Specification languages need big investments essentially - both in technical and educational terms.
Consider something like TLA+. How can we make things such as that - be useful in an LLM orchestration framework, be human friendly - that'd be the question I ask.
So the developer will verify just the spec, and let the LLM match against it in a tougher way than it is possible to do now.
I'd think there'll be a dip in code quality (compared to human) initially due to "AI machinery" due to its immaturity. But over-time on a mass-scale - we are going to see an improvement in the quality of software artifacts.
It is easier to 'discipline' the top 5 AI agents in the planet - rather than try to get a million distributed devs ("artisans") to produce high quality results.
It's like in the clothing or manufacturing industry I think. Artisans were able to produce better individual results than the average industry machinery, at least initially. But overtime - industry machinery could match the average artisan or even beat the average, while decisively beating in scale, speed, energy efficiency and so on.
The issue is that code isn't clothing. It's the clothing factory. We aren't artisans sewing clothing. We're production engineers deciding on layouts for robots to make clothes most efficiently.
I see this type error of thinking all the time. Engineers don't make objects of type A, we make functions of type A -> B or higher order.
Go concrete. In FAANG engineering jobs now what % is this factory designer category vs what % is writing some mundane glue code, moving data around in CRUD calls, or putting in a monitoring metric etc?
Once you look at the present engineering org compositions see what's the error in thinking.
There are other analogy issues in your response which I won't nitpick
> industry machinery could match the average artisan or even beat the average
Whether it could is distinct from whether it will. I'm sure you've noticed the decline in the quality of clothing. Markets a mercurial and subject to manipulation through hype (fast fashion is just a marketing scheme to generate revenue, but people bought into the lie).
With code, you have a complicating factor, namely, that LLMs are now consuming their own shit. As LLM use increases, the percentage of code that is generated vs. written by people will increase. That risks creating an echo chamber of sorts.
I don't agree with the limited point about fast fashion/enthittification, etc.
Quick check: Do you want to go back to pre-industrial era then - when according to you, you had better options for clothing?
Personally, I wouldn't want that - because I believe as a customer, I am better served now (cost/benefit wise) than then.
As to the point about recursive quality decline - I don't take it seriously, I believe in human ingenuity, and believe humans will overcome these obstacles and over time deliver higher quality results at bigger scale/lower costs/faster time cycles.
> Quick check: Do you want to go back to pre-industrial era then - when according to you, you had better options for clothing?
This does not follow. Fast fashion as described is historically recent. An an example, I have a cheap t-shift from the mid-90s that is in excellent condition after three decades of use. Now, I buy a t-shirt in the same price range, and it begins to fall apart in less than a year. This decline in the quality of clothing is well known and documented, and it is incredibly wasteful.
The point is that this development is the product of consumerist cultural presuppositions that construct a particular valuation that encourages such behavior, especially one that fetishizes novelty for its own sake. In the absence of such a valuation, industry would take a different direction and behave differently. Companies, of course, promote fast fashion, because it means higher sales.
Things are not guaranteed to become better. This is the fallacy of progress, the notion that the state of the world at t+1 must be better than it was at t. At the very least, it demands an account of what constitutes "better".
> I don't take it seriously, I believe in human ingenuity, and believe humans will overcome these obstacles
That's great, but that's not an argument, only a sentiment.
I also didn't say we'll experience necessarily a decline, only that LLMs are now trained on data produced by human beings. That means the substance and content is entirely derived from patterns produced by us, hence the appearance of intelligence in the results it produces. LLMs merely operate over statistical distributions in that data. If LLMs reduce the amount of content made by human beings, then training on the generated data is circular. "Ingenuity" cannot squeeze blood out of a stone. Something cannot come from nothing. I didn't say there can't be this something, but there does need to be a something from which an LLM or whatever can benefit.
> it is easier to 'discipline' the top 5 AI agents in the planet - rather than try to get a million distributed devs ("artisans") to produce high quality results.
Your take essentially is "let's live in a shoe box, packaging pipelines produce them cheaply en masse, who needs slow poke construction engineers and architects anymore"
Where have I said engineers/architects aren't necessary? My point is that it is easier to get AI to get better than try to improve a million developers. Isn't that a straightforward point?
