Sure do :) I can provide more if you want, but start there, as it's a good read. Indeed, blind people are much better at understanding the consequences of colours than they are at knowing what colours things are..

Very much unlike an intern (the doing-over part).

You ignored his core point, which is that "rocks don't think" is useless for extrapolating unless you can define some procedure or model for evaluating whether X can think, a procedure that you can apply both to a rock and to a human and get the expected answers, and then apply also to ChatGPT.

Heh. It says a lot about the pace of AI research and discussion that a paper from last year is "old".

This is a common thread I notice in AI criticism, at least the criticism of the "AI isn't really thinking" or "AI can't really do much" sorts... it all references the state of the art from a year or two ago. In most fields that's entirely reasonable. I can read and reference physics or math or biology or computer science papers from last year and be pretty confident that I'm reading the current thinking. If I'm going to depend on it I should probably double-check, but that's just due diligence, I don't actually expect it to have been superseded. But in the AI field, right now, a year old is old. Three years old is ancient history, of historical interest only.

Even the criticism I see that doesn't make the mistake of looking at last year's state of the (public) art tends to make another mistake, which is to assume that you can predict what AI will be able to do a few years from now by looking at what it does now. Actually, most such criticism pretty much ignores the possibility that what AI will do in a few years will even be different from what it can do now. People seem to implicitly assume that the incredibly-rapid rate of change we've seen over the last five years will suddenly stop, right now.

For example, I recently attended the industry advisory board meeting for my local university's computer science department. The professors there, trying desperately to figure out what to teach CS students today, put together a very well thought-out plan for how to use AI as a teaching tool for freshmen, gradually ramping up to using it as a coding assistant/partner for seniors. The plan was detailed and showed great insight and a tremendous amount of thought.

I pointed out that however great a piece of work it was, it was based on the tools that exist today. If it had been presented as recently as 12 months ago, much of it wouldn't have made sense because agentic coding assistants didn't really exist in the same form and with the same capabilities as they do now. What are the odds that the tools won't change as much in the next 12 months as they have in the last 12 months? Much less the next four years, during the course of study of a newly-entering freshman.

The professors who did this work are smart, thoughtful people, of course, and they immediately agreed with my point and said that they had considered it while doing their work... but had done what they had anyway because prediction is futile and they couldn't do any better than making a plan for today, based on the tools of today, fully expecting to revise their plan or even throw it out.

What they didn't say, and I think were shying away from even thinking about, is that their whole course of study could soon become irrelevant. Or it might not. No one knows.

The congenitally blind have never seen colours. Yet in practice, they're practically as efficient at answering questions about and reasoning about colours as the sighted.

One may raise questions about qualia, but the older I get, the weaker the qualia argument gets. I'd argue that I have qualia about abstracts, like "justice". I have a visceral feeling when I see justice and injustice, and experience it; it's highly associative for me. Have I ever touched, heard, smelled, seen, or tasted an object called "justice"? Of course not. But the concept of justice is so connected in my mind to other things that it's very "real", very tangible. If I think about "the colour red", is what I'm experiencing just a wave of associative connection to all the red things I've seen, some of which have strong emotional attachments to them?

What's the qualia of hearing a single guitar string? Could thinking about "a guitar string" shortly after my first experience with a guitar string, when I don't have a good associative memory of it, sounding count as qualia? What about when I've heard guitars play many times and now have a solid memory of guitar sounds, and I then think about the sound of a guitar string? What if it's not just a guitar string, but a riff, or a whole song? Do I have qualia associated with *the whole song*? The first time? Or once I know it by heart?

Qualia seems like a flexible thing to me, merely a connection to associative memory. And sorry, I seem to have gotten offtopic in writing this. But to loop back: you don't have to have experienced something to have strong associations with it. Blind people don't learn of colours through seeing them. While there certainly is much to life experiences that we don't write much about (if at all) online, and so one who learned purely from the internet might have a weaker understanding of those things, by and large, our life experiences and the thought traces behind them very much are online. From billions and billions of people, over decades.

Language does not exist in a vacuum. It is a result of the thought processes that create it. To create language, particularly about complex topics, you have to be able to recreate the logic, or at least *a* logic, that underlies those topics. You cannot build a LLM from a Markov model. If you could store one state transition probability per unit of Planck space, a different one at every unit of Planck time, across the entire universe, throughout the entire history of the universe, you could only represent the state transition probabilities for the first half of the first sentence of A Tale of Two Cities.

