keeda

A relevant article I found from an industry insider (which could indicate bias but also relevance): https://www.whitefiber.com/blog/understanding-gpu-lifecycle

Which has this anecdotal data point:

... when I left Paperspace in mid-2024, our M4000 GPUs, nine-year-old GPUs, were still consistently utilized at near-total capacity. That’s not a typo. Nine. Years. Old. Still booked, still working, still generating revenue.

Also, I won't claim to understand accounting, but in general it seems it is advantageous to accelerate depreciation schedules for high CapEx industries because they lower taxes: https://leyton.com/us/insights/articles/what-is-accelerated-...

As such you'd assume these cloud providers to want faster depreciation of their GPU assets rather than slower? I suppose in this case they do have an incentive to show bigger revenue numbers, but there seems to be a downside to this that is not being discussed?

Computer science is not a great example for this. I could ELI5 most of the terms you listed (and I actually have done so for many of them!) This is because it's pretty easy to map these concepts to everyday physical objects. Once a child understands any of those objects in their lives, it's pretty easy to explain in those terms.

Like, just the concept of "books" gets you very far. E.g. a file is a like a book, a folder is like a shelf to keep books, a stack is literally a stack of books, a heap is just a place you can pile books in willy-nilly, a database is like a library, a cache is books on your desk versus books in the library, replication is having multiple copies of a book so we can afford to lose some copies, indexing/sharding is like arranging books alphabetically, and so on.

Others are trickier but not much: a process is an app that is running on your device, a socket / tcp / http / websocks is a way to exchange information between devices, a namespace is how the name "Tom" in Tom Sawyer is different from "Tom" in Tom & Jerry, DNS is a way to get an address from a name, etc. etc.

You'll also notice that many of the terms you mentioned are already derived from well-known real-world concepts like pool, stream, channel, stack, queue, worker, transactions. You can mix those with other everyday concepts to make useful analogies.

But I could not even begin making analogies for most topics in Mathematics. I guess this is because advanced topics in Mathematics are just too abstract to map to everyday things.

I suspect the market harm angle is not going to work out either based on the one study I know of on the topic: https://www.nber.org/papers/w34777

We document a tripling in the number of new books coming to market between late 2022 and late 2025 that mirrors the use of AI that we detect in new books. The effects of this influx on consumer welfare depend on the quality of the additional books. The average quality of new books has fallen with the LLM-induced influx, and books with detected AI are substantially worse than human-authored books, so that much of the new work is of little value to consumers. Still, the LLM influx has delivered some books in the middle range of the usage/quality distribution, and the LLM-era entry process delivered seven percent more consumer surplus from books than the pre-LLM process in 2025.

...

Moreover, the arrival of LLMs does not appear to have displaced activity by incumbent authors. Despite the controversy surrounding LLMs, their effect on book consumers – like other cost-reducing technological changes in the cultural industries – is positive. However, because the new books are mostly of low quality, the effects are modest

So not only are existing authors unharmed (because most of the new competition is slop) there is even a small improvement for consumers.

That depends on the answer to a question that hasn't been answered yet.

It has been answered in a sense, because the courts (so far) have ruled that training is Fair Use. Whether this is similar to a human learning from a book was not quite the question being answered, but AFAICT there is no other relevant doctrine under Copyright law to address it, largely because the question didn't even exist until LLMs came along.

Also, these are not damages, it's a settlement i.e. a negotiated agreement between both parties.

Relevant sub-thread here: https://news.ycombinator.com/item?id=48997766

If you had some goal to achieve -- whatever it was, professional or personal -- would you rather expend X resources to achieve it or (X - Y), where Y was a significant % of X?

And if those resources had an environmental impact, and if you were truly concerned about the environment, why would you want to choose the option that expends more of those resources?

Think of it this way: if you had a doubt you wanted to resolve, would you drive to the library and spend hours or days scanning through reference material... or would you do a Google search?

That is essentially what the "environmental impact" debate comes down to, but nobody is talking about it that way. We can talk all day about the impact of AI on jobs and society, but the impact on the environment is a red herring, and a destructive counter-productive one at that, as TFA shows.

Ehh, Tomi Ahonen always came across as someone who was letting his emotions cloud his judgment (maybe the N9 was his pet project?) which was not great for a "consultant." Sure enough when I looked around there was substantial criticism to be found, e.g. https://dominiescommunicate.wordpress.com/2014/06/25/top-ten...

