AI Pullback Has Officially Started

AI creates the most spectacular happy path demos. It’s hard not to extrapolate to infinity when you see it.

People roll out a complex and powerful technology without understanding the technology fully, what evals are or updating process to account for the tech, and the rollout fails, news at 11.

Seriously though, "AI fucks up" is a known thing (as is humans fuck up!) and the people who are using the tech successfully account for that and build guardrails into their systems. Use version control, build automated tests (e2e/stress, not just unit), update your process so you're not incentivizing dumb shit like employees dumping unchecked AI prs, etc.

I'm not saying AI is living up to the "hype" or "expectations" - it would largely depend on how you quantify the hype or expectations. Most rational would be to consider how much money is funneled into vs how much ROI would it have within some time range in the future, e.g. 10 years. A wise investor would look ahead 10 years, balance benefits, potential and risks. By that metric it could be too early to say if it's paying off even if it's objectively clearly bringing 10x more expense than income.

But the metrics or facts without context or deeper explanations also don't mean much in that article.

> 95% of AI pilots didn’t increase a company’s profit or productivity

If 5% do that could very well be enough to justify it, depending on for which reasons and after how much time the pilots are failing. It's widely touted that only 5% of start ups succeed, yet start ups overall have brought immense technological and productivity gains to the World. You could live in a hut and be happy, and argue none of it is needed, but none the less the gains by some metrics are here, despite 95% failing.

The article throws out numbers to make a point that it wanted to make, but fails to account for any nuance.

If there's a promising new tech, it makes sense that there will be many failed attempts to make use of it, and it makes sense a lot of money will be thrown in. If 5% succeed, it takes 1 million to do 1 attempt, but the potential is 1 billion if it succeeds, it's already 50x return.

In my personal experience, if used correctly it increases my own productivity a lot and I've been using AI daily ever since GPT 3.5 release. I would say I use it during most of what I do.

> AI Pullback Has Officially Started

So I'm personally not seeing this at all, based on how much I personally pay for AI, how much I use it, and how I see it iteratively improving, while it's already so useful for me.

We are building and seeing things that weren't realistic or feasible before now.

AI (LLM) is useful for coding and I use it to lookup various articles or websites and summarize.

Use it where it works.. ignore the agents hype and other bullshit peddled by 19yo dropouts.

Unlike the 19yo dropouts of the 2010s these guys have brain rot and I don’t trust them after having talked to such people at start up events and getting their black pill takes. They have products that don’t work and lie about numbers.

I’ll trust people like Karpathy and others who are genuinely smart af and not kumon products.

There are a lot of people invested in AI, so they are cheerleaders. There are way more people who didn't invest, who are sour grapes and want to see it fail. I'm neither of these people, but it's a democracy after all. I think AI is due for another winter.

>Back in 2024, 54% of researchers used AI — that figure jumped up to 84% this year

kind of makes me doubt the pullback. Maybe the hype's dying but it's getting along as an everyday tool?

I find these sorts of takes to be tiresome. It is absolutely true there is a lot of hype around AI. Also it is true that many AI companies try to shove AI into everything without necessarily thinking wherer it is a good idea or whether it useful (talking to you Google). Notwithstanding this it is absolutely clear how transformational the technology is. For low skill tasks it can certainly substitute people and save a lot of time. For harder things one has to be more careful and the right model have to be used, i.e. it is not a silver bullet but just a powerful tool which means it needs to be used in conjunction with other tools

Articles like this keep popping up because they're catnip to those who hate AI or feel threatened by it. On coding, I routinely see people trashing vibe coding and jumping on the slightest mistake agents may make, never mind that human devs screw up all the time. And write-ups citing stats on AI coding tend to be written by folks who either don't code for a living or never earnestly tried it.

I use Claude Code regularly at work and can tell it is absolutely fantastic and getting better. You obviously need to guide it well (use plan mode first) and point to hand coded stuff to follow, and it will save you enormous amount of time and effort. Please don't put off trying AI coding out after reading misinformed articles like this.

I personally think a big factor here (i.e., on HN discussions) is that, to programmers, gen-AI seems amazing because we happen to do something it appears to do well and which can be useful if we supervise it. But really, typing up the code has always been the low-skill part of the job! Anyone who's been in the biz 10 years or more knows that that new coders can create code that seems ok but actually creates nightmares long term.

To people who aren't programmers, there isn't really the same kind of easily-verified use case. Most people can't tell at a glance that a business proposal or email is full of errors they need to correct, thus the stuff causes even more damage.

Unfortunately, programmers, as a rule, aren't terribly good at listening to the experiences and perspectives of non-coders, so I don't see this dynamic changing anytime soon.

There's a few bits of information from the original sources that's left out:

- The METR paper surveyed just 16 developers to arrive at their conclusion. Not sure how that got past review. [0]

- The finding from the MIT report can also be viewed from a glass 5% full perspective:

> Just 5% of integrated AI pilots are extracting millions in value. > Winning startups build systems that learn from feedback (66% of executives want this), retain context (63% demand this), and customize deeply to specific workflows. They start at workflow edges with significant customization, then scale into core processes. [1]

[0] https://arxiv.org/abs/2507.09089

[1] https://mlq.ai/media/quarterly_decks/v0.1_State_of_AI_in_Bus...

> To stem the backlash, many journals and universities are starting to resist or have stopped using AI altogether in the peer review process.

… Excuse me, they were doing _what_? The world has gone mad. Unless you think a chatbot that doesn’t know how many ‘r’s are in strawberry is your peer you shouldn’t be using it for peer review, bloody hell.

At least in tech there’s still usually human code review which catches the worst of the magic robot-generated nonsense.

I often wonder how much code people are generating at once to get such poor quality. I'm usually doing one behavior on a class/classes at a time and get great results. It's a bit longer and a little more tedious but by the time I am done the context window is refined enough that it can write my tests with ease.

I’ve been using AI coding systems for quite some time, and have worked in neural networks since the 90’s. The advancements are, frankly, almost as crazy as 90’s neural net devotees like me were claiming could be possible in the eventual future.

That said, the non-tech-executive/product-management take on AI has often been an utter failure to recognize key differences between problems and systems. I spend an inordinate amount of time framing questions in terms of promises to customers, completeness, reproducibility, and contextual complexity.

However, for someone in my role, building and ideating in innovation programs, the power of LLM assisted coding is hard to pass up. It may only get things 50% of the way there before collapsing into a spiral of sloppy overwrought code, but we often only need 30-40% fidelity to exercise an idea. Ideation is a great space for vibe coding. However, one enormous risk in these approaches is in overpromising the undeliverable. If folks don’t keep a sharp eye on the nature of the promises they’re making, they may be in for a pretty wild ride; with the last “20%” of the program taking more than 90% of the calendar time due to compression of the first “80%” and complication of the remainder.

We’re going to need to adjust. These tools are here to stay, but they’re far from taking over the whole show.

We should expect pullbacks, fuckups, plans failing, and rollouts getting canned. It's part of how humans do things. Its actually a pretty effective optimization algorithm.

I'd bet that some sort of exponentiate the learning rate until shit goes haywire then rollback the weights is actually probably a fairly decent algorithm (something like backtracking line search).