Paul Jurjac

Per infinita aspera ad veritatem

[Paul Jurjac @pjurjac]

@SamoBurja ...Madam

That is especially important...

[Paul Jurjac @pjurjac]

@ToKTeacher @AndrewMayne Well said!

[Paul Jurjac @pjurjac]

Neither bothering, nor irritating as long as it is facilitating better communication. It might be a bit hard to keep the thoughts active while focusing on the interlocutor is active and I have to map what he is saying, but it isn't neither bothersome nor frustrating.

I would rather get more annoyed by a slow-rate of progress in creating the shared object-space of the discussion and the movement through them in any way shape of form. That is what is killing me personally and makes things tiresome...

[Paul Jurjac @pjurjac]

@gh0stpen @birdabo You saved me of pointing out the obvious, thank you for your service!

[Paul Jurjac @pjurjac]

Well... we could go technical at any point I guess - I would actually love to know why you think they are a viable option going forward. Or why do you think they are good if you consider the requirements in the limit, and which those are...

For example, assuming a schema of a few million objects (properly typed and non-edge relationships), and a population of a few billions (or could go to trillions or quadrillions to stress-test the instrumentarium), what instruments do you have to deal with the traversal / retrieval problem with maximal granularity with those technologies?

Assume you are willing to write dozens of thousands or millions of custom reads over said schema operands. The question is, which are the instruments you can use, and are they properly embedded into the space itself? AKA are the objects themselves (including the space) designed to empower your all possible finite search warrants? If yes, which are those exactly?

I have found that either of those are extremely limited in the limit quite a while back, and I mean... really limited. They are great for what they were built, but they were built for different problem-sets... Our problem space is totally, totally different going forward.

Or maybe I'm biased -- So let me ask you this; What are your problems in the limit technically speaking and why do you think those are viable options for those?

[Paul Jurjac @pjurjac]

@MillionInt I'm an idiot, I read your question wrong - I though you've asked if the transformer is an evolution or a revolution...

[Paul Jurjac @pjurjac]

And evolution which enabled a revolution. It changed the landscape of computing since it made it possible to program over our own symbolic system (and any arbitrarily defined one going forward as long as its defined) without having to go through the semantic model development anymore. Which was seriously bounded anyway when it came to allowed transformations over it...

However, the revolution will be a Formal Industrial Revolution, and one which will start from epistemic tools, to domain formalizations, to the development of the entire systemic embodiment of whatever is required to operate domains.

This was the great unlock, now the work begins...

[Paul Jurjac @pjurjac]

@grok read all the conversation and explain to person A why he doesn’t understand what a logical fallacy is - and explain to me why I’m such a big idiot engaging in technicalities when the person clearly doesn’t understand that he didn’t make neither an atomic nor a composite proposition which could set the frame for a logical discussion on top of the atomic boolean which he was commenting on…

[Paul Jurjac @pjurjac]

With pleasure. Or rather, a sense of duty. We’ve been suffering with these problems for years now - crossing the bridge between epistemic techne and the tech stack - and it wasn’t fun, so if I can help, I’m glad to do it.

If you have any specific problem set you wanna solve in the area, there is a good possibility we’ve dealt with it in some form or another and I might at least point you toward favourable trade-offs.

Feel free to hit me up if in need

[Paul Jurjac @pjurjac]

No, it isn't...

However, RDF or SPARQL is going to kill you once you are successful in doing everything else. That is to say, after the hardest parts are done.

So my unsolicited advice, is to not even go there, Its a dead-end. It doesn't allow you to properly control bounded infinities with granularity as inputs.

Run the problem, you'll see.

TypeDB is a great vanilla open source alternative for what you're doing. I hope it helps.

And congrats for the problem!

May you be worthy of it...

[Paul Jurjac @pjurjac]

@ccatalini Christian, we're gonna have so much fun together! Great thread, as always.

Looking forward to sending you some work which is quite on your lane, and I'm extremely sure you're gonna like - a lot!