Also wasted spending is not quite the same as "not wanting to spend" -- it's more, to GP's point, "spending a lot unsuccessfully." I got the sense a lot of the friction Nokia and Windows Phone faced were due to Google (and to some extent Apple) using the market dominance of their properties (Android, YouTube, Search, Maps) to suppress competition.

I suppose it's fair play for what MSFT did in the OS and browser wars, but they got dinged pretty hard by Antitrust and played nice for a decade+ after that. Google is starting to see the antitrust blowback for it's actions only now, long after the competition has been crushed.

Or maybe "complimenting your commodities"?

The environmental impact discussion is always incomplete because it only looks at AI in isolation. LLMs do exactly nothing by themselves, they are always used to facilitate some human endeavor, so we need to take a holistic view of the system.

I had previously done some napkin math (https://news.ycombinator.com/item?id=46984659) which found LLMs + humans are way more efficient than humans alone, but that analysis is by definition simplistic, and as some point out, does not account for outliers like coding. Fortunately, I just came across this source: https://gist.github.com/mdodkins/9b49624855cc41570c9d1012e0d...

That looks accurate and lines up with my previous numbers. A heavy day of Claude Code usage is equivalent to running 0.7 of a dishwasher cycle. Already that sounds pretty decent, but let's put that into context and refine the simplistic analysis a bit. Discussing footprints in terms of [electricity / water / CO2] tuples, that is approximately:

1.3 kWH / 2.6L / 540g

Compare that to a human's daily footprint (US averages, scaled down to an 8-hour workday; the numbers are actually higher if you look only at office use):

10 kWH / 102L / 4KG

As a % of the human footprint, that is ~ 13% / 2% / 13%.

Right off the bat water usage is negligible. But at a minimum if LLM usage saves a human 13% time at a task, we're at breakeven i.e. the LLM resource consumption is compensated by the savings in human consumption. Anything more than 13% and we actually start conserving resources on average!

Now I know the productivity impact of LLMs is a contentious topic, but my past comments include studies (from 2024, in the era of spicy autocomplete, long before coding agents!) showing ~30% boost in throughput and/or time savings. Interestingly the survey-based https://www.genaiadoptiontracker.com/ finds a consistent 30% time savings across industries. A human completing a task 30% faster consumes ~30% less resources i.e. we save ~3.3 kWH / 34L / 1.3KG per day.

This gives us a net savings of 2 kWH / 31L / 760g per human per day!

Even with 2024 estimates of productivity numbers we are conserving resources! I would wager that a full day's use of Claude Code today offers a much, much higher boost.

You could say coding is a small part of software engineering, but then resource consumption scales down correspondingly. The upshot remains that for a given task overall resource consumption goes down with increased LLM usage! In fact, if LLMs were adopted across all knowledge workers to do all possible tasks, we could recoup the enormous environmental cost of training models in a few business days! Subsequent ongoing savings can more than compensate for a lot of the slop produced elsewhere.

So -- and tongue only partly in cheek -- if you really want to make an environmental impact, use LLMs more.

It doesn't even have to be monied interests. Media as a whole is getting severely disrupted by AI and they (somewhat understandably) see the technology as being built on top of their content without recompense. As such they will latch on to any topic -- supported or otherwise -- that lets them push a negative narrative.

On the flip side a lot of people's jobs are likely being threatened by the technology, so there is sizeable receptive audience already.

They don't need any more incentive or backing from monied interests, really. However I'm also pretty sure a lot of players are engaging in submarine warfare as well (https://paulgraham.com/submarine.html).

I did the napkin math (https://news.ycombinator.com/item?id=46984659) -- LLMs + humans are way more resource efficient than humans alone, even if they speed you up by only 5%.

It's not because LLMs are that efficient, it's because humans are that inefficient.

OSS does not necessarily mean the contributions are from "goodwill or part-time contributions". In fact, I would wager the most widely used OSS software is largely written by contributors paid to do so by corporations. At least for Linux, about 80% - 85% of contributions are from developers paid to do it (https://newsletter.pragmaticengineer.com/p/how-linux-is-buil...)

