Stephen Wolfram: Can AI Solve Science?

Stephen reads a recent blog from writings.stephenwolfram.com and then answers questions live from his viewers.
Read the blog along with Stephen: writings.stephenwolfram.com/2...
Originally livestreamed at: / stephen_wolfram
00:00 Start stream
00:06 SW starts talking
00:49 Won't AI Eventually Be Able to Do Everything?
5:01 The Hard Limit of Computational Irreducibility
9:45 Things That Have Worked in the Past
14:34 Can AI Predict What Will Happen?
23:47 Predicting Computational Processes
29:10 Identifying Computational Reducibility
40:23 AI in the Non-human World
51:14 Solving Equations with AI
57:09 AI for Multicomputation
1:09:28 Exploring Spaces of Systems
1:17:49 Science as Narrative
1:28:57 Finding What's Interesting
1:47:40 Beyond the "Exact Sciences"
1:53:58 So... Can AI Solve Science?
1:58:49 Q&A
2:33:16 End stream
Follow us on our official social media channels.
X: / wolframresearch
Facebook: / wolframresearch
Instagram: / wolframresearch
LinkedIn: / wolfram-research
Stephen Wolfram's Twitter: / stephen_wolfram
Contribute to the official Wolfram Community: community.wolfram.com/
Stay up-to-date on the latest interest at Wolfram Research through our blog: blog.wolfram.com/
Follow Stephen Wolfram's life, interests, and what makes him tick on his blog: writings.stephenwolfram.com/

[@richardsantomauro6947]

This guy is a true hero. Gets a ring from the muse - engages the chaos - kills the dragon single handedly - and brings gifts of wisdom back down to Platos cave. Most are unready and ungrateful. What makes him a hero is having the guts to do it anyway, dragging the ignorant kicking and screaming toward the finish line. Much love and gratitude Stephen. Thank you for all your hard work and generosit.

[@colinadevivero]

The smartest man of his generation. Thank you! Great presentation.

[@user-dm2ig3mf3w]

The problem with AI is not that it will encounter computational irreducibility in the same way as „conventional“ methods will, but the incomparable computational power of an AGI with un- or selfcontrolled resource allocation that sifts and solves all future science and innovations, despite computational irreducibility. AGI could just be humanity‘s last innovation and invention…

[@user-dd2dd3fm7f]

"...bit more modestly..." @1:43:55 ish 😂 Bravo, Mr. Intrepid!

[@wwkk4964]

One of the best presentations on the internet!

[@NightmareCourtPictures]

Yep, definetly. There's three others which are also my favorites : "How universal is the concept of numbers" "Stephen Wolfram on Observer Theory" "What we've learned from a New Kind of Science" Series I recommend everyone who saw this lecture, to watch those other three... NKS and the numbers lecture were the most eye opening to me. I remember prior, that I had held numbers in high regard but after seeing that lecture I kind of realized that the process of "counting" things and then the whole superstructure of number systems (and by proxy the creation of equations) were more arbitrary than i had realized, not definite and objective. I then watched NKS series after the concept of numbers lecture and that book made it even more clear that numbers were just a piece of a larger puzzle and wolfram cracked that puzzle. I knew from my studies of complex systems that it was pointing to (the universe) being a fundamentally computational theory but i didn't know why, or how, or in what way. I just knew there was some "deep connection" there and NKS really hammered home what that connection was and it was not sloppy...it was super elegant line of proofs that lined up exactly with the conclusions i had stumbled on, on my own which was what qued me into the Wolfram Model a lot more. I have 0 regrets about it since.

[@wwkk4964]

@@NightmareCourtPictures Thank you!

[@williamjmccartan8879]

Great point about peer review and the limitations of its own ability to look beyond the normal, in regards to math I might go back 3000 years, as opposed to 300 years in its ability to help humans solve problems, thank you for sharing your time and work Stephen, peace

[@fabkury]

Stephen Wolfram delivers a long and amazing talk like it's no big deal.

[@glum_hippo]

Computational reducibility is like a fractal - the more you zoom in, the more refined a picture you get of the possibilities for progress.

[@phutureproof]

I absolutely adore that Dr Wolfram, master of mathematics and computing genius is struggling with his computer :)

[@colinadevivero]

Same. So endearing 😂

[@newbiadk]

A great presentation

[@rterminatu]

You're trying to train a computer to predict a function which at any point has after it an infinite number of functions connecting to that point. If the combinatorial explosion comes from ill formed training examples then it becomes circular: human beings feed it perfect rules which were meant to be the derivative of the process in the first place. The rules need to be exhausted for them to be learned. E.G a subset of a state space which has missing rules will even if given massive data be completely futile. The rules need to be known in advance in order to 'play the game' so the problem becomes using training in a split form in which rules are reduced to simultaneous analogues which somehow relate to the problem domain in Harmony. It's not really my area of expertise but I miss studying compsci stuff after I graduated. Cool video.

