لغة تاو: توليف البرامج مستقبل |عرب اورج

Is neurosymbolic AI the same idea?

Everyone wants society to be logical, until they examine the logical contradictions which underpin everything we are accustomed to. We live in a world of a million cognitive dissonances. It might be quite nice to have the language aspect he discusses, but tying a "logic engine" into operating real-world social systems is a doomed endeavor.

“Here’s a project. Does it have a token? Yes -> Stupid.” – George Hotz

My favorite passtime is walking through the forest listening to MLST.

🎯 Key points for quick navigation:

00:00 🧮 Machine learning is a mathematical miracle that learns from examples, but accuracy will never reach 100% and error will never be zero.
00:55 🔄 Tau language is designed as the "endgame of all blockchains" where users can simply state what they want, and the system automatically implements it.
01:23 💼 The concept of an "automatic businessman" where you tell the system to make profit for you is introduced as one powerful aspect of Tau.
01:59 🧠 The speaker has spent 15 years in AI, initially focused on machine learning but shifting to the intersection of logic, mathematics, and computer science.
04:43 📈 Machine learning is approaching its peak capabilities and won't improve much further without incorporating non-ML methods.
05:40 🎲 For complex logical problems beyond a certain size, ML models perform no better than random guessing (coin toss).
07:06 📉 Overfitting in machine learning is compared to polynomial interpolation – fitting everything perfectly leads to random predictions.
08:57 🧪 Statistical ML methods always provide regression to the mean (general answers) rather than specific predictions for unique situations.
10:21 🔍 The three curses of machine learning: optimization curse, statistical curse, and approximation curse limit its capabilities.
13:09 🗣️ Humans are fundamentally obsessed with living in a conceptual world of language, which is often more central to us than physical reality.
17:21 🔄 Words like "tomato" are abstractions that point to but never fully touch physical reality; language approximates but never completely captures the world.
19:19 📚 The speaker defines language broadly as any system of expressing things symbolically, not just specific languages or verbal communication.
20:47 🧩 For logical reasoning, it's more direct to use logic itself rather than encoding it indirectly through linear algebra as ML does.
21:48 🏗️ Synthesis is more advanced than verification; instead of verifying if a system meets requirements, synthesis automatically generates systems from requirements alone.
22:17 🧪 Software synthesis is like writing only tests and having the computer automatically create a program that passes those tests.
23:09 🔐 Mathematical frameworks enable proving guarantees about infinite possible inputs, such as ensuring a phone never sends passwords over the network.
26:01 ⏱️ Software synthesis requires writing tests in languages where the halting problem is decidable to guarantee correctness.
32:36 🧩 This specific form of logical AI achieves self-reference by abstracting sentences into Boolean algebras.
34:02 🎯 Unlike 1980s expert systems that tried to model the entire world, Tau focuses on making computers do specific tasks users want rather than attempting to accurately describe everything.
35:57 🔢 Mathematical logic faces fundamental limits because language contains only countably many sentences, while mathematical realities extend beyond countable cardinality.
39:26 📚 The AI winter in the 1970s was triggered by the Lighthill debate, which recommended defunding logical AI due to the perceived complexity of logical reasoning.
40:47 🚀 SAT solvers have shown Moore's Law-like improvement, becoming better every year and providing empirical proof that logical reasoning is practically feasible.
43:44 🧩 Most real-world systems need both logical components and some "guessing" capabilities (machine learning), and logic provides a framework to describe the complete system.
45:37 🔍 In synthesis, you describe the tests that software must pass rather than describing the software itself.
46:07 ♾️ There are infinitely many programs that could meet a given specification; if you're not comfortable with the system choosing one arbitrarily, you must add more constraints.
47:06 🔄 Specifications need updating either because of bugs in the original specification or because reality changes, requiring adaptation.
49:29 🔍 To overcome Tarski's undefinability of truth, the speaker abstracted sentences to be Boolean algebra elements, viewing them not by structure but as logical elements.
54:15 🧠 The Tau language extends standard Boolean algebra by allowing reference to any element (not just 0 and 1), using its own sentences as the Boolean algebra, and adding a temporal dimension.
57:04 📊 Initially, Tau could only interpret specifications but couldn't compile them into standalone programs; the full synthesis algorithm was only finalized about a month before this interview.
58:24 ⚙️ A challenge with generated programs is efficiency; they may not be the most parsimonious or optimized implementations.
59:19 🧠 The system can identify easy cases to optimize by doing logical solving beforehand, then outputting straightforward functions that don't need runtime logical solving.