Corporations have had many reasons to invest their money in open source software -- custom requirements, marketing / developer mindshare, commoditizing complements -- but as cutting edge LLMs get more and more expensive to train, you'd be hard-pressed to find corporations who will put in that kind of money if they cannot recoup their investments.

Yeah, but the topic of discussion is presumably FSD that actually works ;-)

And before people jump in to say "have you tried it, it works" yes I have a Tesla and FSD has tried to kill me multiple times already. Including jumping a damn red light after slowing down to a stop so I was absolutely unprepared to react.

Credit where it's due, it's a pretty nice car and even the self-driving technology is decent, despite being hobbled by Elon's nonsensical allergy to LIDAR or other sensors. But it is being marketed as "FSD" and it is very much not FSD, and I can't believe more is not being done to prevent the needless deaths that are inevitable because of this.

I'd say it's more of a rent vs buy situation, in that there is a particular company-specific scale at which the TCO's cross over. For smaller companies, public clouds absolutely make more sense, but at larger scales on-prem is much cheaper (see e.g. BaseCamp's recent shift with numbers.)

The reason I say that a primary driver is a focus on core competencies because there are a surprisingly high number of companies with 10M+ public cloud spends... and even they sometimes complain about being treated as small potatoes. (For reference BaseCamp's spend was 3.2M/year when they decided to switch back.) Plus there are reports showing that much of these cloud deployments are at abysmally low levels of utilization (like 30% at best!)

So not only are there many companies at the scale where on-prem would make sense, they are likely even wasting a significant % of their cloud spend, yet they continue and even expand their usage.

Rather than attributing this to widespread corporate incompetency, a more rational explanation is that they have a sensible calculus that is based on more than solely financial numbers, i.e. the organizational distration is more expensive than the $$$ cloud premium.

Somewhat related, I once met the author of a book about sleep [1] and asked about my specific case. At that time I was splitting half my time across two continents 10 time zones apart but working largely in the same timezone, about 2 - 3 months at a time. That is, every 2 - 3 months I switched from a 1st shift sleep schedule to a "2.5th shift" one. So I asked him how bad that was for my health.

He said that as long as people are mostly regular in their sleep hours, the actual timings don't matter much. However, extended periods of irregular sleep schedules are actually (his words) "classified as a type of carcinogen."

I briefly looked into the evidence at the time but did not find it very concerning. TFA does make a more compelling case by linking it to all-cause mortality.

[1] I believe it was The Sleep Solution by M. Chris Winter, though I may be wrong. I can only remember it had a blue cover, but turns out pretty much ALL books about sleep have a blue cover. ¯\_(ツ)_/¯

The underlying MBA principle is that companies strongly prefer to focus on their "core competencies." This makes sense because the distraction to a business could be much costlier than the extra money spent in outsourcing non-core functions.

This is basically also why the cloud business was thriving even when on-prem is much cheaper in monetary terms, and why the trend will probably extend to cloud AI providers even when open weight models get better. (As an example, Linux is free, but MSFT makes a ton of money off it by renting out the hardware it runs on.)

That said, there will likely also be a very large volume of internal SaaS-y apps that enterprises will vibe-code simply because nothing on the market meets their needs and/or price point.

Another possibility is enterprises "remixing" existing SaaS apps for custom functionality by adding their own vibecoded layers on top of the SaaS endpoints. Either invoking official APIs where available, or by less official means like using custom browser extensions. Now that could lead to some interesting dynamics...

That's more the fate of consumer services though, where the product is typically given away for free and the need for revenues and growth eventually leads to the enshittification you describe. TFA is about SaaS products, which tend to be subscription-based and so usually are immune from those pressures.

SaaS products do have their own problems sometimes, such as feature creep and bloat and uptime, but those are less insidious.

