I think your assessment is spot on.
But I also think there's a bigger picture that's getting lost in the sauce, not just in your comment but in the general discourse around AI progress:
- We're currently unlocking capabilities to solve many tasks which could previously only be solved by the top-1% of the experts in the field.
- Almost all of that progress is coming from large scale deep learning. Turns out transformers with autoregression + RL are mighty generalists (tho yet far from AGI).
Once it becomes cheap enough so the average joe can tinker with models of this scale, every engineering field can apply it to their niche interest. And ultimately nobody cares if you're playing by the same rules as humans outside of these competitions, they only care that you make them wealthy, healthy and comfy.
The problem with respect to what? The end-goal of self-driving cars (and humanoid robots) is to work in the environments created for humans. Otherwise we can just put down rails across all cities and call it a tram, or design purpose-built robots for all tasks.
Edit: Stated more explicitly: the human world is the way it is because of many reasons and can't always be changed naively (it's not like nobody in Cairo has thought about improving the traffic situation, or architects haven't thought about the ease of cleaning different flooring material). Robots which are general purpose with respect to their human-like capabilities must necessarily also accept a world in which humans live.
The end goal of is to make them work in reasonable environments. If it works fine in 90% of cities but doesn't work in Cairo, then fuck Cairo, no driverless cars for them.
I hate this comic because it is profoundly lazy, and I hate it when people hand-wave away meaningful security advances with it.
Hitting people with wrenches leaves marks that can be shown to the media and truth & reconciliation commissions. Wetwork and black-bagging dissidents leaves records: training, operational, evidence after the fact. And it hardly scales – no matter what the powers at be want you to think, I think history shows there are more Hugh Thompsons than Oskar Dirlewangers, even if it takes a few years to recognize what they've done.
If we improve security enough that our adversaries are _forced_ to break out the wrenches, that's a very meaningful improvement!
Yes: if you have half of a billion dollars in BTC, sure – you're a victim to the wrench, be it private or public. If you're a terrorist mastermind, you're likely going to Gitmo and will be placed in several stress positions by mean people until you say what they want to hear.
Extreme high-value targets always have been, and always will be, vulnerable to directed attacks. But these improvements are deeply significant for everyone who is not a high-value target – like me, and (possibly) you!
In my lifetime, the government has gone from "the feds can get a warrant to record me speaking, in my own voice, to anyone I dial over my phone" to "oh, he's using (e2e encrypted platform) – that's a massive amount more work if we can even break it". That means the spectrum of people who can be targeted is significantly lower than it used to be.
Spec-fiction example: consider what the NSA could do today, with whisper.cpp & no e2e encrypted calls.
Would you be willing to enter a 100'000'000-to-1 bet? If we make it before the century, you make me a millionaire and if we don't I'll give you a dollar. Should be within your risk tolerance given you think the chance is 0.
We expect this approach to work because it's currently the best working approach. Nothing else comes close.
Using symbolic language is a good idea in theory, but in practice it doesn't scale as well as auto-regression + RL.
The IMO results of DeepMind illustrate this well: In 2024, they solved it using AlphaProof and AlphaGeometry, using the Lean language as a formal symbolic logic[1]. In 2025 they performed better and faster by just using a fancy version of Gemini, only using natural language[2].
Note: I agree with the notion of the parent comment that letting the models reason in latent space might make sense, but that's where I'm out of my depth.
The example is a bit of an unfair comparison though. It takes questions already posed in natural language and, as far as I can tell, expects results in natural language, too.
This means that whatever system is evaluated in this challenge necessarily has to deal with natural language. And indeed, a big part of the AlphaProof system was a neural network to convert from natural language to Lean.
None of this has anything to do with reasoning ability.
I think it would be interesting to present an inverse challenge, where the problems are already posed in a formal language. Would a network that first converts them into natural language, then does chain-of-thought on that, then translates the result back into formal language still be better than a simple symbolic reasoner that could operate on the formal language directly?
Different in the sense that they consume more alcohol? https://en.wikipedia.org/wiki/List_of_countries_by_alcohol_c... That it is legal for children to drink under parental supervision also doesn't necessarily mean that parents will allow it, so the legal situation isn't necessarily the deciding factor.
For what it's worth, in both Denmark and the UK, my experience has been that children are indeed allowed it on occasion, often celebrations like Christmas where they will have something like bucks fizz or a little cider or something alongside the adults.
- We're currently unlocking capabilities to solve many tasks which could previously only be solved by the top-1% of the experts in the field.
- Almost all of that progress is coming from large scale deep learning. Turns out transformers with autoregression + RL are mighty generalists (tho yet far from AGI).
Once it becomes cheap enough so the average joe can tinker with models of this scale, every engineering field can apply it to their niche interest. And ultimately nobody cares if you're playing by the same rules as humans outside of these competitions, they only care that you make them wealthy, healthy and comfy.