Sabitlenmiş Tweet
tom cunningham
685 posts
tom cunningham
@testingham
Economics & AI @ @METR_Evals (ex-openai) https://t.co/FZobuYjdOc
San Francisco, CA Katılım Mart 2009
3.1K Takip Edilen10.3K Takipçiler
@TomDavidsonX Thank you Tom, this makes my day. Huge fan of your work and we’re all working on a trail you blazed.
English
Enjoyed reading this paper!
Some highlights:
- using local elasticities throughout, so they don't need to assume an overall functional form
- allowing for capabilities at AI RnD to improve much faster than capabilities at other economic tasks, which I think is very likely
- defining a unit increase in "ai capabilities" via Epoch's ECI and plugging this into the condition for rsi
- this allows them to add nuance to the "r>1" criteria I've previously outlined, breaking r into three components
- a very long list of things ai companies can measure to forecast rsi!
I hope the authors work with frontier ai companies to identify practical ways to make these measurements happen
tom cunningham@testingham
Very happy to share the first paper from @ElasticityInst: The Economics of Recursive Self-Improvement. Two parts: (1) a graphical representation of feedback loops, to formalize a variety of RSI-related arguments, where each arrow represents responsiveness (elasticity); (2) a survey of existing evidence with a loose calibration & a “wish list” of evidence that would help us calibrate better.
English
Bottleneck on what though?
I should've been clearer in the original tweet but I'm really just thinking about production efficiency, not GDP. We've gone through a few orders of magnitude in input-efficiency in producing food, transport, lighting, computation. Each of those may have a few more orders of magnitude left.
English
@testingham @JimDMiller I think social limits are underrated, limits from pure physics are not necessarily the main bottleneck here
English
When economists talk about "post-AGI futures" they mean abundance: 10X growth.
When AI people talk about "post-AGI futures" they mean omnipotence: growth until physical limits are hit.
A natural question to ask the economists: if AI gets you a century's technological progress in a decade, why will it stop there?
Additionally, from the perspective of omnipotence, the economists considerations all start to seem *provincial*. Considerations about redistribution, inequality, democracy, power concentration, dignity, meaning.
(This observation comes from me trying to reconcile post-AGI conversations at Lighthaven, and post-AGI conversations at Asilomar)
English
@Afinetheorem Fair -- but the 2100 timeframe isn't obvious.
The recent "we must act" statement is talking about an industrial revolution in 10 years.
But if AI gives us one industrial revolution, why not 2, why not 10? We're probably far from the physical limits on efficiency.
English
@JimDMiller fair -- though i think we're pretty far away from the physical limits right now.
English
Economist here: It could be that economic growth stops because AI solves science and social science and figures out optimal arrangements of everything, and we adopt them. There was almost no economic growth for most of humanity's existence because we weren't figuring out new scientific principles that allowed us to get more growth, and we could be in the same situation after we solve science. Of course, at some point you'll have to consider the resources of the entire universe and what kind of growth that would give us although the lightspeed limit greatly limits yearly growth.
English
marvellous work, one couldn’t hope for a better treatment. I think Tom should get a “Such microeconomic effects are self-evidently already occurring” bumper sticker.
tom cunningham@testingham
Very happy to share the first paper from @ElasticityInst: The Economics of Recursive Self-Improvement. Two parts: (1) a graphical representation of feedback loops, to formalize a variety of RSI-related arguments, where each arrow represents responsiveness (elasticity); (2) a survey of existing evidence with a loose calibration & a “wish list” of evidence that would help us calibrate better.
English
tom cunningham retweetledi
New paper on the economics of RSI – and a new org, the Elasticity Institute:
A list of my own takeaways…
tom cunningham@testingham
Very happy to share the first paper from @ElasticityInst: The Economics of Recursive Self-Improvement. Two parts: (1) a graphical representation of feedback loops, to formalize a variety of RSI-related arguments, where each arrow represents responsiveness (elasticity); (2) a survey of existing evidence with a loose calibration & a “wish list” of evidence that would help us calibrate better.
