abathologist

> Everyone else is getting internal productivity lifts that are matched by competition

The main metric analyzed in the note is

> Revenue per worker

This can increase regardless of competition, but I guess would require either expansion into new domains or layoffs. So maybe we will see this raise as a lagging indicator if the LLMs are actually delivering the goods.

> when they are deploying inference features they are usually ones their customers didn’t ask for and don’t value, OR that they expect everyone to have and are not a differentiator

I am definitely seeing lots of this happening. Tons of energy going into side quests that aren't critical for anything. Meanwhile, strangely, the central features we actually need to ship, but that require deep redesigns and careful refactors in legacy code, are not moving quickly at all.

In case you are not being sarcastic, layoffs obviously don't "cause" productivity increase. [Edit: but they would help the main metric (Revenue per worker) analyzed in the note. so perhaps that is what we will see.]

But actually 40% of the S&P 500 are tech companies, and they are saying only the mag 7 are showing clear revenue increase. Would would suggest that even in most tech we are not seeming measurable revenue growth, right?

One plausible explanations that has occurred to me is this: While programmers etc. can produce way more code with LLMs, the declines in quality and coherence of the work, and a widening gap with a coherent strategy and aim for value means the increased activity is mostly not valuable. I suspect this may just be an intensification of https://en.wikipedia.org/wiki/Productivity_paradox#Little_pr...

I'm seeing estimates of a %95-91 failure rate on attempts in generally (or, conversely, a max of ~10% success, but likely lower, since we can only go based on how many attempts they are willing to mention in their statements). E.g., from analysis in https://medium.com/@mehmet.ozel2701/openai-tried-about-4-000...

I wonder how many of those founder at the "paper" to formalization step, or at earlier or later steps (e.g., proof doesn't turn out to prove what the paper said, hits a bug in the kernel, etc.).

This kind of seems to me like exactly what a reasonable person would expect, based on the nature of the technology and the dynamics of the market bubble: getting plausible-looking-text-generators to do text munging on web apps and servers doesn't seem likely to increase base productivity, but the companies who can collect on the hype will make a killing.

Look at how many tech or tech-adjacent companies are in the S&P: https://en.wikipedia.org/wiki/List_of_S&P_500_companies

Like %40 of the companies in the index are tech or IT. SO how does the supposed flat line across all those tech companies fit your thesis?

> displacing the advantages of unearned social or financial power

In what way? Sure looks like pay to win to me.

I've worked in multiple nix-heavy environments, and I've come to really dislike it, despite thinking that Dolstra's thesis and core ideas are an important development in the space of package management. Everywhere it shows up it infects projects, creates constant breakage, and requires at least 1 person to become a nix expert, memorizing a bunch of esoteric nonsense. Basically it forces all of the adhoc nonsense of unix systems to be right in your face constantly, and leads to perpetual dependency limbo.

Totally on point, except that people don't get to understanding and configuring, they just probably correct configs that they don't understand.

This is exactly what is bad about LLM-driven AI: it is perfect for papering over the deep problems with unwieldy messes like nix. In a rational world, the pain and fragility of nix would keep it from getting mass adoption, until a deeply revised, and much more refined radicalization of purely functional software delivery appeared that had the upside without all the warts and inscrutable complexity. Instead, we'll just get a metastasized growth of incomprehensible nix configs, and be stuck in that local optimum for ever.

LLMs make worse is better even better, for the worse.

Depends on the company, obviously.

Well... look at the behavior, it is bad. They use the things you say here to justify it, soo... yes.

Read moar https://en.wikipedia.org/wiki/Effective_altruism#Other_criti...

Who wrote this? I haven't been able to find an author yet, and the supposed source repo (https://github.com/quantitative-finance-with-ocaml as per https://qcaml.com/appendices/appendix-e-setup/ ) is a 404.

In my read, the people trying to "make AI do something" are also slotted in the lost/distracted group in the comment. They are also addicted, and at best just following a profit motive (which is also just a stimulus response programming).

The last alternative, to think if you still can, is not tied to AI at all (which is not to say it can't make some use of or explore it).

The definition I use is shipping LLM generated code you don't understand

We use LLMs extensively on the projects I work in. We don't "vibe code", and we understand every commit.

Yet another case of https://en.wikipedia.org/wiki/Neijuan

https://sel4.systems/ is relevant in this space!

How come all the vibe coded stuff I've tried is totally bug ridden and unmaintainable then?

I predict a different outcome: the rate of vulns identified and fixed will be more than matched by the rate of new vulns introduced by irresponsible use of LLMs on top of brittle and unwieldy tech stacks.

The result will be an overall increase in turbulence and the normalization of steadily intensifying security crises in nearly all software systems.

The only projects that will escape this fate are those which have either been already developed from ground up with rigorous and principled, verified (or verifiable) design, or those which are rewritten to gain this.

Why would this be good for anything but scams and disorder? What is the actually prosocial use case?

No it isn't.

People are just paying attention to what is happening the world.

People are also being trained/acculturated to LLM speak, so even if they are trained on human output, they could be getting reinforcement for their LLM-tinged crap-speak.

I appreciate this line of questioning. It's really interesting to see how many excuses people need to reach for to avoid the conclusion that the "secret unheard of power" is B.S...

> We are going from the era of manual, line-by-line mental model transcription to one where software engineers can focus on data structures, software architecture and algorithms.

This is what declarative programming gives us, not what LLM-based generation offers.

> I love being able to quickly bring out the program that is already running in my head without having to worry about the grind of typing it into a format that the compiler understands.

Using natural instead of a formal language to get probabilistic results based on token fields is not bypassing arbitrary constraints of the compiler, it is dereliction of the responsibility to know and articulate precisely what you are specifying.

During the crypto bubble, it was common for "crypto insiders" to talk about how it was poised to transform the entire economy, on the cusp of making all finance distributed, about to revolutionize ownership, etc., etc. .

The people in the most inflated parts of bubbles don't tend to have the most clear eyed assessment of the real impact and potential of the dynamics contributing to the bubble: their perception is warped by the bubble, and they cannot help but see everything filtered thru it.

Yes. Maybe much more:

> Such intensive use of AI doesn't come cheap. In a post on X, LisanBench, an LLM benchmark evaluator, estimated that the output tokens alone would cost about $6.5 million at OpenAI's average consumer price. Including the far larger volume of input tokens, the post estimated the total could reach $10 million to $40 million.

https://www.businessinsider.com/openai-math-problem-solved-t...

> The retail price is not the cost.

True. The cost is probably much higher, since they are still subsidizing as part of the first phase of the enshitification playbook.

This could fund 10 top income mathematicians for 8 years (based on https://careers.usnews.com/best-jobs/mathematician/salary ). Imagine what kinds of results we'd have to transform the foundations of science if we were giving brilliant minds this kind of funding to do nothing but research for most of decade....

Instead, we get slop proofs that are technically correct as PR stunts to enable corrupt kleptocrats, and most likely will drive research into culs-de-sac.