Operationalising the bitter lessons in compute and cleverness

1 Intelligence as resource scarcity

See Agency and bounded compute for a more detailed discussion of the problem of bounded rationality and the various ways it has been formalized.

2 At scale

The above ideas don’t directly address the economic motivation at the start — the substitution between training and inference, the amortization of a foundation model across millions of users, the trade-off between data acquisition cost and model complexity. That seems to require something genuinely economic, not just decision-theoretic.

How much can one kind of computation substitute for another? It might be something “deeper” like a fundamental theory of intelligence.

On the other hand, it feels like just treating the problem as an economic one might have some mileage. It looks like some factors are fungible (accuracy, storage, compute, generalization…) but also we don’t have anything like the “utility” of the decision maker; in fact we just spent a lot of effort arguing our way out of having a conventional utility. Would we be making progress by keeping utility and augmenting it with a compute-regularized meta utility? Do we need to resign ourselves to intrinsic motivations and/or utilities as local fitness approximations and it will still all work out?

ANYWAY, had we the right framing we might be able to express foundation models as an ingenious amortization strategy — maybe an interesting financial instrument denominated in the currency of “cognition,” or maybe just a classic production cost curve with unusual returns to scale. I mean, most of these things still cash out in the margins in dollars (or joules?).

It seems like, since the different ways of computing clearly have some kind of substitutability, microeconomic models of them are not hopeless. But some stuff is still weird; the exchange rate between FLOPS and synaptic updates I suspect arises from some deeply contingent market equilibrium rather than anything fundamental about the physics of computation.

Maybe at this point I need to impersonate a real economist?

3 Amortization

Amortization is what Bayesians call the trade-off involved in learning inferential shortcuts; there is a whole parallel literature on this in probabilistic NNs and variational inference.

Just because that is where I ran into it doesn’t mean we are committed to Bayes though. Case in point: SGD is a kind of amortized optimization. We could notionally solve the optimization problem at inference time by doing a full-batch gradient descent, but that would be very slow. Instead, we amortize the cost of optimization into a training phase, where we do many cheap updates to learn a good set of parameters that can be evaluated cheaply at inference time.

RL, if you squint at it just so, can be seen as a kind of amortized dynamic programming. We could spend a lot of compute at training time to learn a policy that is very cheap to execute at inference time. In many RL algorithms, notably ones trained on pure physical models, this is a naked attempt to speed up a thing that we could in principle already calculate. We could notionally calculate the optimal action at each time step by solving a dynamic programming problem via some expensive simulator, but that would be very slow. Instead, we learn a policy that approximates the optimal action — amortizing the cost of dynamic programming, and maybe even physical simulation, into a single forward pass.

In practice RL+SGD in RLHF adds a second compute budget on top of pretraining. Toby Ord argued that RL training was nearing its effective limit by entering an unfavourable scaling regime: “we may have lost the ability to effectively turn more compute into more intelligence.” That doesn’t seem to have tanked progress in market substitution of machine for human labour, though, so I suspect something else is going on.

🚧TODO🚧: training vs. inference substitution more generally; memorization as a form of amortization.

4 The bitter lesson and its discontents

Sutton’s bitter lesson is now folk wisdom:

The biggest lesson that can be read from 70 years of AI research is that general methods that leverage computation are ultimately the most effective, and by a large margin.

In the substitution language above, this is a claim about a particular slope on the curve: at the margin, compute substitutes for cleverness at a generous exchange rate. The slope is well-attested in the data-rich regime. Algorithmic progress has outpaced classical hardware efficiency (Grace 2013); the compute needed to reach AlexNet-level ImageNet performance has halved every 16 months since 2012 (Hernandez and Brown 2020), faster than Moore’s Law would predict. Scaling laws tell the same story in a different vocabulary.

But that slope is the local gradient in one regime. In others it flips. On my last hydrology project a single data point cost about USD 500,000, because drilling a thousand-metre well cost that much in that spot. Telling me to collect a billion more data points instead of “being clever” wasn’t an option; collapsing the global economy to collect them would not have been “clever” in a useful way. The same goes wherever data acquisition is the constraint — narrow scientific domains, embodied agents, real-time decisions in novel settings. Foundation-model pretraining moves us along the substitution curve: pay compute up front to manufacture training-data-like output where data was scarce. Pretraining a geospatial foundation model on hydrology-adjacent tasks for example, is SOTA on geospatial domains. That gets us a long way, but it is still not clear how much water is in the damn well.

The folk bitter lesson should be read not so much as a homily as an empirical claim about the slope of a substitution curve in the data-rich regime. The slope might be different elsewhere. What we generally want is the curve itself, parameterized by data cost, compute cost, and the kind of generalization we need. That’s the best lesson; the bitter one is a corollary in one of its regions.

Proposal: we should call the substitution curve a slop slope.

Ermin Orhan gives a tighter and more concrete reading of where in the data-rich regime to look — probing capability limits as data scales, learning from new modalities, designing architectures that exploit both data and compute.¹

5 Human bottlenecks

A second implication of the bitter lesson: even the best human minds aren’t very efficient as general-purpose symbol processors; the quickest path to intelligence prioritises bypassing that human bottleneck.

Things work better when they can scale up at a favourable rate. Humans don’t scale up at many points — not at the scale of an individual organism, and not at the scale of a society either.

🚧TODO🚧: this is a stub; the human-substrate bits seemed to me like I had an angle of attack, but, I am not sure what this was supposed to be

6 Bitter lessons in career strategy

At my point on the substitution curve, I get the most from the limited compute in my skull by figuring out how to use the larger compute on my GPU.

Career version:

…a lot of the good ideas that did not require a massive compute budget have already been published by smart people who did not have GPUs, so we need to leverage our technological advantage relative to those ancestors if we want to get cited.

7 Incoming

• Will the Need to Retrain AI Models from Scratch Block a Software Intelligence Explosion?

• IsoFLOP curves of large language models are extremely flat – Severely Theoretical

• A Better Lesson – Rodney Brooks

• Compute Goes Brrr: Revisiting Sutton’s Bitter Lesson for Artificial Intelligence

• Reflections on ‘The Bitter Lesson’

• Gwern

  • On GPT-3: Meta-Learning, Scaling, Implications, and Deep Theory
  • The Scaling Hypothesis

• Scale, schlep, and systems

• Trading off compute against humans — partially in economics of LLMs

• Epistemic bottlenecks — information transmission and compression

• Thermodynamics of computation: material costs, statistical mechanics of statistics

8 References