Comment by augusteo
1 day ago
The framing of AI risk as a "rite of passage" resonates with me.
The "autonomy risks" section is what I think about most. We've seen our agents do unexpected things when given too much latitude. Not dangerous, just wrong in ways we didn't anticipate. The gap between "works in testing" and "works in production" is bigger than most people realize.
I'm less worried about the "power seizure" scenario than the economic disruption one. AI will take over more jobs as it gets better. There's no way around it. The question isn't whether, it's how we handle the transition and what people will do.
One thing I'd add: most engineers are still slow to adopt these tools. The constant "AI coding is bad" posts prove this while cutting-edge teams use it successfully every day. The adoption curve matters for how fast these risks actually materialize.
What makes you think that they will just keep improving? It's not obvious at all, we might soon hit a ceiling, if we haven not already - time will tell.
There are lots of technologies that have been 99% done for decades; it might be the same here.
From the essay - not presented in agreement (I'm still undecided), but Dario's opinion is probably the most relevant here:
> My co-founders at Anthropic and I were among the first to document and track the “scaling laws” of AI systems—the observation that as we add more compute and training tasks, AI systems get predictably better at essentially every cognitive skill we are able to measure. Every few months, public sentiment either becomes convinced that AI is “hitting a wall” or becomes excited about some new breakthrough that will “fundamentally change the game,” but the truth is that behind the volatility and public speculation, there has been a smooth, unyielding increase in AI’s cognitive capabilities.
> We are now at the point where AI models are beginning to make progress in solving unsolved mathematical problems, and are good enough at coding that some of the strongest engineers I’ve ever met are now handing over almost all their coding to AI. Three years ago, AI struggled with elementary school arithmetic problems and was barely capable of writing a single line of code. Similar rates of improvement are occurring across biological science, finance, physics, and a variety of agentic tasks. If the exponential continues—which is not certain, but now has a decade-long track record supporting it—then it cannot possibly be more than a few years before AI is better than humans at essentially everything.
> In fact, that picture probably underestimates the likely rate of progress. Because AI is now writing much of the code at Anthropic, it is already substantially accelerating the rate of our progress in building the next generation of AI systems. This feedback loop is gathering steam month by month, and may be only 1–2 years away from a point where the current generation of AI autonomously builds the next. This loop has already started, and will accelerate rapidly in the coming months and years. Watching the last 5 years of progress from within Anthropic, and looking at how even the next few months of models are shaping up, I can feel the pace of progress, and the clock ticking down.
I think the reference to scaling is a pretty big giveaway that things are not as they seem - I think it's pretty clear that we've run out of (human produced) data, so there's nowhere to scale to in that dimension. I'm pretty sure modern models are trained in some novel ways that engineers have to come up with.
It's quite likely they train on CC output too.
Yeah, there's synthethic data as well, but how do you generate said data is very likely a good question and one that many people have lost a lot of sleep over.
Which technologies have been 99% "done" for "decades?"
Bicycles? carbon fiber frames, electronic shifting, tubeless tires, disc brakes, aerodynamic research
Screwdrivers? impact drivers, torque-limiting mechanisms, ergonomic handles
Glass? gorilla glass, smart glass, low-e coatings
Tires? run-flats, self-sealing, noise reduction
Hell even social technologies improve!
How is a technology "done?"
It's not! But each one of your examples is in a phase of chasing diminishing returns from ever-expanding levels of capital investment.
It's done when there is no need to improve it anymore. But you can still want to improve it.
A can opener from 100 years ago will open today's cans just fine. Yes, enthusiasts still make improvements; you can design ones that open cans easier, or ones that are cheaper to make (especially if you're in the business of making can openers).
But the main function (opening cans) has not changed.
Technology is just a lever for humanity. Really would like an AI butler, but I guess that's too hard (?). So many things AI could do to make my life better, but instead the world is supposedly over because it can summarize articles, write passable essays, and generate some amount of source code. In truth we haven't even scratched the surface, there is infinite new work to be done, infinite new businesses, infinite existing and new desires to satisfy.
This is a really good question.
What convinces me is this: I live in SF and have friends at various top labs, and even ignoring architecture improvements the common theme is this: any time researchers have spent time to improve understanding on some specific part of a domain (whether via SFT or RL or whatever), its always worked. Not superhuman, but measurable, repeatable improvements. In the words of sutskever, "these models.. they just wanna learn".
Inb4 all natural trends are sigmoidal or whatever, but so far, the trend is roughly linear, and we havent seen seen a trace of a plateau.
Theres the common argument that "Ghipiti 3 vs 4 was a much bigger step change" but its not if you consider the progression from much before, i.e. BERT and such, then it looks fairly linear /w a side of noise (fries).
Even if the technology doesn't get better, just imagine a world where all our processes are documented in a way that a computer can repeat them. And modifying the process requires nothing more than plain English or language.
What used to require specialized integration can now be accomplished by a generalized agent.
The trade-off is replacing deterministic code with probabilistic agents. I've found you still need a heavy orchestration layer—I'm using LangGraph and Celery—just to handle retries and ensure idempotency when the agent inevitably drifts. It feels less like removing complexity and more like shifting it to reliability engineering.