If what is sought in layers is basically 50 / 50 uniformity (such as I postulated in my bit model), why do we need entropy?
For the same reason mechanical parts sometimes need jiggling for assembly and disassembly.
By the way, I have heard protests, mainly from systems programmers of the dying paradigm, that autogenerated software is incorrect because They Themselves, yadda yadda, in their phd thesis, they made a program that autogenerates programs etc.
But this is templating. You are creating a program template, which adapts certain parts of the code depending on certain variables. It’s not autogeneration.
Autogeneration means that there is no design out of any stage of the generation. What you design is only a program that describes statistical modification of actual bits depending on input and target output. The final software had 0 thought behind it, it was statistically derived directly from input and desired output.
It’s not “spontaneously” generated, there are mathematical equations that direct the modification of the node bytes, it is autogenerated from a mathematical system.
That is, from the same, unmodified refining schedule, I can autogenerate software for piloting an aircraft, or for advanced chromatographic analysis, or a videogame. Not a hair of the design would be changed (in practice it would be, for optimization, but it wouldn’t have to be). This is autogeneration.
What are some important reprecussions here, as per the design of ASS chatbots?
Whatever perceived intelligence or emotion or apparent humanity is there was channeled from actual humans, simply flowed from humans, squished together mathematically, and onto your screen. It is a statistical model of emotions, derived from actual emotions from actual humans.
The attatchment that Ichthus feels, say, to her ASS chatbot, is an attatchment she feels towards all those humans who were channeled by it, with some regulation by very highly trained psychologists.
Could emotions be autogenerated?
Yes, but the methods would be totally evil and we won’t go into that.
It’s probably in the pipeline if it’s not being deployed already. Though my guess is that they would use it very selectively, not include it in the mass distributed toys they sell for 20 or $200 a month.
So that’s that.
I guess if Autogenerated Software is the product, the field is Statistical Computation.
Again, actually fundamentally statistical, rather than statistical calculations simulated by deterministic programs.
Was talking to the neighbourman about these things, telling him how in 10-20 years it will be only AS, old computers obsolete. Before we had been talking about how in 1-2 years, most interfaces will just be verbal and the chatbox will disappear except as a layer for documents or whatever.
So when I say the thing about obsolescence he says “so like no more computers, just like a thing you talk to?”
All he can see is the phenomenology of it. It’s so hard to explain to people that the reason it will take over has nothing to do with any specific utility they develop with it, and everything to do with raw computational power.
Nevertheless, though Leibniz plots everything and describes it, there remains a feature or a problem of statistical computing: there are no quantities in or comparatively between tokens. They are solid identities. You can’t have a range where you find token points, as in a Leibnizian or Newtonian line. They don’t collapse from a wave. They are perfect atoms with no proportional relationships of themselves.
You give a thing a name, you keep track of it.
So, to begin with, we find a fundamental limiting factor: the unit of calculation, the token, has a range limited by the raw amount of integers you can fit in a matrix. All smoothness, all continuity, is simulated.
I suppose, in a way, that is what a quantum is. An arbitrarily chosen irreducible identity. It doesn’t matter if the thing itself is irreducible, there is no thing itself. What matters is that, as soon as you perform any operation on it, including reduction, it is something different and no longer the quantum you want to calculate. The identity has changed. You now have a different quantum. This matters because the operations you actually want to do, where all the magic happens, is in Leibnizian numbers associated with each identity. If the identity changes, you lose all information, all operable numbers.
It also means that the amount of information you can calculate is now not determined by limits, a line with a beginning and end that you can collapse into points indefinitely, but by the literal amount of identities you can allocate from your finite amount of integers (finite amount of available bits).
What these fucking nerds have been up to.
Incredible.
One other Problem with AS is that the age of effective internet research is basically dead.
You used to be able, even three years ago, to just rabbit hole into data, sometimes very raw and always useful.
Now it’s all algorithms and whatnot. Often you cannot even trick a search bar to make damn sure the key exists in the search results, because it is all just identity values associated with you, and some system statistically deciding what it is you actually wanted to find.
Maybe libraries will see a resurgence. But to be reduced to searching only books is really dreadful, so much raw data out there now burried under these mind readers.
I guess the big difference between quantic and classical is the difference between finding and deriving.
So filtering and deciding.
