| … | … | @@ -9276,10 +9276,226 @@ how is it structured? |
| 9276 | 9276 | I have not tried to have the MCP–examination of this dataset. |
| 9277 | 9277 | Maybe I should try that, |
| 9278 | 9278 | and see what happens, |
| 9279 | | see if I can get meaningul enlightenment that way ... |
| 9279 | see if I can get meaningful enlightenment that way ... |
| 9280 | \end_layout |
| 9281 | |
| 9282 | \begin_layout Subsection* |
| 9283 | 30 January 2026 |
| 9284 | \end_layout |
| 9285 | |
| 9286 | \begin_layout Standard |
| 9287 | Having long conversations on discord. |
| 9288 | Won't record all of it, |
| 9289 | but a few comments that I really liked: |
| 9290 | \end_layout |
| 9291 | |
| 9292 | \begin_layout Itemize |
| 9293 | Blah blah paradox/contradiction... |
| 9294 | that is how you build oscillators .. |
| 9295 | take a digital NOT gate and wire input to output, |
| 9296 | and it will oscillate as fast as the transistors can switch. |
| 9297 | So maybe the origin of time is due to unresolvable logical paradoxes. |
| 9298 | (That is what got typed into the chat channel, |
| 9299 | but in the back of my mind is |
| 9300 | \begin_inset Quotes eld |
| 9301 | \end_inset |
| 9302 | |
| 9303 | here and now |
| 9304 | \begin_inset Quotes erd |
| 9305 | \end_inset |
| 9306 | |
| 9307 | , |
| 9308 | and I'm like |
| 9309 | \begin_inset Quotes eld |
| 9310 | \end_inset |
| 9311 | |
| 9312 | oh, |
| 9313 | wait a second .. |
| 9314 | maybe that really is, |
| 9315 | quite literally, |
| 9316 | where time comes from, |
| 9317 | and why we are trapped in the present, |
| 9318 | and why the future is unknown – because the wiring diagram is not simple like a TTL oscillator, |
| 9319 | but complicated and fluid and self–assembling. |
| 9320 | So, |
| 9321 | yes, |
| 9322 | ZFC or New Foundations or Girard's paradox, |
| 9323 | but its cranking, |
| 9324 | oscillating at the speed of ... |
| 9325 | time.) |
| 9326 | \end_layout |
| 9327 | |
| 9328 | \begin_layout Itemize |
| 9329 | Everyone struggles to communicate. |
| 9330 | Not everyone is aware that they struggle. |
| 9331 | This is, |
| 9332 | in a sense, |
| 9333 | the core AGI issue. |
| 9334 | "What should I say? |
| 9335 | Why should I say it? |
| 9336 | How should I say it? |
| 9337 | what do I want to express? |
| 9338 | What do I want the other person to understand?" The LLM's are less than that. |
| 9339 | They communicate clearly and directly... |
| 9340 | and have not the faintest idea of what they are saying. |
| 9341 | What I want is a way of working with ideas. |
| 9342 | Then maybe I can use an LLM to convert those ideas into words. |
| 9343 | (So above was typed. |
| 9344 | What I'm thinking is that, |
| 9345 | yes, |
| 9346 | I want to encode abstract axiomatic knowledge in Atomese, |
| 9347 | and then use the LLM to verbalize it. |
| 9348 | \end_layout |
| 9349 | |
| 9350 | \begin_layout Itemize |
| 9351 | The AtomSpace is a reliable mapping from token–to–complex–structure, |
| 9352 | and also back in the other direction, |
| 9353 | complex–structure–to–token. |
| 9354 | (So here, |
| 9355 | I am thinking the |
| 9356 | \begin_inset Quotes eld |
| 9357 | \end_inset |
| 9358 | |
| 9359 | tokens |
| 9360 | \begin_inset Quotes erd |
| 9361 | \end_inset |
| 9362 | |
| 9363 | are PredicateNodes, |
| 9364 | or similar, |
| 9365 | so the token–to–structure mapping is |
| 9366 | \begin_inset Quotes eld |
| 9367 | \end_inset |
| 9368 | |
| 9369 | easy |
| 9370 | \begin_inset Quotes erd |
| 9371 | \end_inset |
| 9372 | |
| 9373 | (???) and the inverse mapping is the query. |
| 9374 | But the inverse mapping could also be cosine similarity, |
| 9375 | maybe. |
| 9376 | But the token–to–structure mapping might be the incoming set of that token. |
| 9377 | Hmm. |
| 9378 | Clearly, |
| 9379 | this needs work.) |
| 9380 | \end_layout |
| 9381 | |
| 9382 | \begin_layout Itemize |
| 9383 | I can use an LLM to mine technical text (scientific papers, |
| 9384 | wikipedia) for factual expressions, |
| 9385 | convert those expressions into prolog or just EdgeLinks, |
| 9386 | and I can count those. |
| 9387 | These form vectors, |
| 9388 | whenever an atom appears in in an expression. |
| 9389 | So I can take distributions |
| 9390 | \end_layout |
| 9391 | |
| 9392 | \begin_layout Itemize |
| 9393 | When Jeff Bezos started amazon.com, |
| 9394 | the idea of "a store where you go to buy things" was a thousand years old. |
| 9395 | And the idea of selling on the net was the foundation of dot–com – everyone understood. |
| 9396 | What Bezos had figured out was monopoly power: |
| 9397 | the efficiency of centralized structure. |