For LLMs to function, they have to "think", for some definition of thinking. You can debate over terminology, or how closely it matches our thinking, but what it's not doing is some sort of "the most recent states were X, so let's look up some statistical probability Y". Statistics doesn't even enter the system until the final softmax, and even then, only because you have to go from a high dimensional (latent) space down to a low-dimensional (linguistic) space, so you have to "round" your position to nearby tokens, and there's often many tokens nearby. It turns out that you get the best results if you add some noise into your roundings (indeed, biological neural networks are *extremely* noisy as well)

As for this article, it's just silly. It's a rant based on a single cherry picked contrarian paper from 2024, and he doesn't even represent it right. The paper's core premise is that intelligence is not lingistic - and we've known that for a long time. But LLMs don't operate on language. They operate on a latent space, and are entirely indifferent as to what modality feeds into and out from that latent space. The author takes the paper's further argument that LLMs do not operate in the same way as a human brain, and hallucinates that to "LLMs can't think". He goes from "not the same" to "literally nothing at all". Also, the end of the article isn't about science at all, it's an argument Riley makes from the work of two philosophers, and is a massive fallacy that not only misunderstands LLMs, but the brain as well (*you* are a next-everything prediction engine; to claim that being a predictive engine means you can't invent is to claim that humans cannot invent). And furthermore, that's Riley's own synthesis, not even a claim by his cited philosophers.

For anyone who cares about the (single, cherry-picked, old) Fedorenko paper, the argument is: language contains an "imprint" of reasoning, but not the full reasoning process, that it's a lower-dimensional space than the reasoning itself (nothing controversial there with regards to modern science). Fedorenko argues that this implies that the models don't build up a deeper structure of the underlying logic but only the surface logic, which is a far weaker argument. If the text leads "The odds of a national of Ghana conducting a terrorist attack in Ireland over the next 20 years are approximately...." and it is to continue with a percentage, that's not "surface logic" that the model needs to be able to perform well at the task. It's not just "what's the most likely word to come after 'approximately'". Fedorenko then extrapolates his reasoning to conclude that there will be a "cliff of novelty". But this isn't actually supported by the data; novelty metrics continue to rise, with no sign of his suppossed "cliff". Fedorenko argues notes that in many tasks, the surface logic between the model and a human will be identical and indistinguishable - but he expects that to generally fail with deeper tasks of greater complexity. He thinks that LLMs need to change architecture and combine "language models" with a "reasoning model" (ignoring that the language models *are* reasoning - heck, even under his own argument - and that LLMs have crushed the performance of formal symbolic reasoning engines, whose rigidity makes them too inflexible to deal with the real world)

But again, Riley doesn't just take Fedorenko at face value, but he runs even further with it. Fedorenko argues that you can actually get quite far just by modeling language. Riley by contrast argues - or should I say, next-word predicts with his human brain - that because LLMs are just predicting tokens, they are a "Large Language Mistake" and the bubble will burst. The latter does not follow from the former. Fedorenko's argument is actually that LLMs can substitute for humans in many things - just not everything.

Greater Fool scheme finds Greatest Possible Fools (goverments) to keep pumping money into their zero-sum game.

Them: Why don't you act your age?
Me: Well, I've never been this age before.

Like buying booze, renting a car, purchasing a handgun, buying a lottery ticket, getting a tatoo?

(some of these vary by state)

I don't see how you're too immature to order a Chianti with your steak dinner but you're mature enough to go $200K in debt based on a sales pitch of returns after investment.

These aren't even reasonable equivalents from a neuroscience perspective.

A read is supposed to be fine. At read time the firmware *should* rewrite the cell if the read is weak.

The firmware also *should* go out and patrol the cells when idle and it has power.

you can dd if=/dev/sdX of=/dev/null bs=2M once a year if your firmware behaves.

If your drive is offline you could
dd if=/dev/sdX of=/dev/sdX bs=2M iflag=fullblock conv=sync,noerror status=progress

to be sure, though write endurance is finite.

If you're running zfs you can 'zpool scrub poolname' to force validation of all the written data. This is most helpful when you can't trust the firmware to not be buggy crap. Which only applies to 90% of drive firmware out there.

Well 0-clicks and OTA attacks but yeah, to your point, device compromise lets you use apps on the device.

News at 11!

> I have found that streaming directly to my Plex home server over TLS is generally smoother without going through Wireguard. Not quite sure why.

I recently had to solve this.

Wireguard should work with a regular 1500byte MTU connection at 1440 or 1420 bytes (the default) --- however --- if your ISP is routing your IPv4 using 4-in-6 internally (like my major cable company) everything goes to hell.

Try dropping your wg MTU to 1360, MSS at 1320, and set up a mangle table to clamp MSS to PMTU (e.g. iptables rule).

I got a 10x bump in TLS over wireguard throughput.

Total pain in the ass and lightly documented.

Cite?

What conflict of interest? DCar Studio has money and wants advertising, the influencer wants money and can provide advertising. I see an alignment of interests, not a conflict.