Economics of large language models

Various questions about economics of social changes wrought by ready access to LLMs, the latest generation of automation. This is the major “short”-“medium” effect whatever those words mean. Longer-sighted persons might also care about whether AI will replace us with grey goo.

Thought had in conversation with Richard Scalzo about Smith (2022):

“Snowmobile or bicycle?” / “Complement or substitute?”

Is the AI we have complementary or competitive tech? At an individual level it can be either, augmenting or replacing human endeavour.

At a social scale we can have other dynamics come into play.

Economics of collective intelligence

How do foundation models/ large language models change the economics of knowledge production? Of art production? To a first order approximation (valid at 03/2023) LLMs provide a way of massively compressing collective knowledge and synthesising the bits I need on demand. They are not yet directly generating novel knowledge (whatever that means). But they do seem to be pretty good at being “nearly as smart as everyone on the internet combined”. There is no sharp boundary between these ideas, clearly.

Deploying these models will test various hypotheses about how much of collective knowledge depends upon our participating in boring boilerplate grunt work, and what incentives are necessary to encourage us to produce and share our individual contributions to that collective intelligence.

Historically, there was a strong incentive to open publishing. In a world where LLMs are effective at using all openly published knowledge, we should perhaps expect to see a shift towards more closed publishing, secret knowledge, hidden data, and away from reproducible research, open source software, and open data, since publishing those things will be more likely to erode your competitive advantage.

Generally, will we wish to share truth and science in the future, or will the economic incentives switch us towards a fragmentation of reality into competing narratives, each with their own private knowledge and secret sauce?

Consider the incentives for humans to tap out of the tedious work of being themselves in favour of AI emulators: The people paid to train AI are outsourcing their work… to AI. Which makes models worse (Shumailov et al. 2023). Read on for more of that.

Organisational behaviour

There is a theory of career moats, which are, basically, unique value propositions that only you have that make you, personally, unsackable. I’m quite fond of Cedric Chin’s writing on this theme, which is often about developing skills that are valuable. But he (and organisational literature generally) acknowledges there are other ways of making sure you are unsackable which are less pro-social — attaining power over resources, becoming gatekeeper, opaque decision making etc.

Both these strategies co-exist in organisations generally, but I think that LLMs, by automating skills and knowledge, will tilt incentives towards the latter. It is rational in this scenario for us to think less about how well we can use our skills and command of open (e.g. scientific, technical) knowledge to be effective, and rather, for us each to focus on how we can privatise or sequester secret knowledge to which we control exclusive access if we want to show a value add to the org.

How would that shape an organisation, especially a scientific employer? Longer term, I would expect to see a shift (in terms both of who is promoted and how staff personally spend time) from skill development and collaboration, and more towards resource-control, competition and privatisation: less scientific publication, less open documentation of processes, less time doing research and more time doing funding applications, more processes involving service desk tickets to speak to an expert whose knowledge resides in documents that you cannot see, etc.

Is this tilting us towards a Molochian equilibrium?

What to spend my time on

Economics of production at a microscopic, individual scale:

GPT and the Economics of Cognitively Costly Writing Tasks

To analyze the effect of GPT-4 on labor efficiency and the optimal mix of capital to labor for workers who are good at using GPT versus those who aren’t when it comes to performing cognitively costly tasks, we will consider the Goldin and Katz modified Cobb-Douglas production function…

Alternate take: Think of Matt Might’s iconic illustrated guide to a Ph.D..

Imagine a circle that contains all of human knowledge:

By the time you finish elementary school, you know a little:

A master’s degree deepens that specialty:

Reading research papers takes you to the edge of human knowledge:

You push at the boundary for a few years:

Until one day, the boundary gives way:

And, that dent you’ve made is called a Ph.D.:

Here’s my question: Does the new map look something like this? If so is that a problem?

Now OpenAI have shipped an LLM and where is the border?

Content landscape, culture landscape, spamularity, dark forest, textpocalypse

Hmm let me see; devastate, develop or devour?

Charlie Stross’s 2010 Spamularity stuck with me:

We are currently in the early days of an arms race, between the spammers and the authors of spam filters. The spammers are writing software to generate personalized, individualized wrappers for their advertising payloads that masquerade as legitimate communications. The spam cops are writing filters that automate the process of distinguishing a genuinely interesting human communication from the random effusions of a ’bot. And with each iteration, the spam gets more subtly targeted, and the spam filters get better at distinguishing human beings from software, in a bizarre parody of the imitation game popularized by Alan Turing (in which a human being tries to distinguish between another human being and a piece conversational software via textual communication) — an early ad hoc attempt to invent a pragmatic test for artificial intelligence.

We have one faction that is attempting to write software that can generate messages that can pass a Turing test, and another faction that is attempting to write software that can administer an ad-hoc Turing test. Each faction has a strong incentive to beat the other. This is the classic pattern of an evolutionary predator/prey arms race: and so I deduce that if symbol-handling, linguistic artificial intelligence is possible at all, we are on course for a very odd destination indeed—the Spamularity, in which those curious lumps of communicating meat give rise to a meta-sphere of discourse dominated by parasitic viral payloads pretending to be meat…

Maggie Appleton’s commentary on Dark Forest theory of the Internet by Yancey Strickler is another neat framing:

The dark forest theory of the web points to the increasingly life-like but life-less state of being online. Most open and publicly available spaces on the web are overrun with bots, advertisers, trolls, data scrapers, clickbait, keyword-stuffing “content creators,” and algorithmically manipulated junk.

It’s like a dark forest that seems eerily devoid of human life—all the living creatures are hidden beneath the ground or up in trees. If they reveal themselves, they risk being attacked by automated predators.