This study finds increased sales and value per customer from GenAI integration at a large Chinese online retailer (all the way back in 2023-24!) The customer Q&A scenario is one of those covered, except the customer talks directly to the LLM rather than an employee with a subscription:

https://arxiv.org/abs/2510.12049

We quantify the short-term impact of Generative Artificial Intelligence (GenAI) on sales performance through a series of large-scale randomized field experiments involving millions of users and products at a leading cross-border online retail platform. Over 2023-2024, the platform integrated GenAI into seven business workflows spanning customer service, consumer-product matching, advertising, and seller services. We find that GenAI adoption increases sales in most workflows, with effects ranging from no detectable impact to 16.3%, depending on GenAI's marginal contribution relative to baseline firm practices. Across the four GenAI applications with positive sales effects, the implied annual incremental value is roughly $5 per consumer−an economically meaningful impact given the retailer's scale and the early stage of GenAI adoption. The gains operate primarily through higher conversion rates rather than larger cart values, consistent with GenAI improving the shopping experience by reducing search, information, communication, and personalization frictions. Importantly, these effects are not associated with worse post-purchase outcomes, as product return rates and customer ratings do not deteriorate. Finally, we document substantial demand-side heterogeneity, with larger gains for less experienced consumers. Our findings provide novel, large-scale causal evidence on how GenAI shapes sales productivity in online retail, highlighting both its immediate value and broader potential.

Impact on profitability itself is hard to determine due to caveats listed in the paper (which are important to read!) but offhand I would guess that incemental $5 margin per customer is much more than what their prompts cost.

Glad to see he's recovered from his little flesh wound.

I think I can put the finger on how it differs and why the AI version seems more "generic." I think the difference is in the "post production" that became more common in pop music in the 2010's. I'm not at all a musician so I don't know the right terms for it, but the AI version has a bunch more of little "flairs" and auto-tuning and audio tweaks which I assume is put in during post-production that makes it sound "slick."

However, it seems to me that all pop music uses the same post-production tricks these days, and so it all sounds somewhat formulaic even if the individual songs themselves are very different. As such the original version may sound more interesting simply by being different.

So the 2001 version sounds more like an "MTV Unplugged" performance whereas the AI version sounds more like the professional and polished version that gets released commercially.

Each has their allure, however. I suspect the AI version will do better with younger crowds.

Alternate viewpoint: the ideas and insights in the content are more important than the voice. The voice absolutely is critical in getting your ideas across and making a point effectively, but there is value in having fresh ideas being broadcast into the world.

To the extent that LLMs help this where otherwise people would simply have not spoken up, I think it's OK. Of course, the slopisms are an instant turn-off and limit the reach of those ideas, but at least they're out there now.

Personally I've developed a knack of quickly skimming through the bland language to get a sense of whether there is anything interesting enough to re-read more closely. It's become so ingrained that I don't mind wading through all the noise from all the other lazy, content-free slop to get to that little bit of signal.

If I'm working on something open source I actually would not mind as much. (I can understand why many folks don't like the models being trained with impunity on ALL the open-source code out there, but personally I think it's fine for my open-source code.)

However, I am working on some projects where I think I've stumbled on genuinely interesting and valuable technical approaches, and I'm still figuring out how to capitalize on them. As such I am a bit paranoid about my ideas leaking via some training dataset.

Heck, it could end up as nothing more than a blog post that nobody reads, but at least I get to publish it.

Training included trillions of tokens of Cursor data which capture a wide-range of user interactions with codebases and software tools.

This -- training on work done on hard, real-world tasks -- seems to be how most frontier models are making capability gains these days. In fact people make decent money doing that for data companies like Mercor. However it's also striking that Cursor managed to gather so much of such data.

Turns out Cursor will train on everything you do unless you opt-out, even if you're already paying for it with cash! Are that many people really not opting out?

This is why it seems like a significant concern to me: It's very clear that typical, run-of-the-mill coding has been completely commoditized, so the primary value remaining is either in novel use-cases and applications, or novel technical solutions to hard problems.

Presumably the value for novel use-cases could be captured by building a business around it via the usual moats (distribution, relationships, network effects, first mover advantage, etc.) so the code and techniques do not matter as much.

However novel technical solutions, which are already hard to monetize without building a whole damn business around it, could at least be capitalized on by simply being able to claim credit for it. I'd at least like the option of being "paid in exposure" if I'm not getting paid in cash. But having them "leaked" unwittingly via the training corpus to whosoever happens to prompt the model with the same problem removes even that option.

I know people have been calling out this risk forever, and I don't use any tool that I can't opt-out of training completely, but the scale at which this is happening -- on an ongoing basis, mind you, after training on the data of the whole world, and that too after paying for the product -- is surprising. I'm bullish on the technology but we really should be way more careful handing these AI companies even more of our intellectual crown jewels.

