Incentives and coordination to solve hard problems \ stacker news ~mostly

Incentives and coordination to solve hard problems

2905 sats \ 21 comments \ @elvismercury 9 Mar mostly_harmless

Here's an article you might have come across in the last couple days. It's about how a group came up with a way to train quite large LLMs (70b params) on equipment that a normal person could have in her basement. It turns out there was like a 10x price / performance gain possible but that nobody was incentivized to do it:

Solving this problem is hard. It requires understanding many separate libraries (e.g bitsandbytes, PEFT, Transformers, Accelerate, and PyTorch), and computer science and math concepts (e.g discretization, distributed computing, GPU programming, linear algebra, SGD concepts such as gradient checkpointing), and how they all interact.

So an eclectic set of skills requiring deep education and high intelligence. But the world is full of those problems and those kinds of people! And yet something about this was different:

Academia is full of brilliant people that solve hard problems. But academia hasn’t solved this particular problem. That’s because it’s difficult for university researchers to justify spending time on this kind of work. Combining existing tools and techniques together isn’t generally considered “novel” enough to result in publication in a high impact journal, but that’s the currency that academics need. Furthermore, academics are generally expected to become highly specialized within their field, making it challenging to bring together so many pieces into a single solution.

And, of course, big tech companies are also full of brilliant people that solve hard problems. But this particular problem, training models with consumer GPUs, isn’t a problem they need to solve – they’ve already bought the big expensive GPUs!

Many startups are also full of brilliant people that solve hard problems! But, as Eric Ries explains, “today’s financial market forces businesses to prioritize short-term gains over everything else”. It’s extremely hard for a startup to justify to investors why they’re spending their funds on open source software and public research.

So essentially, there was a problem, it was worth solving, it would be useful to lots of people to have it solved, but the nature of the problem, and the coordination required to solve it, and the system of incentives operating across the various groups that would have to coordinate to solve a problem like this, did not align. The problem dropped between the outfielders of the existing structure.

This happens all the time, of course. It's just interesting to see it happen about this topic, right now, and for these particular reasons. It makes you ask the obvious question of what other important yet solveable problems are embedded in an ecosystem that can't solve them, even though it would be good for everyone if it did.

And the next obvious question, of how one might introduce a meta-system atop the existing system that could do better.

view all related items

372 sats \ 7 replies \ @davidw 9 Mar

So true, great post. There are many examples of this in society. Like @kr discussed in a recent podcast we could have had flying cars already. We could also have had abundant chemical & salt-free water from our oceans and water-powered engines & vehicles.

It’s not just the incentives to build these innovative products don’t exist, it’s that the existing power structures in place have very very strong incentives to ‘neut’ or prevent them from taking off & gaining traction.

The reason we have so many plastics in our society is not because we can’t create biodegradable packaging, just like the reason we can’t have natural pharmaceutical products with no petrochemicals in them is not because it’s hard but because the oil giants have every incentive to monetise their waste products and impose their usage on society.

When you start thinking about it, there is sooo much suppression of innovation and technology in today’s society. When the cap is lifted and these dark incentives cease to retain such power, there’s going to be such a huge economic boom IMHO.

150 sats \ 4 replies \ @k00b 9 Mar

I haven't listened to the podcast, but I don't feel like anyone in particular is wielding the technology suppression gun. It's like the suppression gun emerges from the system design and then people on the right side of the gun all fight to maintain and strengthen it. So the question for me is: how to design a system that doesn't create technology suppression guns?

21 sats \ 3 replies \ @elvismercury OP 9 Mar

This is my overwhelming feeling, also. Systems work to propagate themselves at every level, from the cellular to the civilization. Which, if true, has sobering implications for what we can expect even if btc succeeds wildly.

Certain problems will be eternal, that's my prediction. But it's hard to say exactly which ones.

216 sats \ 2 replies \ @davidw 9 Mar

Economies of scale has been a way to fast-forward efficiency and the side effect has been some of these themes & symptoms.

More money = more firepower = more economies of scale = more marketing = more customers = more money.

In the future, it may actually be a drag having all of that capital overhead, when people in their teenage basements and with a creative mind can will any one of their 1,000+ creations into existence.

I’m specifically wondering whether energy fusion (or abundance), nanotech, AI, open source and hard money (bitcoin) combined will break down the economies of scale in products & business, that we have witnessed for the past 100 years.

Not sure how true it is but I heard a report last week that drones costing several thousand dollars each were overwhelming military equipment worth tens of millions. That’s just one anecdote but I think it represents a shift in balance. Same for ‘journalism’ and individual researchers / writers / analysts punching above their weight and out-competing these big newsrooms. But I digress.

I wonder if the wheels are already in motion to less suppressed technology. If so, just think what that unleashes.

50 sats \ 1 reply \ @elvismercury OP 9 Mar

I’m specifically wondering whether fusion, nanotech, AI, open source and hard money (bitcoin) combined will break down the economies of scale in products & business, that we have witnessed for the past 100 years.

That's a fascinating thing to think about -- the more you can pull disparate things into one single brain, s.t. the coordination cost is basically zero, the more advantage you have. LLMs do this more than anything has in my lifetime, so I think the cocktail you describe could be a plausible erosion -- at least temporarily -- roll back the transaction cost advantages of the firm, at least for a while.

10 sats \ 0 replies \ @davidw 9 Mar

Yes, civilisation has peaks and troughs, ebbs and flows of increasing centralisation vs decentralisation. So it certainly won’t be forever.

