Pencils Down, Eyes Open: A Rails Developer After Rails World
8 points by Caio
8 points by Caio
I hate the programming sector so much right now. I remember the hype around scrum, where the message was "you have to adopt scrum, it makes you super productive!" "but you also need a scrum master who makes sure you do it right" "if it doesnt work you doing it wrong".
And now with "AI", when i read "handwriting code is already uneconomical for most programmers", its the same thing in a different color all over again. Adopt it or get left behind, which means if your resume doesnt include "Did work with agents" you will have a tough time to get hired.
It almost feels like programmers dug their on graves with AI. Managments wet dream, having a bare minimum of programmers working 10x faster with agents promised by AI companies.
It almost feels like programmers dug their on graves with AI. Managments wet dream, having a bare minimum of programmers working 10x faster with agents promised by AI companies.
It's a big con game by OpenAI and the ex-OpenAI people who left to start Anthropic and, like any con game, it works because it's telling executives what they want to hear so badly. These managerial-class people want so badly to just do nothing and become rich for it, and want so badly to not have to deal with the complexities of human interaction.
There's a reason David Gerard used the name Pivot to AI for his equivalent to Molly White's Web3 is Going Just Great... it's because, when NFTs crashed, he saw the same 5%-ers trying to pry back open the door to becoming 1%-ers just pivot from NFTs to GenAI... just as they'd pivoted from cyptocurrencies being the future of payment processing to NFTs.
That's also why they pivoted from "AI will replace all your jobs" to "AI is the ultimate coder" (when that plateaued) and now that AI coding is plateauing, they've pivoted to AI for cybersecurity by committing computer crimes that the U.S. government won't prosecute them for while simultaneously pleading to be regulated so they won't "end the world" so they'll have a justification for when "line stops going up" because they're running out of sources of new venture capital to set on fire in their "charge subscribers $1 per $5-20 of resource spend in order to remain competitive" machines and companies are starting to actually accumulate the data to realize that LLMs cost more than coders for equivalent work and it's better to just use LLMs to justify a "lay off and quietly re-hire with lower pay" cycle.
The only sin programmers not in those circles committed was being too "smart but not wise" to remain silent instead of impulsively saying "I think I see the problem" when their turn came to stick their neck in the guillotine which "God" was preventing from killing people.
...and, as someone who has to be careful how much time he spends privately goofing around with a local copy of Stable Diffusion, I don't blame them. Generative AI is not only designed to demo well but have a low potential for growth (a white van speaker scam for executives who AI probably could replace), it's also built on gambling addiction.
TL;DR: The only sin programmers committed was falling into the same excitement for new technology that we've been falling into since time immemorial... just this time, it was tailor-made to glom onto the intrinsic laziness inherent in human neural architecture. (telegraphy networks and the Marconi company being staffed by a subculture of enthusiastic tech nerds, radio enthusiasts paying top-dollar for just one of the newfangled transistors, IMSAI Altair, TV Typewriter, etc. etc. etc.) This is all "capitalism is eating itself desperately trying to find new sources of growth in a post-Moore's Law world on a finite Earth where all the low-hanging Windows 95s and Internets and iPhones have already been invented".
Hell, look into Gamers Nexus's recent coverage of ACR, including the new Microsoft patent. Peak example of companies desperate to keep line going up at any cost.
(TL;DR: The EULA you agreed to on your smart TV allows it to spy on you with cameras and microphones for advertising purposes and use Bluetooth with other devices to silently sandbox-break. It's already implemented with smart TVs and Microsoft just patented doing the same thing with games (or any other application) for the purpose of making sure it only holds your game hostage behind a "watch an ad to get more credits" mechanism when it wouldn't interrupt something that can't be paused like an active online play session.)
I think you're taking the opposite extreme view but missing good supporting facts.
and now that AI coding is plateauing
OpenAI is releasing a $500/mth plan that a number of people I know are going to jump on immediately so they don't have to juggle multiple $200/mth subscription. Neither the improvements, nor the use of AI coding is plateauing yet.
and companies are starting to actually accumulate the data to realize that LLMs cost more than coders for equivalent work
You can look at this in various ways, so anyone trying to generalise this too hard will run into problems... but yeah, unless you do stupid things or try to force it where not necessary, it's just true. Just this week I did 3 side projects which cost under $5 in total, had positive impact on revenue, and were entirely in the "it would take weeks and we can't spend that much dev time on this" territory a few years ago. And I don't believe I'm some weird exception here.
