Comment by bob1029

6 days ago

It was embarrassing a year+ ago when much of this technology barely worked at all.

If you did not perceive a step change in model capabilities in late 2025, it's possible you are working on problems that are just too hard or not well aligned (e.g., arts and crafts). If you are working on software and need to move the occasional mountain, it's really hard to argue with the capabilities of contemporary models and harnesses.

This space is evolving rapidly. Faster in some areas than others, but you should still keep an open mind and adapt to the new realities as they emerge. Wishing for a thing to fail can blind you to more opportunities than you would probably be willing to reject if they were presented directly.

15 comments

bob1029

Reply
surgical_fire  6 days ago

A year+ ago, LLMs worked nearly as good as they work today.

I say this as someone that uses LLMs on a daily basis. It is a moderately useful tool for software development, but it was already useful in January 2025 (to pick a random date). Check what models were available by then. There were improvements since them, but they were mostly marginal.

The thing is, ever since ChatGPT 4 days there are people claiming on a daily basis exactly what you are claiming now, and it feels pretty silly. Sometimes it's hard to not imagine the accounts breathlessly claiming 100x productivity as bots.

In other words, I presume that if you are 100x more productive now, you were a 1/100x developer before LLMs. I actually question the capabilities of anyone making claims like that.

  • JeremyNT  6 days ago

    The "100x productive" thing is some kind of bogeyman. It's pretty clear these things are useful and helpful and the degree to which they are such continually improves, but putting some specific large number on it is pointless.

    If you didn't notice a difference between AI assisted coding in Jan 2025 and now beyond "marginal" improvements... well to be frank I don't think you've really been paying very close attention. It's a combination of the way we interact with the models and the models themselves, and this really did get much more effective in late 2025.

    You can say the model improvements in that year were incremental, but the degree to which the overall capabilities have changed and what that means for software development is more than the sum of those parts.

    • surgical_fire  6 days ago

      > The "100x productive" thing is some kind of bogeyman.

      > If you didn't notice a difference between AI assisted coding in Jan 2025 and now beyond "marginal" improvements... well to be frank I don't think you've really been paying very close attention.

      Those are the same claims, just phrased differently. If the first is a bogeyman, so is the second.

      And of course, if someone disagrees that the improvements since 2025 were not dramatic, they can only possibly not be paying attention. Except we have been seeing the same claims on a daily basis for years.

UpsideDownRide  6 days ago

Sure it's better. 100x productivity is pure nonsense and there are other claims of this order.

Rejecting exaggerated claims doesn't mean that someone is not seeing any value.

  • QuantumGood  6 days ago

    One of the things not represented well by the hype cycle is how hype is a moving target. When "ai defenders and hype" assign a number to things, the early numbers can become closer to the later realities.

archagon  5 days ago

Ok but none of those articles are remotely interesting. At least if you’re into the science part of computer science — algorithms, data flow, software and hardware architecture, intimate knowledge of the machine — and not just productivitymaxxing to make a messy pile of CRUD widgets that nobody asked for.

It’s all influencer and hustle culture now. The joy of technical creation and understanding has evaporated from the mainstream in mere months. We’re surrounded by AI meat puppets wearing the masks of engineers.

customguy  6 days ago

We have access to and use the same tech, the difference is elsewhere.

> It was embarrassing a year+ ago when much of this technology barely worked at all.

The things some people say, how they talk about "outraged AI critics" what they would prefer those to be, would be fucking stupid even in first grade.

These, and things like "AI-generated news sites to attack critics" really speak for themselves. There is something worth really calling out here, and it fell out naturally.

xeonax  6 days ago

Nov'25-Dec-25 something changed. I believe that was the time when Opus 4.5 was launched. (we went from function auto-complete to module auto-complete) People came from the vacations, and were like, make AI do your work. One could actually start offloading work to LLMs.