I actually agree. At some point, a RSI system has to interact with real-world, and that imposes serialization constraints. It is harder to know how much that slow-down would be and how much speed-up we will get before that. But a RSI cannot simply be a exponential growth forever.
It’s not advances on the underlying operation of matrix multiplication that have driven ai progress to date. It’s the layers above that; trying different neural architectures (transformers w/attention mechanisms), and also different data and training regimes (different ways of doing reinforcement learning) that are the main drivers of improved performance. Perpetual motion is a physical impossibility. Whereas Ai is already being used to improve the workflow of ai researchers, thus speeding up improvements in said research. It’s not hard to see that AI could well be spun up to continue to try new arrangements of the aforementioned levers that drive ai progress on its own.
Presumably there's more efficient hardware foundations to perform these efficiently, and potential at the various abstraction layers for more efficiency. Obviously this is not unbounded - simple things would seem to have a physical limit to the potential improvement.
But if you think of the optimization space: different physical representations, different approaches (photo, quantum, etc), more parallelism - there's undoubtedly a lot of headroom even on the matrix multiplication side. I would imagine there's a lot left on the table when it comes to the abstractions we've built. Infinite? No, but lots of potential.
And what does a machine with a few orders of magnitude more power come up with? I'm not readily able to predict what something like that could create (maybe it's tapped out, but I doubt it).
It seems to come down to an article of faith (as referenced in the article) that there's a lot more potential to be extracted in our current exploitation paths. Which I think is probably reasonable.
Heck, even if a theoretical machine tops out at 3-5 orders of magnitude faster/more complex, I'm sure that could do some amazing things that look like magic to us.
The logical business opportunity in the current LLM-boom is to create a bunch of AI-less services and products, and then charge money to access them.
Think of premium branding analogy: masses get cheap AI slop, wealthy get high quality human-curated and human-created produce. Like organic vs regular food.
This only works if your business is large enough that you and all your competitors aren't expected to have humans doing everything already, but small enough that going AI won't boost your valuation by much. My intuition is that the intersection is pretty much empty.
Mark Fisher in Capitalist Realism touches on this concept, where there's a constantly shifting opposition to the market that itself becomes engulfed in its own market, to be advertised.
So for example all the productivity/digital detox channels and videos are themselves a consumer demand to be watched on YouTube, on phones. And now we have anti-AI products marking themselves higher for a feature that didn't previously exist. It's like the tree of capital gets split at every turn.
Blowing up on the pad is incredibly worse from a design data collection perspective, a risk to life perspective, and a downstream impact to future launches perspective (nobody can use that site for a couple of months).
To be fair the last Starship to blowup on launchpad/ground was less than a year ago. It is a set back but it appears nobody has avoided this issue yet.
I think at least one approach that can work is de-globalization of social media into smaller, reputation/trust-ranked social networks. Discord is pretty good in this regard.
If you have suggestions for good Discord servers, please share. The bullshittery is coming out of my ears. I don't quite know where to turn to anymore. There's HN. Reddit is getting a bit crazy, all other social media is a pass for me.
In order to make friends you kind of need to have spaces where you can meet people and trust them enough to make connections
In person is obviously the safest for this. For online friendships I feel the places to meet people you can trust aren't AI or Scammers has shrunk a ton
Actually, for most of its existence, Cyrillic has had a θ (theta) like in Greek, used only in loanwords and pronounced either as ф (f) or т (t) because the th sound is not part of Slavic phonetics. θ was dropped fairly recently - in the 20th century.
It absolutely doesn't as pronounced now, yet Thomas is Фома, Theodor is Феодор, etc. Just like Hertz is Герц, even through Г and H are as far from each other as one can get.
In, say, civil or aerospace engineering, science is understood well enough to allow your building or airplane to be modelled and tested using computer modelling, CAD software, FEM and CFD algorithms and so on. You can design a house or an aircraft without ever building a single physical model, and it will stand (or fly), 99 times out of 100. It is oversimplification to a degree, but sufficiently close approximation.
No such thing exists in biology, pharmaceutics, biotech and so on. The accuracy of computer models and simulation is not sufficient to produce results with single-digit percent accuracy for any metrics, hence long and complex Phase I-II-III trials. Maybe 1 out of 100 candidate drugs works.
Why? Because we do not have the same level of understanding for biological systems as we do for buildings or aircraft, or software. Amount of information is much larger, complexity is far greater, enzymes and cell signalling network make biochemistry extremely non-linear. This makes the problem space vast. It is practically untapped domain and it can eat any amount of computational power and biologists, data scientists and software devs (manpower-wise).
Any incremental improvements in simulation, modelling and interpretation of biological system behaviour will generate downstream improvements in medicine, pharma, biotech. But general-purpose LLM AIs are not that useful in biology, you need more specialized solutions to improve both accuracy and performance of large number of algorithms that have tremendous computational complexity: computational chemistry, molecular dynamics, genomics->proteomics->interactomics->metabolomics (all of that for just intra-cellular behaviour - tissues, organism and organisms are multiple orders of magnitude harder).
But fundamentally it is a problem of missing software to better model biological systems (AI or non-AI). Once created, such a solution will enable large amount of very big breakthroughs in almost every biology-connected discipline.
What would be a way to recursively self-improve algorithms for matrix multiplication (foundations of machine learning and inference)?
reply