"We Really Don't Know How to Compute!" - Gerald Sussman (2011)

>LLMs are not "compelled" by the training algorithms to learn symbolic logic. I think "compell" is such a unique human trait that machine will never replicate to the T. The article did mention specifically about this very issue: "And of course people can be like that, too - eg much better at the big O notation and complexity analysis in interviews than on the job. But I guarantee you that if you put a gun to their head or offer them a million dollar bonus for getting it right, they will do well enough on the job, too. And with 200 billion thrown at LLM hardware last year, the thing can't complain that it wasn't incentivized to perform." If it's not already evident that in itself LLM is a limited stochastic AI tool by definition and its distant cousins are the deterministic logic, optimization and constraint programming [1],[2],[3]. Perhaps one of the two breakthroughs that the author was predicting will be in this deterministic domain in order to assist LLM, and it will be the hybrid approach rather than purely LLM. [1] Logic, Optimization, and Constraint Programming: A Fruitful Collaboration - John Hooker - CMU (2023) [video]: [link] [2] "We Really Don't Know How to Compute!" - Gerald Sussman - MIT (2011) [video]: ▶ this video [3] Google OR-Tools: [link] [4] MiniZinc: [link]

in “LLMs aren't world models” · View on HN

It seems that in order for robotics and automation to work properly, AI models including LLMs, YOLO, RL and others need helps from their distant cousins namely logic, optimization and constraint programming that can be attributed as intelligent automation or namely IA [1],[2],[3],[4]. [1] Logic, Optimization, and Constraint Programming: A Fruitful Collaboration - John Hooker - CMU (2023) [video]: [link] [2] "We Really Don't Know How to Compute!" - Gerald Sussman - MIT (2011) [video]: ▶ this video [3] Google OR-Tools: [link] [4] MiniZinc: [link]

in “We accidentally solved robotics by watching 1M hours of YouTube” · View on HN

> We found that LRMs have limitations in exact computation: they fail to use explicit algorithms and reason inconsistently across puzzles. It seems that AI LLMs/LRMs need helps from their distant cousins namely logic, optimization and constraint programming that can be attributed as intelligent automation or IA [1],[2],[3],[4]. [1] Logic, Optimization, and Constraint Programming: A Fruitful Collaboration - John Hooker - CMU (2023) [video]: [link] [2] "We Really Don't Know How to Compute!" - Gerald Sussman - MIT (2011) [video]: ▶ this video [3] Google OR-Tools: [link] [4] MiniZinc: [link]

in “The Illusion of Thinking: Strengths and limitations of reasoning models [pdf]” · View on HN

>Software 2.0 are the weights which program neural networks. >I think it's a fundamental change, is that neural networks became programmable with large libraries... And in my mind, it's worth giving it the designation of a Software 3.0. I think it's a bit early to change your mind here. We love your 2.0, let's wait for some more time till th e dust settles so we can see clearly and up the revision number. In fact I'm a bit confused about the number AK has in mind. Anyone else knows how he arrived at software 2.0? I remember a talk by professor Sussman where he suggest we don't know how to compute, yet[1]. I was thinking he meant this, Software 0.1 - Machine Code/Assembly Code Software 1.0 - HLLs with Compilers/Interpreters/Libraries Software 2.0 - Language comprehension with LLMs If we are calling weights 2.0 and NN with libraries as 3.0, then shouldn't we account for functional and oo programming in the numbering scheme? [1] ▶ this video

in “Andrej Karpathy's talk on the future of the industry” · View on HN

I'm really surprised out of current nearly 300 comments and counting, nobody really is mentioning constraint programming or CP, a seemingly forgotten deterministic sibling of the stochastic data-driven AI [1],[2],[3]. "Out of these 14 tasks, we guessed that only 6 require abstract reasoning alone to perform. Strikingly, we classified only one of the top five most important tasks for medical scientists as relying solely on abstract reasoning. Overall, the most critical aspects of the job appear to require hands-on technical skills, sophisticated coordination with others, specialized equipment use, long-context abilities, and complex multimodal understanding." Almost all of 14 R&D tasks listed in the table including that are not suited for data-driven AI with abstract reasoning can be solved by CP. Provided that we are allocating enough compute resources with at least similar to that we're currently providing data-driven AI with the crazy amount cloud networks of massively parallel compute CPU/GPU/TPU/etc. Fun facts, the modern founder of Logic, Optimization, and Constraint Programming is George Boole, the grandfather of Geoffrey Everest Hinton, the "Godfather of AI" and "Godfather of Deep Learning". [1] Constraint programming: [link] [2] Logic, Optimization, and Constraint Programming: A Fruitful Collaboration - John Hooker - CMU (2023) [video]: [link] [3] "We Really Don't Know How to Compute!" - Gerald Sussman - MIT (2011) [video]: ▶ this video