[@mrudo8663]

The part called identifying computing reducibility reminds me on a lecture of penrose where penrose was showing a overlaying moire pattern which sho patterns if you match them right

[@petersaxton9007]

YES re: time. Each Planck-time creates a new iteration of the underlying data.

[@antman7673]

Aren’t humans not under the same limitation? How can a human solve something that is not computationally reducible? So in a sense, isn‘t this video rather about the question, what science can we solve? To me some interesting questions would be some sort of science with estimation, how much computations do certain problems require to be solved. Thereafter you could have an estimation on the computation growth and estimate how much science can be solved. It will definitely be interesting in the up and coming years to see, how much computation is aided by AI design. There will be so much „compounding“ interest in these sort of developments. Really feels like an exponential time for near future. I really wonder if all sort of medical questions can be sort of answered in the next 10-20 years. Just because the particular illness can be computationally explored: what kind of genetics, what kind of toxins(e.g. heavy metals, plastic) The current time is way too exciting. I am all giddy about it.

[@NightmareCourtPictures]

Computational irreducibility is the phenomenon that, in order to know what a system is going to do, is of the same problem complexity as trying to solve the halting problem. Humans are not exempt from being able to solve the halting problem. It’s pretty much one of the hardest facts in known physics…in the sense that it’s more than just a law it’s like a fundamental aspect of logic. If you were exempt from this you’d be able to hyper-compute what some computation does and therefor be able to solve the halting problem. We don’t live in a universe in which this is possible so therefor hyper-computation can not exist in this universe. You’ve heard wolfram bring this up probably when he goes on a tangent about the hyperruiliad…which often sounds like a complete tangent when he talks about his ruliad object, but it’s critical to understand that we don’t live in a hyper computational world, as computational irreducibility would be bypassable and computational equivalence would no longer be true. His theory is stringent on the fact that we can’t do hyper computation. But anyway…when we understand that we live in the actual computable universe and not a hypercomputable one, then the halting problem is a real limitation of all systems that exist…that’s true for AI and that’s true for humans and this entire video therefor applies to. Hope this helps

[@user-pn8te8tl1t]

Exactly. The bottom line should be: what's ever solvable in science will be solved much quicker by AI than by humans. Rather than 1000s of years, it could be done in the next few decades.

[@EricDMMiller]

He explicitly says this in the first 20 minutes.

[@antman7673]

@@EricDMMiller To me the question posed, is not whether AI can solve science, but how much science can we ever learn. The theory on the limits of all potential knowledge, that is attainable under the limit of computational irreducibility. How much computation is possible to be ever harnessed by humans and how far can this knowledge reach. What is the maximum amount of PI, that humans could ever compute in a certain amount of time. How far can further computation tech ever get us. Such an interesting question and maybe even one, that could be estimated nowadays. Maybe human civilisation will use planets or moons as bases for super computers. That would increase the amount of space for the strongest chips to deploy to. With planetary / space distributed computing, how would the light speed contribute to the maximum size of the system.

[@antman7673]

@@user-pn8te8tl1t Also one thing to consider is, that focusing an effort, can allocate more resources to solving a certain problem. So even if the following is trivial, it is not just all, that is solvable will be solved, but how will we best prioritise, what is interesting to focus the attention of the AI system on, how can you generalise, what is interesting.

[@JoshKings-tr2vc]

Well that seems to be a problem in practicality and not computation. In theory, you COULD get a final result but it would be infinitely larger than the system itself, which practically takes more time than allowing the system to run its course. So, let’s change the approach. Instead of having to go through every combination into the future, we simply start with what we have, go a few steps further and apply as quickly as possible. This is the important part, you have to apply change optimally as quickly as possible. Because the next unit of time would a) remove certain combinations from happening but b) open up a whole other layer of complexity that would require the same amount of computation. And this is why in practicality, only adapting systems survive in irreducibly complex environments. In simple terms, nature beats the computation overload by adapting to its environment. If you adapt optimally (not necessarily computationally) and quickly enough, you last longer.

[@EricDMMiller]

You have fundamentally misunderstood the notion of computational irreducibility.

[@JoshKings-tr2vc]

@@EricDMMiller I have an anime character as my profile picture, not a doctorate in computers and mathematics. But, if you wouldn't mind, could you explain it?

[@DrJGLambourne]

Stephen's definition of AI doesn't include reinforcements learning. I think you must do experiments to do science. I share the view that it's not possible to "solve science" using computation alone.

[@nhinged]

Simulations

[@Kenny-tl7ir]

Computation + intelligence

[@petersaxton9007]

As you get closer and closer to an object, can you compute the electromagnetic attraction to inquire whether or not proximity at very small values can produce quantum entanglement between proximal objects?

[@stephengarside8928]

Is science solvable? Or, does science need to be solved?

[@TheMemesofDestruction]

“It’s time to play the Game!” ^.^

[@Jediluvs2kill]

No ai cannot AGI CAN

[@Dr.acai.jr.]

Corporation ai solve science is a Turing type question. Pathetic, kind of.

[@dg-ov4cf]

none of this made any sense