01:00:12 📈 There's a relationship between specification complexity and program complexity; as specification size grows to infinity, program complexity will also increase.
01:00:42 🔧 Strong simplification algorithms can significantly reduce complexity even with suboptimal requirements specifications.
01:01:40 👁️ Human readability of normalized or simplified formal statements remains an ongoing challenge; simplification doesn't always lead to better intelligibility.
01:02:09 🔄 The Tau language workflow involves a human-readable interface, intermediate normalization, and then optimization during compilation.
01:02:38 🧩 The general belief revision problem is unsolvable according to established research; there's no optimal way to choose which elements to modify when contradictions arise.
01:04:30 🔍 Pointwise revision allows changing only small components without rewriting the entire program; you write only what needs to change and the rest remains intact.
01:05:28 🧮 Quantifier elimination can make complex logical statements more intelligible by removing existential operators during normalization.
01:06:25 🤔 No one can mathematically define "intelligibility," making it impossible to guarantee that transformed specifications will remain intelligible to humans.
01:10:15 🔓 The speaker believes there's no other way to make software truly controlled by users except through the Tau approach.
01:12:04 🌐 Tau enables collective control of peer-to-peer networks and blockchains by all users, unlike current blockchains controlled by small groups of developers.
01:13:58 🔗 Blockchain doesn't need Tau, but blockchains need Tau to be redefined over time by users in a sound way.
01:17:36 💼 The concept of an "automatic businessman" allows users to specify assets, goals, and acceptable deals, then tell the system to make profit autonomously.
01:22:54 🪙 The Agoras token (AGRS) is currently available as a temporary token on Ethereum that will be swapped when the Tau blockchain launches.
01:24:16 🧠 Tau could enable collective intelligence by allowing thousands or millions of developers to specify requirements that the system integrates, potentially creating software of unprecedented scale and quality.
01:25:45 🎯 The best software is defined as one that does exactly what users want it to do; by this definition, Tau aims to be the best software.
01:26:17 🏇 The speaker acknowledges Steve Jobs' famous quote about people wanting "faster horses," suggesting that users might not initially know what they truly want.
01:27:15 💬 Tau is described as "a discussion about Tau" where users collectively discuss what the system should be, creating an "opinion map" that tracks which opinions contradict others.
01:27:46 📊 The consensus from user discussions becomes the software update, with even the definition of "consensus" being modifiable through the "laws of changing the laws."
01:28:13 📚 There will be a learning curve for users to define requirements in the logical language, though the language itself is minimalistic; understanding the domain is harder than learning the syntax.
01:29:08 🔒 While Tau is designed for change, users could potentially lock the system by implementing changes that prevent future modifications.
01:30:34 🛡️ The risk of system entrenchment is mitigated by logical AI providing guarantees about what can happen, monetary incentives, and the collective nature of decision-making.
01:31:55 👥 Many users might delegate their influence to trusted experts, potentially resulting in a form of meritocracy rather than direct democracy.
01:33:51 🧪 The system will start with a testnet using fake coins, allowing users to experiment with rules and governance; if deadlocks occur, they can restart from scratch before implementing on the mainnet.
01:40:30 🧠 The speaker's primary stance is against "illogicalism" in all contexts (LLMs, social structures, etc.), believing that bringing logic to our lives through computer mechanization can fix many problems.

Made with HARPA AI

Smart contracts can already be written by anyone. It will be interesting TAU bridges the gap from unskilled to skilled. And future llm type beat does bave a rhythm to it

Ohad Asor has a novel theory about AI. His explanations are precise. Imagine him writing a book on AI to expound his perspective — that we live in language and perceive the world through it — , such book would redefine the AI revolution, I presume.

Definitely not your conventional AI approach, he's talking about using logical methods, some of which are his own breakthroughs, to achieve new forms of software implementation and decentralizations, as I understand it

So using his declaractive programming language with unique logical properties to implement a blockchain or other software that can evolve itself in a way that reflects its users changing requirements. Sounds pretty next level

Was there some kind of dare to incorporate the phrase 'calcium cannons' into the interview? That made me spit out my drink.

John 1:1, in the beginning was the word…

I think this guy has a good shot at being right, though the future is hard to predict.

I like learning about "calcium cannons".

>A token
Yeah I'm out. See you next episode.

I am all in for new languages and stuff. I went to the website to look for examples, demos etc to see how software is written in this new tau languages. There was nothing. Only flashy images and promises. Not a single example or demo.

Hybid Camp