I used to think Claude Code was released much earlier too, and my initial theory was that OpenAI as a follower had the benefit of more powerful models... but when I looked it up, Codex was first released in April 2025 [1], whereas CC Beta was released in Feb 2025 [2] -- only a couple of months apart!

I'm sure each lab is keenly aware of what the other is doing (how else could they time so many of their releases so close to each other?) so it's highly probable they started developing each app about the same time, and likely even knew the technical details involved. Which is why it's additionally interesting that OpenAI started off with Rust while CC used Electron.

1. https://techcrunch.com/2025/04/16/openai-debuts-codex-cli-an...

2. https://github.com/jqueryscript/anthropic-claude-timeline

That analogy only works partially, because when IE6 was released, it was the best browser by far. IE only became terrible once MSFT actively stopped developing it, and other browsers kept getting better.

On the other hand, Claude Code was the best coding agent when it was released, but there's no way Anthropic is going to let its cash cow stagnate. Like, I think pretty much all of Anthropic's revenue spike in the last few months was driven by the tokenmaxxing mania.

My take is most of Claude Code's problems originate from insufficient compute capacity and all kinds of workarounds they're doing to mitigate that fundamental limitation.

Interestingly I just skimmed through a video linked in TFA: https://www.youtube.com/watch?v=SlGRN8jh2RI and Boris Cherny explains that they chose TypeScript and React for Claude Code primarily because it was "on distribution" and the models back then just weren't good enough at other languages.

So it's interesting that Codex is written in Rust. Amongst other things it could mean OpenAI had more powerful models that could handle Rust, or their engineers had to handhold the agents a lot more up front, or Rust has structural advantages that could overcome being less represented in the training data.

... Dario bizarrely copying Altman’s 2023 fire-and-brimstone playbook that had already massively backfired.

I've said this before, they always knew it was terrible marketing, but they just can't help themselves because they actually believe it.

From multiple accounts, the people working at these labs, who are most exposed to the latest models' capabilities and how they're being used out in the world, are simultaneously excited and terrified about what they're building.

In a way that's even scarier than the "Capitalist sociopaths marketing AI to other Capitalist sociopaths" rationale everybody assumes.

I last looked into this a decade+ ago so my memory is fuzzy, and there are a lot of economists who looked at a lot of different things, but the authors I recall from the time were Kenneth Sokoloff, Petra Moser, Adam Mossoff, Zorina Khan, Bhaven Sampat, and Bronwyn Hall. Undoubtedly there are dozens more, but I just happen to remember these offhand.

This book gets a lot of airtime in discussions of IP but the authors have a narrative they are trying to push and they don't let inconvenient things like facts or history get in the way.

The book cherry-picks its sources, and even then contains several mischaracterizations and exaggerations of those works. There are many other economists who have shown significant beneficial aspects of patents with empirical data but they conveniently don’t get mentioned at all.

As an example, see this: https://www.researchgate.net/publication/46556404_Watt_Again...

Yep, the very first chapter of the book starts with patents and steam power, and they got called out on it by actual experts on the topic. Note the “still” in the title – this is after the book was already “revised” once. The book was not revised after this last note, so the exaggerations still stand.

The rest of the book had many similar issues. Once I started digging into their sources, I could not get past the first few chapters, but anecdotally others on HN have also pointed out glaring inaccuracies. I remember thinking wryly that it should be called “Against Intellectual Honesty.”

The original study itself had at least one developer who later revealed that he had filtered out tasks he prefered not to do without AI: https://xcancel.com/ruben_bloom/status/1943536052037390531 -- given the N was 16, and he seems to have been one of the more AI-experienced devs, and we don't know if the other devs did this, the results of the first study itself could be questioned.

Agreed, there's nothing sticky about the models right now, but I see the big AI players making moves that hint at long-term success, and even dominance, in the enterprise space, which is where the real money is. As Microsoft has shown you can create stickiness for things that are already heavily commoditized. For instance, the FDE play itself could be a huge business.

Plus the big factor in my mind for why these frontier labs will succeed -- and this is very fuzzy and hand-wavy -- is that they are very business- and government-savvy and execute extremely fast. They have the most powerful AI models at their fingertips with sharp people who know how to use them, along with insane levels of funding, and are showing the world how a truly AI-empowered organization can operate. I suspect they will thrive despite all the forces arrayed against them.