The reality is likely somewhere in the middle. There's oversaturation of weird AI people teaching their ideas and news on YT and I'm sure they'll find their way to enterprise environments too. But that's going to happen to every popular trend - we'll see extremes that scale with how big an idea is. And it would be surprising is we didn't see AI scammers.
OpenAI is releasing a $500/mth plan that a number of people I know are going to jump on immediately so they don't have to juggle multiple $200/mth subscription. Neither the improvements, nor the use of AI coding is plateauing yet.
My bad. Since their projected profit for AI coding is plateauing relative to what they need it to be, given their expenses and contractual obligations. (Seriously. Take a look at some of Ed Zitron's investigation into their financial situation.)
(Though I do also believe that the progress on LLMs themselves is also plateauing, forcing a shift to pushing further innovation into the harnesses.)
You can look at this in various ways, so anyone trying to generalise this too hard will run into problems... but yeah, unless you do stupid things or try to force it where not necessary, it's just true. Just this week I did 3 side projects which cost under $5 in total, had positive impact on revenue, and were entirely in the "it would take weeks and we can't spend that much dev time on this" territory a few years ago. And I don't believe I'm some weird exception here.
It's the "for equivalent work" part. Companies were sold a story that they could replace coders with it, and now they're starting to get the data to determine that, yes, you do still need skilled humans to keep them on track and, yes, companies like Anthropic are moving agents off their loss-leader plans and onto APIs with metered billing and, yes, when you actually look at the effects on the balance sheet, it's getting harder to make a compelling case that it's saving or creating value greater than it's costing... maybe once the A.I. bubble bursts and the hardware for on-site/on-prem models becomes more affordable.
Ed Zitron has significant bias and is selective with facts. Financial issues are there, but I would not refer to Ed as a verified source. Dan Luu did the last write-up we'll need on that topic: https://danluu.com/zitron/
I don't get the equivalent work note. If you can get the same work done by decreasing work hours and increasing LLM spend by a smaller amount, it holds true. You don't have to replace whole people - someone needing to drive the process doesn't invalidate the idea.
Luu's analysis was sufficiently biased and cherry-picked as to completely undermine all of his other writing; he burnt all of his reputation attacking Zitron. For example, Luu completely ignores any discussion of Oracle and Stargate; previously, on Awful, some year-and-a-half ago, I reused Zitron's logic to predict that Stargate would run out of true believers before it could build enough capital to become a plausible venture, and a year later, on Awful, we can see how well that prediction held up. For another example, Luu's response to the idea that GPT-based products haven't improved since 2023 is to claim that confabulation rates are empirically lower, which is an out-of-touch approach that doesn't talk about the technology as it actually exists; previously, on Lobsters, I pointed out that there are concrete techniques (PEFT, soft prompts, steering vectors, grammars) which have emerged in the three-year period under discussion but not yet manifested in popular products.
I don't take Zitron's advice for my finances, but I'm not going to be able to take Luu's advice on anything ever again. The whole mild-mannered-engineer persona, once revealed as a deception, is not trustworthy.
Luu completely ignores any discussion of Oracle and Stargate
A broken clock can be right twice a day, does that mean you trust a broken clock? Ed has been substantially wrong on the vast majority of his predictions about LLMs, you can't point to the minority and say "well he's been right sometimes, that rescues him".
For another example, Luu's response to the idea that GPT-based products haven't improved since 2023 is to claim that confabulation rates are empirically lower
What? I can't see anything like that in his post. This is what he says about models improving:
If someone's actually looked at what's happening, I don't think there's anything you can really do to convince someone who's denying reality at that level, but for anyone who's just hasn't seen how things have changed, for a visual example of improvements over that time period, here's a comparison of 2023 and 2025 video generation and here's an example from August 2026. Video isn't a great example since models have improved a lot more at coding, e.g., with on the order of minutes of human time, it's possible to create a new regex engine with an interpreter and an native code compiler and then fork ripgrep to make it faster for codex's actual ripgrep calls on my machine, a project that would probably cost 7 figures pre-LLM if you price out how much people with the expertise for that are paid.
The closes I can get by searching for "hallucinations" is:
Zitron's argument at the time was that hallucinations were as good as they were going to get, which meant that AI performance is capped at 2024 levels. Both the overall prediction and the mechanism were wrong.
... Which is not what you're saying at all? Are you hallucinating?
