iia/1100 Tegmark Pro Human Future

• A Pro Human Future

  • A keynote on the path toward a future where AI amplifies human agency, creativity, and flourishing — not replaces it.
  • meta
    • [Keynote]
    • [Converging Intelligence]
      • 11:00 AM
    • Person/John Werner
      • Senior Fellow at MIT
      • Founder & CEO of Imagination In Action
    • Person/Max Tegmark
      • MIT
      • Professor of physics & AI research
    • Person/Ramesh Raskar
      • MIT Media Lab
      • Associate Professor of Media Arts and Sciences
  • topics
    • Person/Max Tegmark
      • My Side-Note
        • tegmark is wearing a black bomber jacket and some kind of nerdy "progression" tshirt with stages of development
      • right now ai is the only industry that is less regulated than sandwich shops
      • 3 superpowers - AGI is in the center of the triple Diagram/Venn/3d
        • autonomy
        • generality
        • domain intelligence
          • Alpha/Fold and protein folding
      • each person has all three superpowers
        • as long as an AI only has two of the three, then each person is more powerful than an ai
      • his claim is that "pro-human AI" is AI that's only two out of three
      • 11:09 tool ai case study: AI for "vericoding
        • Bugs and Security
          • just by doing pip/install or npm/install you can get all your Cred/en/tials stolen
        • Formal Verification
          • import evolution of coding Diagram

            •                  The evolution of coding
              
                 More bugs ↑
                           |
                           |        [ vibe coding ] ←──── [ trad coding ]
                           |              |                       \
                           |              |                        \
                           |              ↓                         \
                           |         [ vericoding ] ←──── [ formal auto-verification ] ←──── [ formal verification ]
                           |            🏰
                           |
                           +--------------------------------------------------------------→ More effort

          • it's much harder to verify that the proof is correct, than to prove that it's correct
          • FMxAI "virtuous circle
            • larger benchmarks -> better ML-powered verifiers and vericoders
            • greater FV adoption
          • FV Logseq/Entity/Software/Projects
            • Dafny
            • Rust / Verus
            • Lean
          • Signal/Messenger
            • wouldn't it be very cool if there was a mathematical proof that nobody was reading your messages
          • gzip has now been part of FV
          • having a formal Proof is different than having 150k test cases
          • the only way you can be sure that it will never mess up is using this AI/LLM/Technique
        • Relevance to AI/Safe/ty
          • this is about making AI/Models safer too
          • If I asked diana rouse can you make algo for flying to international space station, I wouldn't put diana in a spaceship and fly her up, I'd have her make a piece of software that runs the place
            • she can export knowledge into code
          • if a cat finds a new way to hunt mice, we can't put that into a computer
          • we can export a lot of knowledge
          • Conjecture
            • fraction of Knowledge that can be extracted grows with the Intel/ligence of species the Diagram
              • TODO import diagram

              •    What fraction of its knowledge and algorithms can a neural network export?
                
                   Extractable fraction ↑
                                       |
                        100% ──────────+----------------------------------------------.
                                       |                                   _________  |
                                       |                                __/         \_|
                         50% ──────────+                           ___/               |
                                       |                        __/                   |
                                       |                     __/                      |
                                       |                  __/                         |
                          0% ──────────+____ ____ ____ __/                            |
                                       |   🐱 Cat        AGI              ASI         |
                                       +------------------------------------------------→ Intelligence

  • Decentralized society and AI Safety
    • 11:23 Person/Ramesh Raskar asks a question about Dist/ri/but/ed architectures and AI/Safe/ty
      • Person/Max Tegmark
        • FV software is going to take off not because someone makes you use it but because people will pay more for it.
    • 11:27 Person/Ramesh Raskar more about decentralization