English
@Apoorva__Lal @METR_Evals @ConstellOrg Interesting! Our research is too spicy for corporate firewalls
English
@testingham @METR_Evals @ConstellOrg Oh weird only on the work laptop: "elasticity.institute refused to connect"; probably vpn rejecting a new domain. Found it; this looks great.
English
Very happy to share the first paper from @ElasticityInst: The Economics of Recursive Self-Improvement. Two parts:
(1) a graphical representation of feedback loops, to formalize a variety of RSI-related arguments, where each arrow represents responsiveness (elasticity);
(2) a survey of existing evidence with a loose calibration & a “wish list” of evidence that would help us calibrate better.
English
English
English
English
English
Very happy this statement eventually turned into something far bigger, thanks to @erikbryn , @professor_ajay , & @akorinek:
x.com/erikbryn/statu…
To me the core claim is (1) it's big; (2) it's confusing.
We don't have a good theory of what's going to happen next -- we're driving in the fog -- & I've been spending time trying to persuade the smartest & hardest-working people I know to spend their time on this.
tom cunningham@testingham
I think many economists agree with the following, but it would be valuable to make this publicly known: 1. There is a substantial probability (>10%) that AI will exceed human-level performance on virtually all non-physical tasks within ten years. 2. This would be an unprecedented shock to human society. 3. The economics profession should treat it with an urgency comparable to WWII or COVID.
English
tom cunningham retweetledi
We offer a simple and transparent model of RSI. It uses basic ‘econ 101’ ingredients because we think simplicity is important when the underlying question is complex, scary, and fraught with disagreement. There’s still so much we don’t understand — I’d love to see more work here!
tom cunningham@testingham
Very happy to share the first paper from @ElasticityInst: The Economics of Recursive Self-Improvement. Two parts: (1) a graphical representation of feedback loops, to formalize a variety of RSI-related arguments, where each arrow represents responsiveness (elasticity); (2) a survey of existing evidence with a loose calibration & a “wish list” of evidence that would help us calibrate better.
English
English
> I think it's also possible to be smart & tenacious but end up creating more noise than signal
Agreed! But, all things considered, I also expect AI to make it significantly easier to filter out noise long-term, meaning individual’s long-run contributions might be dominated by their peak outputs, rather than compounding median-level outputs
English
tom cunningham retweetledi
Artificial Intelligence may become the first technology that accelerates its own capability growth—without additional human labor or capital, which haven been key for making AI better.
When AI improves itself, what will happen to economic growth?
tom cunningham@testingham
Very happy to share the first paper from @ElasticityInst: The Economics of Recursive Self-Improvement. Two parts: (1) a graphical representation of feedback loops, to formalize a variety of RSI-related arguments, where each arrow represents responsiveness (elasticity); (2) a survey of existing evidence with a loose calibration & a “wish list” of evidence that would help us calibrate better.
English
tom cunningham retweetledi
Excited to share Elasticity’s first note: The Economics of Recursive Self-Improvement, by @testingham (lead author), @andrewjkoh, @arjun_ramani3, @BasilHalperin, @brian_jabarian, @cherylwoooo, @AlthoffLukas, @whitfill_parker, and @pawtrammell.
Elasticity will continue to produce rigorous economic research on the most important questions raised by advances in AI.
tom cunningham@testingham
Very happy to share the first paper from @ElasticityInst: The Economics of Recursive Self-Improvement. Two parts: (1) a graphical representation of feedback loops, to formalize a variety of RSI-related arguments, where each arrow represents responsiveness (elasticity); (2) a survey of existing evidence with a loose calibration & a “wish list” of evidence that would help us calibrate better.
English
English
@METR_Evals @ConstellOrg Shout out specifically to foundational work on RSI by @TomDavidsonX , @Ph_Aghion , @ChadJonesEcon , @bfjo
English