[Paul Jurjac @pjurjac]

The worrying times for the world of AI, as well as the reliance on inference as primary driving force, are going to abruptly come to an end soon enough.

And the age of formal industrial revolution at its dawn, and our abundance with it.

May this gift be well received!

[Paul Jurjac @pjurjac]

Roadshow:"We're growing in capabilities so fast that we will eat the entire economy, including your existing portfolio! Almost all of your positions are exposed to us! Make sure to hedge so if your positions lose market share, you are going to win EVEN MORE! Protect the pensions!

[Paul Jurjac @pjurjac]

@velimirgasp @pmddomingos So much, that you should be scared for all others on the losing side 😭

Especially since - apparently - the economy is a zero-sum game

[Paul Jurjac @pjurjac]

Thank the Berlin's Triumverate for this and the 19th century reaction to the insanity of idealism and the radical positivist reaction. Plus the other rave party in the house of philosophy (especially politics and ethics), which made the secession of Natural Philosophy from philosophy, into "science" so they're not associated with the maddies.

Skip to the 1950's epistemic decline was in freefall...

Almost a century later since the freefall, we're now deep in Dante's hell. We've went from the 17th century's epistemic rise evident in Royal Society's "Nullius in verba" to Authoritative SCIENTISM

[Paul Jurjac @pjurjac]

Most of the time, I don't want to take the doomers seriously. Then I remember the decline of epistemic rigor across fields, and when combining that with perverse incentives, I have to concede that they might actually believe what they say.

If you think about it, the entire fear with AI doing damage is fully justified only if you accept one assumptions; That you're going to place an inferential engine, a probabilistic engine, into formal environments and let it loose.

If we accept that assumption, then yes, it is definitely going to do damage. But that is the equivalent of me saying that IF I jump out of the airplane without a parachute, I need an alternative solution to not die, which indeed I might engineer.

However, if one is looking at the LLM as for what it actually is - a programmable constructor over OUR symbolic operands - then it is easy to treat it like electricity. Which by the way, is also dangerous, and deadly.

That is not to say that it's uncontrollable, but it is to be placed within our engineered systems. And ideally, if something goes bad, is OUR error, not its error. It would be obviously idiotic to argue that a machiene performed badly due to the electricity powering it. Even if electricity in said example isthe cause, the underlying one is bad design.

Thus, the solution is rather simple in nature, and not so different. We should treat LLMs as functions in complex systems. And yes, that implies that one system might have millions of them, but the system itself must necessarily have designed state-driven processes, not inference driven ones.

That is painfully obvious, because any serious domain which is operationalized through such systems, might have millions of steps until an output is finished (though in reality there are N intermediary product to be created) and even if you assume a minimal inferential error per step, those are quickly compounding to super-exponentials.

Hence, that's why they believe AI is not strong enough to deal with domains yet in any serious economic sense. In fact, they are wrong. The ability to program an LLM as a function today is already well beyond the operability criteria.

The lack is our ability to define such systems. And since we're here, that lack is semantic, ontological and formal in the first place. As you mentioned in your recent paper, the hidden objects and processes which we take for granted and implicit are vast. I would argue -from serious experience as well as purely theoretical - that they are MOST of a domain, especially if we enter into Groethendieck-like structures to represent spaces, objects, processes, bridges,etc.

But this phase shall pass as well, and we're going to move toward a formal industrial revolution this century, we have no choice....

I'm happy to see that they're going in the right direction.

In the end, formal systems will win irrespective of temporary volatility in any dimensions. If for nothing else, civilization always demands responsibility from any actor at a certain point, and actors in turn will need assurances, and those are providable only by formal means.

That is not WHY formal methods will succeed - obviously they will due to their advantages and efficiency gains - but rather why they will be demanded

[Paul Jurjac @pjurjac]

@bnielson01 @velimirgasp Yes, that was the post to which Velimir was responding to, and I perceived your response - wrongly - as a defence of that, or an expansion of computational theory outside of what you were actually referring to.