Ed Zitron has significant bias and is selective with facts. Financial issues are there, but I would not refer to Ed as a verified source. Dan Luu did the last write-up we'll need on that topic: https://danluu.com/zitron/
Fair point... though I'd call a lot of those points more sloppy and biased in the aggregation than inherently biased. Things like "none of these companies anymore really know how to grow ... in the desperation to try to reignite growth in a dying ecosystem the tech industry is going to shove this [AI] shit into everything" are perfectly consistent with stuff like how Windows 11 is artificially forcing a hardware replacement cycle despite Microsoft having marketed Windows 10's rolling-release nature as the last Windows there ever would be, how they're finding more and more places to shove ads, and how most of the things most people do with their computers can still be done perfectly comfortably on a Windows 7 machine from 2011 and a lot of such machines are still alive and kicking in the post-capacitor-plague era. I wouldn't call the ecosystem "dying", but I would say that these companies really don't want to accept that they're running out of novel "things people are willing to buy" to be discovered within the current economic context. It's the same kind of motivation that birthed Windows 8 and Windows Phone in response to a perception that desktop/laptop sales were plateauing while mobile and tablet sales continued to rise.
Hell, aside from goofing around with Stable Diffusion, running WhisperX to transcribe conversations, Core Keeper, and improving my Rust build times, everything I do CAN be done on a couple of older machines I have (my old Athlon II X2 270 from 2012 with a GTX750 and RAM maxed out to 32GiB and an Intel hand-me-down from 2012 with Core i7-2600S and a GTX670 that's stuck with only 8GiB of RAM which I use as a "game console") and I game a lot on that GTX670... just not the AAA stuff that's some mix of pushing the boundaries of photorealism and not having enough time to optimize before the deadline.
(Core Keeper's world simulation is a little too demanding for even the Core i7-2600S that I splurged and spent $55 CAD on to order from China to upgrade that machine from a Core i3-2120 that already ran basically everything perfectly, and I'd probably need a $110 CAD GTX980 instead of the $35 CAD GTX670 to consolidate both machines into one since, IIRC, Maxwell is the first generation that can drive three DVI monitors without an active DisplayPort-to-DVI hub and the GTX750 isn't exactly a gaming powerhouse.)
EDIT: Also, I was thinking more about how datacenter buildouts seem to be overly optimistic about demand going forward, and estimates of how much OpenAI, Anthropic, and xAI (companies without the kind of fallback revenue sources that Microsoft, Amazon, and Meta have) are spending as a proportion of the demand are worryingly high.
I don't get the equivalent work note. If you can get the same work done by decreasing work hours and increasing LLM spend by a smaller amount, it holds true. You don't have to replace whole people - someone needing to drive the process doesn't invalidate the idea.
My point is that the initial scam was "you can replace people" and now things like the switch to token-based billing for APIs are piling on additional reasons to make it harder to justify the token spend.
(Though I do also believe that the progress on LLMs themselves is also plateauing, forcing a shift to pushing further innovation into the harnesses.)
You are wrong. LLMs have gotten materially better in the past few months. Harnesses are doing fewer things now than they used to. You can use pi, a harness that is deliberately minimal, with the newer models and get better results.
I'll take your word for it since I haven't personally experienced a noticeable difference in the ones I interact with yet.
I think it’s much more likely that LLMs are legitimately powerful and useful. People really aren’t that dumb. It’s pretty infantilizing to think that millions of people are simply being conned.
Are there people trying to maximize the value out of overhyping the technology? Yea definitely. But that doesn’t make it a con.
I never said they weren't powerful and useful.
Fundamentally, the con is the claims about how big an industry they'll form within any relevant-to-investors-right-now time-frame, and downstream claims about fitness for purpose. The con is things like claiming that things like Artificial Superintelligence and Recursive Self-Improvement are right around the corner. Altman and Amodei are shysters.
Reading the "quick review exercise" at the end of the post, I have a hard time shaking off the thought that the entire thing is written by a clueless high school graduate who would post this on linkedin to get back pats from their peers.
"A Rails Developer"? This isn't even a person. I would prefer to know what an actual human thinks about the subject, rather than this machine-extruded content. Even the "author" headshot on the homepage is slop.
Sir, I’m a person. And I’ve written this article. You should focus on the content of the article instead of scanning it for tells. It just makes you look petty and pedantic.
FWIW Pangram agrees with me: "100% of this text is AI". Did you write it, or did you "write" it?
That’s a radical choice for one company, not an instruction that every developer has to follow tomorrow.
That’s his economic bet, not a measured deadline on my career.
The goal isn’t to defend every line I would have written. It’s to help a team ship software it can trust.
I don't have to say it. You know.
We can't agree to unionize or on anything really hard or controversial, but I wish people could agree to take a really tiny step together: don't upvote obviously not-human-authored drivel, even if you like the message.