But it might be a pretty cool time to be a small vendor or small business owner again. At least that’s my hope.

Super excited to see some of these ridiculous and groundbreaking inventions go mainstream. Not just turning something analogue digital (which is basically the last 30 years). Hoprfully real sci-fi-looking progress.

21 sats \ 1 reply \ @Bell_curve 9 Mar

Suppression of innovation by government or monopolies?

118 sats \ 0 replies \ @davidw 9 Mar

By both. So much misallocation of capital. It may not be directly desired, but because of that one simple intervention… it’s my suspicion that it reinforces the existing power structures and prevents them from needing to compete against potentially superior alternatives.

Pulling down such vast resources from the future, extends existing giants reach and allows them to market their inferior products to a wider audience, grow their share of the pie and change consumer attitudes/behaviour. That in a hard money standard would be extremely difficult or at least significantly harder to accomplish.

308 sats \ 3 replies \ @Undisciplined 9 Mar

academics are generally expected to become highly specialized within their field, making it challenging to bring together so many pieces into a single solution.

That problem is the root of my nym.

Another, similar problem, is that academics focus on small questions that can be answered precisely, because that's how you get published. Attempting big, paradigm shifting work upsets potential reviewers and can't usually be done with the same precision in a single article.

Basically, everyone is just tweaking the parameters of the prevailing model their field adopted decades ago, rather than exploring different models.

26 sats \ 2 replies \ @elvismercury OP 9 Mar

Excessively Pedantic Guy (was that his name?): not everyone is just tweaking -- innovation still occurs, but it's not what's generally incentivized. Even so, it eventually breaks through (see e.g., the rise of deep learning, from Hinton and others relentlessly pushing it forward, through decades of neural methods being in disfavor).

But yeah, the incentives are universally bad, as you say. If innovation breaks out, it's in spite of them.

27 sats \ 1 reply \ @Undisciplined 9 Mar

I agree with your point. I just remember how shocking it was when someone pointed out that pretty much no scientific field has had a major breakthrough in the past 50 years, but many had their major breakthroughs in the 50 years before that. Progress in our understanding of the universe has largely stagnated.

Needlessly Pedantic Man: The "basically" that preceded "everyone" in my previous comment implied that the statement was an inexact summary.

26 sats \ 0 replies \ @elvismercury OP 9 Mar

Implication is irrelevant to Needlessly Pedantic Man.

218 sats \ 5 replies \ @k00b 9 Mar

I missed this article. Reading about the gap their org fills was like a whiff of smelling salts. When did we largely retire private, independent labs like this?

I'll be reading their Lessons from history’s greatest R&D labs soon. OpenAI started in a similar lab form but they were pulled another way.

how one might introduce a meta-system atop the existing system that could do better

I always reach for incentives when coordination problems arise, but its not clear to me how they'd work here. Loosely, financial incentives gives us companies, reputation incentives give us academia, and R&D labs are both. If we aren't creating R&D labs, our culture/economy is leaving out blended incentive systems.

You could incentivize creation of new incentive systems so these gaps never form. But how would you design such a thing to produce R&D labs? Is the good that R&D labs do measurable in a short enough time period to reward them?

100 sats \ 0 replies \ @elvismercury OP 9 Mar

I think there might be a Goedel's (tried for one whole minute to make the umlaut, failed) analogue here, that any incentive system is 'incomplete' in the kinds of expressions that can arise from it. My working theory is that one can't solve this problem, the best one can do is come up with some Schumpeterian arrangement that encourages the wrong ones to die, but also doesn't make the death of a company into the death of the people who constitute it.

Easier said than done, though. The failure modes on both extrema are clear, and bad.

108 sats \ 3 replies \ @k00b 9 Mar

I left out intrinsic incentives, which we don't seem to account for in our societal system design much, but maybe that's part of the problem.

I'm completely untrained here, but my understanding is that we partition motivation into intrinsic and extrinsic forms. Financial and reputational motivations are extrinsic subforms. What are the intrinsic subforms?

100 sats \ 1 reply \ @elvismercury OP 9 Mar

Ha! Yes. This is an extra-hard problem, I think, because:

a) economic incentives are not only generally useless for vitally important classes of endeavor, but counter-productive to it, and

b) we have no good way of non-economic incentive-design that our civilization can manage anymore -- religion being both perverted and largely ineffectual to current sensibilities; and nothing else seriously on offer.

0 sats \ 0 replies \ @k00b 9 Mar

R&D labs are mostly motivated by these non-extrinsic incentives too as far as I can tell. If a stable financial life is harder to achieve now than it was in the golden age of R&D labs, maybe we are scrambling too much to pursue intrinsic things. That's probably my bias though.

Another view: the hard problems, absent AI, were so hard R&D couldn't reliably spinoff extrinsic things so we stopped building R&D labs.

This is sort of consistent with one of Theil's theories for why we are stagnating.

0 sats \ 0 replies \ @Bell_curve 9 Mar

It’s hard to separate intrinsic and extrinsic motivation

108 sats \ 2 replies \ @random_ 9 Mar

And the next obvious question, of how one might introduce a meta-system atop the existing system that could do better.

0 sats \ 1 reply \ @k00b 9 Mar

Interesting. What's the through line between these two? It seems like they might connect but how?

10 sats \ 0 replies \ @random_ 9 Mar

IMO the through line is the proof-of-useful-work.

I attended a talk about this paper and from what I remember, it's a way to coordinate auctions for CO problems you need solved. Not sure how it could be applied to abstract problems like 'develop a method to train llm on consumer grade hardware', but what if?