in “Most AI value will come from broad automation, not from R & D” · View on HN

>I believe that solving autorouting will be a massive unlock for physical-world innovation and is a key piece to enable the “vibe-building” of electronics. I strongly believe that the CAD world including EDA is at the verge of disruption by an AI or more correctly Machine Intelligence (Constraint Programming - CP to be exact) very similar to how LLM disrupting the Chatbot technology [1],[2]. The path to this is most probably by solving the autorouting mechanism with CP, a deterministic logic, optimization and constraint programming machine based intelligence [3], [4], [5],[6]. Fun facts, the modern founder of logic, optimization, and constraint programming is George Boole, the grandfather of Geoffrey Everest Hinton, the "Godfather of AI" and "Godfather of Deep Learning". [1] Most AI value will come from broad automation, not from R & D (313 comments): [link] [2] Diagrams AI can, and cannot, generate (68 comments): [link] [3] Constraint programming: [link] [4] Logic, Optimization, and Constraint Programming: A Fruitful Collaboration - John Hooker - CMU (2023) [video]: [link] [5] "We Really Don't Know How to Compute!" - Gerald Sussman - MIT (2011) [video]: ▶ this video [6] High-Quality Ultra-Compact Grid Layout of Grouped Networks [PDF]: [link]

in “Things I would have told myself before building an autorouter” · View on HN

Reminds me of Gerald Sussman's talk "We Really Don't Know How to Compute!" (2011) ▶ this video

in “Scoped Propagators” · View on HN

As Gerald Sussman says in "We really don't know how to compute!": > It's not dataflow! It's closely related, but it's not. ▶ this video

in “Scoped Propagators” · View on HN

"Suppose you get a salamander. You know, a salamander can regrow its limbs if you cut them off. So all these nasty biology students do things like, they'll cut a limb of the salamander -- a hand, the arm off, just between the shoulder and the elbow, and at the same time, cut the same one between the elbow and the hand, flip around that segment, and re-sew it back on. Nasty thing to do to a salamander. What do you think you get? Does anybody here know? What you get is three elbows. It grows two new elbows to make up for the fact that the wrong thing is connected to the wrong thing." --Gerald Sussman, "We Really Don't Know How To Compute!" ( ▶ this video ) These images look like someone spliced salamander regeneration DNA into a human, cut them into many pieces, and connected the wrong thing to the wrong thing all over the place. Maybe that's what the AI wants to do to us, AM style (I Have No Mouth and I Must Scream)?

in “The 'Compact' Version of Stable Diffusion 3 Is Generating Monstrous Human Bodies” · View on HN

It's not clear how many classical calculations a single human neuron is equivalent to. There's a strong analog component in multiple domains (strength, frequency, timing) and each neuron can connect to up to 15,000 other neurons. Assuming the brain's neurons are (probably unrealistically) fairly 'digital' we get an estimation of the human brain being equivalent to 1 exaflop (this is the currently accepted lower bound, and rather disputed as being too low). Current TPUv4 pods currently provide approximately 9 exaflops. I don't think we're currently reaching human-level learning rates. There’s currently no accepted “upper bound” on estimates of FLOP equivalency to a human brain. ▶ this video is a recently posted video to HN Frontpage which was summarized as such: > Though we have been building and programming computing machines for about 60 years and have learned a great deal about composition and abstraction, we have just begun to scratch the surface. > A mammalian neuron takes about ten milliseconds to respond to a stimulus. A driver can respond to a visual stimulus in a few hundred milliseconds, and decide an action, such as making a turn. So the computational depth of this behavior is only a few tens of steps. We don't know how to make such a machine, and we wouldn't know how to program it. > The human genome -- the information required to build a human from a single, undifferentiated eukariotic cell -- is about 1GB. The instructions to build a mammal are written in very dense c

in “StarCoder and StarCoderBase: 15.5B parameter models with 8K context length” · View on HN

There are lots of people with misconceptions of LLMs. It will take time to adjust. I reached the same conclusion as yourself, but do see a totally different path to take regarding information propagation (how GPT works). For example, cells merge information monotonically. This is how neural networks balance too, but could be applied in new/undiscovered ways. ▶ this video

in “Tech’s hottest new job: AI whisperer. No coding required” · View on HN

Oh what a fun question! So, the first thing is, what's the essential difficulty versus the noise: Lisp has two features which enable the evaluator to work. The first is quotation, which means that you can put a little apostrophe before some code and then it saves that code as a data structure, rather than implicitly evaluating it. The second goes by the overfancy name “homoiconicity”, that data structure is actually just singly linked list consumed by “car” (“first”) and “cdr” (“rest”) operations. You can focus on understanding those two aspects pretty quickly, and then there are many tutorials about building your own Lisp that will get you there in an afternoon, if you'd like. Or, you can take the royal road to this expression, in all of its cheesy 1980s glory: [link] It's lecture 7A where this specific program gets explicitly dissected by Sussman, wearing a fez. If he seems familiar, you might have watched Strange Loop 2011’s “We Really Don't Know How To Compute”, ▶ this video

in “Lisp-stick on a Python” · View on HN

This also reminds me of Gerry Sussman’s talk “We Really Don’t Know How to Compute!”: ▶ this video Granted, the assumption he made is that computing power would grow increasingly cheap (which is basically the opposite of the point here), but I also wonder if techniques like those shown in that talk could be used effectively with a network of smaller, lower-power computers. (I mean, the talk does mention the use of small low-power computers specifically, but more in the context of e.g. sensor networks.)

in “Permacomputing Wiki” · View on HN

I think he says that in every talk he gives. ▶ this video , [link] , [link] , etc. There's a 2002 paper where he uses the term: [link]

in “Thinking in an array language” · View on HN