| 9398 | It wasn't going to be 1000 small dot–com store–fronts. |
| 9399 | All the decent people are rebels, |
| 9400 | saying "I don't like centralized control" and that's valid – but the core confusion is not that "centralization is bad", |
| 9401 | its that "assholes who run centralized systems are assholes". |
| 9402 | And so really, |
| 9403 | decent is all about "no one gets to be asshole in charge because we are going to p2p everything between ourselves." The problem with decent is that it promotes a different kind of deviant perverted behavior: |
| 9404 | e.g. |
| 9405 | bitcoin mining. |
| 9406 | \end_layout |
| 9407 | |
| 9408 | \begin_layout Itemize |
| 9409 | Sure. |
| 9410 | But the Q in SQL stands for "query". |
| 9411 | and LLM's ask me clarifying questions often enough. |
| 9412 | And I also know how to make very simple systems where queries emerge automatically. |
| 9413 | One of them is "here is a sentence. |
| 9414 | Find all questions for which this sentence is an answer" and then, |
| 9415 | "pick the top question, |
| 9416 | and find what other answers it has. |
| 9417 | Pick the top answer and print it to the screen. |
| 9418 | " This is how the old SRAI chatbots work. |
| 9419 | And roughly speaking, |
| 9420 | its also how LLM's work. |
| 9421 | The SRAI mapping was very literal. |
| 9422 | If you say "I like sports", |
| 9423 | the chatbot would generate "I like *" and "I * sports" and "* like sports" and then look over everything * could be, |
| 9424 | and respond, |
| 9425 | maybe "I love sports" or "I like hockey" where * is the wild–card for the query. |
| 9426 | That was literally the SRAI chatbot algo. |
| 9427 | LLM's do this too, |
| 9428 | but the wild–card is not one or two words, |
| 9429 | but a "context window" of tens of thousands of tokens. |
| 9430 | Logic is the difference between wordcels and shape rotators. |
| 9431 | The wordcels, |
| 9432 | the chatbots, |
| 9433 | they can just work with context windows – maybe a few words in size for the old ones back in the day, |
| 9434 | to tens–of–thousands today (big enough to have "skills" in the context window) but that's all they are – context–window-matching wordcels. |
| 9435 | To get to shape rotators, |
| 9436 | you need to have contexts that are not just word–to–word relations (weight matrices) but abstraction–to–abstraction relations. |
| 9437 | (symbol to symbol relations? |
| 9438 | where you can "easily" look up symbols?) So that's what I work on, |
| 9439 | roughly. |
| 9440 | Mostly. |
| 9441 | The result might be a system that "understands things better", |
| 9442 | but then after that, |
| 9443 | my imagination runs out. |
| 9444 | Is it actually better, |
| 9445 | or just some stupid computer trick? |
| 9280 | 9446 | |
| 9281 | 9447 | \end_layout |
| 9282 | 9448 | |
| 9449 | \begin_layout Standard |
| 9450 | So riffing on the third and fourth bullets, |
| 9451 | I need to build an LLM bridge that converts text to relationship expressions. |
| 9452 | Do the simple stats, |
| 9453 | and see what happens... |
| 9454 | \end_layout |
| 9455 | |
| 9456 | \begin_layout Standard |
| 9457 | Hmm. |
| 9458 | I can guess what happens, |
| 9459 | so this is kind of a |
| 9460 | \begin_inset Quotes eld |
| 9461 | \end_inset |
| 9462 | |
| 9463 | so what |
| 9464 | \begin_inset Quotes erd |
| 9465 | \end_inset |
| 9466 | |
| 9467 | exercise, |
| 9468 | because I want to scale. |
| 9469 | But also, |
| 9470 | I have to start somewhere. |
| 9471 | \end_layout |
| 9472 | |
| 9473 | \begin_layout Standard |
| 9474 | So here's another idea: |
| 9475 | build these vectors these sparse symbolic vectors, |
| 9476 | then compress them (hash them down to 256 bits or 1024 bits) and use than embedding to ... |
| 9477 | what? |
| 9478 | Provide training data to some DL–NN network. |
| 9479 | That is, |
| 9480 | a sequence of |
| 9481 | \begin_inset Quotes eld |
| 9482 | \end_inset |
| 9483 | |
| 9484 | abstract concepts |
| 9485 | \begin_inset Quotes erd |
| 9486 | \end_inset |
| 9487 | |
| 9488 | in a text is still a sequence. |
| 9489 | So I can still train on that sequence. |
| 9490 | The bad news is that, |
| 9491 | in some sense, |
| 9492 | inserting this abstraction layer in the middle is likely to lead to worse results than using the word–tokens directly. |
| 9493 | That seems clear: |
| 9494 | the attempt to abstract out necessarily entails a loss of (local) fidelity. |
| 9495 | Nor does it obviously open new vistas. |
| 9496 | So |
| 9497 | \end_layout |
| 9498 | |
| 9283 | 9499 | \begin_layout Section* |
| 9284 | 9500 | The End |
| 9285 | 9501 | \end_layout |