrong, and it could be a spectacular miscalculation. Maybe the robots will make me eat my hat.
I agree that most of the AI companies describe themselves and their products in hyperbolic terms. But that doesn't mean we need to counter that with equally absurd opposing hyperbole.
"Your code does not compile" and "Your tests fail"
If you have to tell an intern that more than once on a single task, there's going to be conversations.
If it costs them even just one more dollar than that revenue number to provide that service (spoiler, it does), then you could say AI has had no positive economic impact.
Considering we know they’re being subsidized by obscene amounts of investment money just like all other frontier model providers, it seems pretty clear it’s still a negative economic impact, regardless of the revenue number.
What HAS indisputably changed is the cost of hardware which has driven accessibility and caused more consumer facing software to be made.