What the role of an engineer in the new context - I am not speculating on.
> My point is that it is easier to get AI to get better than try to improve a million developers.
No it's not, your whole premise is invalid both in terms of financing the effort and in the AI's ability to improve beyond RNG+parroting. The AI code agents produce shoe boxes, your claim is that they can be improved to produce buildings instead. It won't happen, not until you get rid of the "temperature" (newspeak for RNG) and replace it with conceptual cognition.
Artisanal clothing is functionally equivalent to mass-produced clothing, but more expensive.
Much of contemporary software is functionally equivalent but more expensive to run and produce than previous generations. Chat, project management, document editing, online stores… all seem to have gotten more expensive to produce and run with little to no gain in functionality.
Complexity in software production and tooling keeps increasing yet functionally software is more or less the same as 20 years ago (obv. excluding advancements depending on hardware like video, 3D rendering, LLMs, etc.
Except I am not talking about clothing. You are guessing when you say "I'd think" based on your comparison to manufacturing clothing. Why guess and compare when you have more context than that? You're in this industry, right? The commodity of clothing is not like the commodity of software at all. Almost nothing is, as it doesn't really have a physical form. That impacts the economics significantly.
To highlight the gaps in your analogy; machinery still fails to match artisan clothing-makers. Despite being relatively fit, I've got wide hips. I cannot buy denim jeans that both; fit my legs, _and_ my waist. I either roll the legs up or have them hemmed. I am not all that odd, either. One size cannot fit all.
In his view - most ML algos are at level 1 - they look at data and draw associations, and "agents" have started some steps in level 2 - doing.
The smartest of humans operate mostly in level (3) of abstractions - where they see things, gain experience, and later build up a "strong causal model" of the world and become capable of answering "what if" questions.
Leslie Lamport built latex, most of distributed systems such as AWS services depend on formal verification. The job of Science here is to help Engineering with managing complexity and scale. The researchers are doing their jobs
What does LaTeX have to do with TLA+? Also I think "most of distributed systems such as AWS" might be an exaggeration. At least the public known examples of formal verification in AWS are scarce.
Willful ignorance is a different process. Consider a food analogy.
Of the food we take - cells accept a % of it as nutrients and such, rest is discarded as waste. The cells know how to get this job done - it's a very complex process for sure.
I think it's the same with information content - a % actually is useful for making life happen - whereas the rest should ideally be discarded because it is meaningless from a life perspective. The mind just knows what's important most of the time.
In this case - willful ignorance would be something like intermittent fasting or regulating food intake carefully, since it is a conscious process.
The former process is unconscious and operates at the "cell level" whereas the latter is a conscious process that operates at the "whole-being" level.
Book sales in general (across all formats) are up I think - so there are still many, many readers around. We just have many new formats (EPUB, audiobooks, reader devices, etc.) and of course population is increasing over the globe. I'm pretty sure we have the highest number of readers on the planet right now than ever before in absolute terms.
I'm not sure that's still correct. There was an uplift because of Covid and people having more spare time, but whatever more recent (2024 - 2025) sources I can find suggest the trend has reversed.
It's worth also considering demographics. If you narrow the focus to just younger generations (who, we can guess, are more addicted to smartphones) then the numbers look pretty bad. E.g.:
My son, who is away at college as a freshman this year, recently phoned me and apologized for calling me a bad dad and thanked me for not allowing him to have any devices in his bedroom after bedtime growing up, as it made him become a reader. He said he was amazed when he got to school and nobody else reads for pleasure.
I can't fault people for feeling that nobody reads anymore. In the US today the majority of Americans can't even understand books written at a 6th grade level and literacy has been trending downward. Only a small number of us are propping up book sales.
Audience matters here. Most book sales have been falling. The one increase has been in romance porn with those books accounting for some 50% of all paperbacks sold at this point (they are dominating for the exact same reason porn dominates internet video content).
Personally, I don't count pornhub traffic the same way I count Youtube or Netflix traffic and I think the same applies here.
Consider something like TLA+. How can we make things such as that - be useful in an LLM orchestration framework, be human friendly - that'd be the question I ask.
So the developer will verify just the spec, and let the LLM match against it in a tougher way than it is possible to do now.
reply