> From what I've seen in these HN discussions, most people are using "determinism" when they really mean "prompt sensitivity", i.e. minor variations in framing leading to different results. This, in turn, confuses people who do understand what determinism is supposed to mean and where it's necessary (build reproducibility for example).
For lack of a better word, I'd also have used "determinism". But to borrow a bit from TFA, what I'd really mean by that would some kind of "semantic determinism": for any input source code in a well-defined language, a correctly working compiler will always produce output that's semantically correct for the input.
Let's say a compiler implementation internally does something random or nondeterministic but that the nondeterminism does not affect the semantics of the output. You could argue that the compiler is technically nondeterministic, but in terms of program semantics it would still be deterministic.
I assume that's what people mean when they say compilers are deterministic in comparison to LLMs.
So in some sense the post is correct, but I think the author is somewhat pedantically misinterpreting the way people use the word "nondeterminism".
IMO prompt sensitivity is something different. A prompt does not unambiguously describe full program semantics in the first place, and the neural network would not contain an explicit mechanism for producing semantically matching output even if it did. Prompt sensitivity comes on top of that but isn't the core matter.
I think you want the impossible, because a) input semantics are non-formal and ambiguous/subjective by definition, and b) the model suffers from the curse of its knowledge being vastly wider than yours and doesn't have enough context to converge on exactly what you want in the huge space of possibilities presented by even the most constraining but still informal inputs.
If you limit your requirement to the difference between your and model's interpretations being small enough, that's probably doable. Which is realistically what most people want, and most good coding models already have, more or less (with caveats that still need to be addressed, of course). But a hard guarantee of output staying unchanged with different inputs is not possible to give (regardless of whether you think they're unambiguous) due to the nature of intelligence, human or machine.
I'm not asking for LLM tools to be similar to compilers, or saying that they can't be useful if they aren't. I know rather well that the two are different, and that's the point.
Because LLMs aren't deterministic in terms of producing semantically correct output,
that just means they aren't similar to compilers. That means you probably can't just start blindly trusting that their output matches the input and thus ignore understanding the code, as most people mostly can with compilers.
I think that's what people mean with "determinism" when they compare LLMs to compilers, or in response to other people suggesting LLMs are no different than compilers.
In my (admittedly limited) experience, a verbose and elaborate writing style is also traditionally more common in humanities whereas scientific or technical writing favours a rather more terse and matter-of-fact style.
I don't know about the structures you mention specifically but if you compare an article on humanities or social matters against the style that's common in science and technical writing, chances are it's going to look more verbose in any case.
I don't necessarily have the best AI-dar but TFA didn't ring any LLM bells to me.
Sounds like the short answer is "because there was no standard for the variable set by MS-DOS from the get go".
The background is that the issue hadn't existed in CP/M because there hadn't been environment variables. Perhaps if the issue had already been seen in CP/M, the developers of MS-DOS might have defined a standard variable to avoid it. Maybe. Other than that it doesn't seem to have much to do with CP/M specifically.
I think #2 is actually circular, or perhaps rather contradictory. In order to be able to have an illusion one would have to be conscious in the first place. Or how would you have an illusion of something if you're not aware enough to experience that illusion? So I don't think the concept of "illusion of consciousness" makes much sense. (It does make sense for others to have an illusion that an AI or some other entity is conscious, but not for the entity itself.)
> Pain isn't a real thing any more than an IEEE float is a real thing. A circuit flips bits and an LED shows a number. A set of neurons fire in a pattern and the word "Ow!" comes out of someone's mouth.
Perhaps, but I think a physical presence is still required for consciousness, at least for any kind of consciousness that resembles ours.
It's perhaps easier to talk about qualia rather than consciousness, but I think qualia are a prerequisite for consciousness anyway.
Basically all of our qualia are somehow related to our needs in the physical world. We feel physical pain because it signals that our body is in danger of being damaged. We feel emotional pain from social rejection because for most of our history humans have needed other people for physical survival. (Or in some cases perhaps because our genes make us want to procreate and we failed at that.) Either way, our needs in the physical world are not being met. Evolution has produced genetic code that produces a brain that somehow makes us feel that subjectively, even if nobody knows how.
Those subjective experiences of course get processed by neurons, assuming you accept materialism. (Neurons are AFAIK significantly more complex than the "neurons" in ANNs, so equating biological neuronal activity with ANNs is wrong. But I suppose in principle any physical process may be represented or at least approximated by some symbolic representation, so in theory that probably doesn't matter.)
We can also express those subjective qualia in terms of language. However, I don't think it's possible to have our qualia (or consciousness) based on language or symbolic manipulation alone if it doesn't have some kind of a connection to our physical needs.
If you could directly simulate an entire human brain and feed it artificial sensory input, I suppose it would actually be conscious without having a physical body. In principle an AI could also evolve consciousness based on survival needs even if it were not biological.
But for example LLMs have been trained only on the symbolic level. Their "neural" structure is not simulating a brain and they don't have a connection to physical needs. I think that makes them incapable of consciousness even if the output they produce successfully mimics human language -- that is, symbolic representations of our qualia and conscious thought.
I'm not sure if that's the point the author is making. But I think the distinction between the purely symbolic "map" and the "actual thing" sort of makes sense.
You could do all of those things on an OS with proper security separation as long as you have full root access.
There's no megacorp stopping you from reading and writing kernel memory. Unless, of course, the computer refuses to run software not signed by the megacorp or some software refuses to run without a digitally signed chain all the way down to the firmware like some game anticheats do.
But that's not really because of things like permission boundaries for processes. You can have those and still be in full control of those boundaries. It may be more convoluted than in a barebones system like DOS, of course.
> I’m still at where when I connect external hard drive or SSD via USB, use it and then eject it, I shut down the MacBook Pro completely before I unplug the cable. Just in case.
That sounds... a bit paranoid? At least on Linux (Gnome), if I click to "safely remove drive" it actually powers off the drive and stops external mechanical drives from spinning. No useful syncing is going to happen anyway once a hard drive no longer spins. A modern OS should definitely be reliable enough that it can be trusted to properly unmount a drive.
> For the laptops that I actually carry around and plug and unplug things to etc, normal amount of time between reboots for me is somewhere between every 1 and 3 days. Cold boot is plenty fast anyway, so shutting it down after a day of work or when ejecting an external HDD or SSD doesn’t really cost me any noticeable amount of time.
I personally don't reboot my laptop that often, but it's not because of a boot taking too much time. It's because I like to keep state: open applications, open files, terminal emulator sessions, windows on particular virtual desktops, etc.
A lot of that probably came down to the motherboard chipset. IIRC Intel made their own chipsets for the Pentium III and they were good and reliable. Athlons were coupled with chipsets from VIA and whatnot.
Some of those chipsets were fine and others were less reliable or compatible. The quality of the drivers for each chipset may also have mattered.
If running low on memory seems to matter less now than it did a couple of decades ago, I'd rather say that's because fast SSDs make swapping a lot faster. Even though virtual memory and swapping were available even on PCs since Windows 3.x or so, running out of memory could still make multitasking slow as molasses due to thrashing and the lack of memory for disk cache. The performance hit from swapping can be a lot less noticeable now.
Of course compression being now computationally cheap also helps.
reply