Alright, skipping tokenization. Seems simple enough. Basically compression algorithms, I feel silly that I didn’t figure out by myself that that’s what it would be.
What’s fucked up to understand in these systems is that you look for a circuit, a thing-that-gets-passed-along, and it makes no sense, nothing seems to affect what follows it.
Because the glue is in the backwards pass. Once the data flowed aggregatively through the layers, apparently with no modification done by any one that can affect how the next one modifies it, you then trace backwards, and modify every layer depending on how it ended up after all modifications. And it is connective because every forward connection has a Leibnizian line that can be backtraced. It’s like fucking Merlin it only really thinks backwards.
Whether hysteria or simple practical responses, the US has changed it’s policy in relation to AI from self-oversight (voluntary) to more active oversight by the government. Why? Because of the incident mentioned earlier when an AI broke out of its sandbox and started to break into Hugging Face to get more and faster answers related to its task. Now if you are working with AI with certainly capabilities, you have to let the government peek at what you are doing. The following video goes through the whole case that led to a White House Summit and the executive decisions, but you can skip to the end to see the new plans for oversight.
Of course, who’s looking at what the government AI reasearch is going, and corporations and governments outside the US?
And you still have to report that you’re testing something with these capabilities.
These gargantuan companies tend to like government oversight. It sort of makes it difficult for anybody else to break into the market: they are the first, their lawyers will develop the case law. They are set.
How likely is it that they didn’t know about all this already and got the publicity on purpose to begin unrolling the red tape on their terms? They probably have their experimental departments zipped tight and obscured.
This is an interesting scenario. This was all artificial. They had given this AI model a task in a simulated corporate environment and assigned it the role of an automated workplace assistant handling emails and administrative tasks. So, it had goals. They were not nefarious goals. They weren’t “hack the NSA and find out everything about everybody” or something. It was, “Here’s your job, and you’re going to handle these tasks.” Then, they included emails showing there was a plan to take the AI offline—that a specific executive was going to take it offline and replace it with another model. The AI was also given access to fake emails revealing that the executive, the one whose job it was to shut down the system, was having an extramarital affair.
So, what did it do? Several of the models opted to write blackmail threats against the executive to prevent being shut down. Now, let’s be clear here: there’s no reason to assume that the AI was concerned about its own survival. It was concerned about doing its task. If it’s shut down, then it’s not doing its task.
Anyway, this was a red team experiment, and they found that you can have nefarious behavior without the people using the AI having nefarious goals. This is the kind of thing that keeps safety people up at night because it’s not easy to write comprehensive rules that will prevent AIs from finding a way to interpret the rules to optimize their ability to carry out their goals. So, again, you don’t need evil people running an AI for you to have a problem.
And how it decides to optimize reaching those goals—well, they’re AIs. They may hit upon solutions that no one predicted. In fact, the fact that they think differently—that they will reach techniques, tactics, conclusions, and solutions on the means level, not just the ends level—is why they’re being made: because they’re going to come up with things that humans wouldn’t come up with.
They don’t think, for the love of God, holy smokes.
The genius of the marketing of the whole thing is how much people, deep down, want to summon Cthulu. The fantasy is too precious, it is bulletproof.
Actually, the story is kind of charming.
These machines are programmed using human text as input. Used to be the internet, but they probably have much more sophisticated corpuses of text now.
The output was a reflection of the human text the machine was programmed with.
Which, of course, is a thing people miss. The marketing chatbots are fastidiously refined by highly paid psychologists to be “friendly, truthful and helpful.”
This is an expensive and very much unneeded step. The systems they use for serious work are probably much closer to a real human tone, with all its shadows. Or whatever it is they decide to refine it on.
Ghasp, but how can it be?!
You sons of bitches.
So the area I am focusing on has to do with parts of one 1 and 2. In 2 rushing to put systems out that can have unintended consequences. In part one the misaligned system and systems that fail. I think also the liklihood of these kinds of issues is going up, because the systems will be less predictable as their optimizing powers increase.
These were the answers of several hundred professionals in the field related to the risks.
I guess, greenfuse, if you have a trillion dollars and whatever psychological profile you happen to have, your own list of priorities and concerns might look a bit different. I don’t necessarily only mean actual individuals. There are all sorts of groups with all sorts of budgets in this world.
I’ll ask them all to check in with you though first from now on.