feanaro

This is unrelated to the above comment, but I figured I'd ask while we're here. You state at the end of your blog post on this topic that you'd now probably recommend the NIST curves over Curve25519. Why is that?

I was responding to a comment from a person claiming this is not an example where the "intention/substance of the work might exist somewhere buried within a wall of impenetrable LLM text". I am saying this is in fact the case, given human experts almost universally agree that the papers are badly written, with inscrutable, hard to follow logic.

I had initially understood you as saying that this inscrutable logic was merely a consequence of superhuman intelligence, which is what I deemed unlikely. I don't see experts agreeing with that position either.

I am not unhappy about this result, but I am somewhat unhappy about it being dumbed down, decontextualised and understood without any nuance.

I mean, this is possible, but it seems quite unlikely if you take a good impartial look. It's nothing I personally dislike, in fact I enjoy LLMs when they are understood as what they are, which is a tool.

What I do not don are these rose-tinted, thought-terminating glasses which make people assume that this just must be an impenetrable product of superhuman intelligence.

No, it's just a word for which LLMs have an inordinately large bias, so it's often found in blog posts written by one.

How is it not true here? Plenty of papers in this release are hard to follow, use non-standard terminology, make hard to understand leaps, and so forth.

Even if it were not, it is a wise aim, unlike ceding understanding to a robot.

Of course, sorry if that came across as more combative than necessary. I was trying to point out two things I guess:

1. "Your experience is not universal from my perspective, and I offer my experience as evidence"

2. The possibility that this adaptation you are observing is neither a stable equilibrium given future developments, nor is necessarily an indication that the people using AI have bought into it fully. I use AI heavily while remaining a sceptic.

What exactly is appalling? Are you referring to something in the sentence you quoted?

It's possible and I dare say even likely that you are missing nuance and observing bias from your bubble. It's possible to avoid being an AI "opponent" while still being a sceptic. It's possible to use AI reluctantly and with tact instead of going all in.

The whole thing is far from over when we haven't experienced the worst of the technology yet: deskilling, concentration of power, extreme imbalance of wealth, military and surveillance uses, etc.

I think this matters very little. These were adaptations to initial advances of small scope which we are rapidly leaving behind. If there is truly no near and practical ceiling to the abilities of AI in mathematics, not even in terms of efficiency, then I think it is likely there is similarly no ceiling in other cognitive endeavours.

In that case, everything stands to be disrupted and those who wield AI will have massive amounts of power in the short term. But this is not a stable situation, and therefore those who wield AI will have to cede control of it or otherwise enter into a great war with those that do not.

I hope for all our sakes that there is a practical ceiling.

Are you advocating for inaction? If so, what do you wish to achieve?

If you consider the data on which an LLM was trained on to be points on a very highly multidimensional object, the claim is that the LLM can interpolate a convex hull spanned by those points, therefore recovering a subset of consequences attainable from those points. Obviously this hull includes completely novel points that were not present in the initial data set, so the output of the LLM goes beyond its initial training. And yet, there are clearly points outside a convex hull spanned by any finite number of points, such that we can imagine not all possible outputs are attainable using this method.

The claim is furthermore that truly original thinking, the infamous leaps in understanding and creativity, happen by attaining points outside such a convex hull.

It's hard to rigorously verify or disprove this claim. Hopefully this helps build an intuition of why the claim is not as shallow and obviously wrong as it may seem initially.

Surveillance is decidedly and completely unlike locks.

> This is the million dollar we'll see answered in our lifetime. Software engineering exists to automate work, are we arrogant to think we are not destined to the same fate? Is this truly a job befitting of a human over an agent?

There is a fine distinction here that I believe is often glossed over, so the two things it's delineating get muddled together. One of those two things is coding—the rote, mechanical encoding of meaning into computer instructions. It can be argued the LLM is fit to take this out of hands hands almost entirely, and it's almost indisputable the LLM is better at this in at least a certain sizeable subset of coding tasks.

But the other thing is the choosing, determination, specification of the intended meaning itself. This I think is squarely the job of the human, because letting this fall through to the AI means it is no longer the human that is making the decisions. This then becomes not merely automating work but ceding control. This, ultimately, is a bad thing.

So if we accept the premise that the specification of the intended meaning is the job of the human, the question is how you do that. Today many of us do it somewhat half-assedly, by writing lots of natural language text at the LLM and hoping it sticks. It is our hope that the text, given that there is a lot of it, will drive the stochastic machine in a sufficiently correct direction. This works to a degree—meaning we've ceded some control but not the majority of it—not least because we still read (most) of the code but cannot work in the limit, if code reading ceases.

A more proper way to specify the intended meaning is to specify (or "model") your system formally in a system that is mechanically verifiable. Then the final artefact produced by the LLM can be validated by verifying that it aligns with the specification. However this type of high-assurance specification looks a lot like a certain type of programming. In my opinion, writing this kind of specification is the future of human software engineering.

I do not accept the approach of simply rolling the dice and hoping the machine knows better than us, though I'm sure that church is also going to have its acolytes.