I mean seriously is this the prediction folks are going with? Ok so we can build something like our SOTA coding agents today, breathing life into these things that 3 years ago were laughable science fiction, and your prediction is it will be worse from here on out? Do you realize coding is a verifiable domain which means we don’t technically even need any human data to improve these models? Like in your movie of 2050 everyone’s throwing their hands up “oh no we made them dumber because people don’t need to take 8 years of school and industry experience to build a good UI and industry best practice backend infrastructure”. I guess we can all predict what we want but my god
That's an INCREDIBLY good point about synthetic training data. During model training, AI agents could pretty much start their own coding projects, based on AI-generated wish-lists of features, and verify progress on their own. This leads to limitless training data in the area of coding.
Coding agents is what I’m talking about, they are also an old idea, everything is an old idea, what is new and a major step change is the realized capability of them in December 2025.
No, I’m saying AS OF Dec 2025. 2025 itself being a step change in that coding agent adoption has undergone a step change as a result of model quality and agent interface being good enough.
Understood, but I still think you're exaggerating. Tool use is a 2024 thing, and progress on model quality this year has been downright vomit-inducing (looking at you, OpenAI...)
What would you have expected model quality to have been this year, it’s greatly exceeded my expectations, I’m genuinely confused by this perspective…considering where we were a very very short time ago
I find a lot of folks share this sentiment but from where I sit it just sounds so much like the “kids these days” crap that spawned all of YOU folks when you were younger. I grew up so inspired by the internet culture of the nineties, people that understood a technology and had a passion for wrangling it to do great things. We had a mixed run and the internet today has simultaneously exceeded these early dreams by orders of magnitude in some ways and has become absolutely Orwellian and backwards in others. Same thing is happening here. It’s just so interesting seeing the same peers have such an identical take on this generations paradigm shift as the folks that we all ridiculed in the 90s. Those hilarious badly aged takes on the internet being a fad or not user friendly enough etc etc, I guess my naivite was to expect this time around we would be able to better recognize it in ourselves
Have you considered that the people in the 1990s were mostly correct, and it's you that has been corrupted by modern marketing influences and external pressures?
There's no shortage of "Chicken Little" technologies that look great on-paper and fail catastrophically in real life. Tripropellant rockets, cryptocurrencies, DAOs, flying cars, the list never ends. There's nothing that stops AI from being similarly disappointing besides scale and expectation (both of which are currently unlimited).
Again another common take; hint: if you’re against AI or the current investment in AI you have so many better and more nuanced arguments at your disposal than “AI is chicken little”. It’s already here. I’ve built so much stuff with Claude and Codex I’d have never have been able to build at a speed that is already incredible and it’s getting better and better every 6 months. Be worried about alignment or centralized unregulated power, worry about what wars will look like and how this is a pre packaged Stasi for any dictatorship. But “this is a fad equivalent in stupidity and hype to cryptocurrency and tripropellant rockets” is just kind of silly
I use AI regularly, it regularly disappoints me. I won't worry about alignment or centralizing the singularity because AI does nothing that we haven't seen already.
The one thing that AI hasn't done that was promised a million times over is make money.
Do you genuinely believe this? That AI is not making money? Maybe you just are referring to another tired refrain of people who don’t appear to understand the strategic play of pure AI companies which is that they operate at a loss?
I don't have to believe anything, I just look at the S&P 500 and see the same old stuff. Nvidia is enjoying the shovel shortage, but none of the gold-rushers have discovered anything better than CUDA. Nothing new under the sun.
> people who don’t appear to understand the strategic play of pure AI companies
Get a load of this guy. Strategy in isolation is worthless; Russia has excellent strategic deterrence that is utterly useless for deterring Ukraine. Pure crypto companies had strategic foresight, but none of it was worth a damn when they had to compete with each other on merit.
The strategic play is perfectly well-understood. The tactical side is not, so far Nvidia is the only company that has gone to war and won.
“Russia has excellent deterrence” is what you’re trying to say sanctions are not working to stop what Russia is doing to Ukraine? That not only demonstrates a bad understanding of geopolitics and how sanctions work, but also distracts from what we are actually talking about.
It’s not really clear to me what you are trying to say. There will be winners and losers and it will be hard to know who they will be. That has nothing to do with Anthropic/OpenAI/etc not being rational in their strategy...
Some of us do, and actively root it out. I’ve never in my life been more excited to sit alone in a room with an editor and a compiler than I am these days.
Idk for me the only issue I have with Rob’s take is that its a pretty overly dramatic one that oversimplifies and casts as black and white something much more complex. Obviously a very real living legend, much respect, and getting one of these emails is icky and distasteful but to make this into what he does is a bit much
I find these really are not only condescending but also really miss the mark and ironically come off as really uneducated in my opinion, and that really is the most infuriating type of condescension. What you call slopware today is becoming less and less sloppy every six months as new coding models drop. In 2 years the “unmaintainable mess” is going to be far better and far more maintainable than anything the engineers behind these snide websites will make. Do folks realize you can also use the same coding models to ask questions and reason about the “slop” that these code models are writing that somehow is able to do something I would never have been able to do before? I don’t really care if it’s 100% accurate, hit it with a hammer until everything makes sense. Yell at Claude and learn how to wrangle it to get what you want, that skill is an investment that’s going to pay you back far more than following the advice of these folks, that’s my opinion.
Like “you will learn better without AI” is just a bad short sighted opinion dressed up in condescension to appear wise and authoritative.
Learn your tools, learn the limitations, understand where this is going, do the things you want to do and then realize “hey my opinions don’t have to be condescendingly preached to other people as though they are facts”
Opus 4.5 is not better and less sloppy by an order of magnitude than sonnet 3.7? Amazing how people can’t look at what’s happening right in front of them