Humans who want to engage in informal, unoptimised, personal interactions have to hide in closed spaces like invite-only Slack channels, Discord groups, email newsletters, small-scale blogs, and digital gardens. Or make themselves illegible and algorithmically incoherent in public venues.

I feel like I’m going to lose this battle, but for the record, I despise the term “textpocalypse”.

Information landscape

Fake news, credibility, cryptographic verification of provenance, etc. TBC

Economic disparity and LLMs

PR, hype, marketing implications

George Hosu, in a short aside, highlights the incredible marketing advantage of AI:

People that failed to lift a finger to integrate better-than-doctors or work-with-doctors supervised medical models for half a century are stoked at a chatbot being as good as an average doctor and can’t wait to get it to triage patients

The Tweet that Sank $100bn

Google’s Bard was undone on day two by an inaccurate response in the demo video where it suggested that the James Webb Space Telescope would take the first images of exoplanets. This sounds like something the JWST would do but it’s not at all true So one tweet from an astrophysicist sank Alphabet’s value by 9%. This says a lot about how a) LLMs are like* being at the pub with friends, it can say things that sound plausible and true enough and no one really needs to check because who cares? Except we do because this is science, not a lads’ night out and b) the insane speculative volatility of this AI bubble that the hype is so razor thin it can be undermined by a tweet with 44 likes.

I had a wonder if there’s any exploration of the ‘thickness’ of hype. Jack Stilgoe suggested looking at Borup et al which is evergreen but I feel like there’s something about the resilience of hype: Like crypto was/is pretty thin in the scheme of things. High levels of hype but frenetic, unstable and quick to collapse. AI has pretty consistent if pulsating hype gradually growing over the years while something like nuclear fusion is super-thick (at least in the popular imagination) – remaining through decades of not-quite-ready and grasping the slightest indication of success. I don’t know, if there’s nothing specifically on this, maybe I should write it one day.


  • What Will Transformers Transform? – Rodney Brooks

  • Tom Stafford on ChatGPT as Ouija bourd

  • Gradient Dissent, a list of reasons that large backpropagation-trained networks might be worrisome. There are some interesting points in there, and some hyperbole. Also: If it were true that there are externalities from backprop networks (i.e. that they are a kind of methodological pollution that produces private benefits but public costs) then what kind of mechanisms should be applied to disincentives them?

  • C&C Against Predictive Optimization

  • Stanford CRFM

    In this post, we evaluate whether major foundation model providers currently comply with these draft requirements and find that they largely do not. Foundation model providers rarely disclose adequate information regarding the data, compute, and deployment of their models as well as the key characteristics of the models themselves. In particular, foundation model providers generally do not comply with draft requirements to describe the use of copyrighted training data, the hardware used and emissions produced in training, and how they evaluate and test models. As a result, we recommend that policymakers prioritize transparency, informed by the AI Act’s requirements. Our assessment demonstrates that it is currently feasible for foundation model providers to comply with the AI Act, and that disclosure related to foundation models’ development, use, and performance would improve transparency in the entire ecosystem.

  • Is it time for the Revenge of the Normies? - by Noah Smith

  • I Do Not Think It Means What You Think It Means: Artificial Intelligence, Cognitive Work & Scale


Andrus, McKane, Sarah Dean, Thomas Krendl Gilbert, Nathan Lambert, and Tom Zick. 2021. AI Development for the Public Interest: From Abstraction Traps to Sociotechnical Risks.” arXiv.
Barke, Shraddha, Michael B. James, and Nadia Polikarpova. 2022. Grounded Copilot: How Programmers Interact with Code-Generating Models.” arXiv.
Bowman, Samuel R. 2023. Eight Things to Know about Large Language Models.” arXiv.
Danaher, John. 2018. Toward an Ethics of AI Assistants: An Initial Framework.” Philosophy & Technology 31 (4): 629–53.
Grossmann, Igor, Matthew Feinberg, Dawn C. Parker, Nicholas A. Christakis, Philip E. Tetlock, and William A. Cunningham. 2023. AI and the Transformation of Social Science Research.” Science 380 (6650): 1108–9.
Métraux, Alfred. 1956. “A Steel Axe That Destroyed a Tribe, as an Anthropologist Sees It.” The UNESCO Courier: A Window Open on the World IX: 26–27.
Pelto, Pertti J. 1973. The snowmobile revolution: technology and social change in the Arctic. Waveland Press.
Shanahan, Murray. 2023. Talking About Large Language Models.” arXiv.
Shumailov, Ilia, Zakhar Shumaylov, Yiren Zhao, Yarin Gal, Nicolas Papernot, and Ross Anderson. 2023. The Curse of Recursion: Training on Generated Data Makes Models Forget.” arXiv.
Smith, Justin E. H. 2022. The Internet Is Not What You Think It Is: A History, a Philosophy, a Warning. Princeton: Princeton University Press.
Spector, Jonathan Michael, this link will open in a new window Link to external site, and Shanshan Ma. 2019. Inquiry and critical thinking skills for the next generation: from artificial intelligence back to human intelligence.” Smart Learning Environments 6 (1): 1–11.
Susskind, Daniel, and Richard Susskind. 2018. The Future of the Professions.” Proceedings of the American Philosophical Society 162 (2): 125–38.

1 comment

I asked ChatGPT to summarize your piece. Since it was not able to "read" the title as it is an image, it misunderstood LLM for "Master of Laws". Hence its first pass summarizing went as follows--verbatim:

"The post is a personal account of the author's experience pursuing a Master of Laws (LL.M) in Economics program. The author discusses the challenges they faced, such as the steep learning curve of economics and the demands of the program. They also highlight the benefits of the program, including exposure to new economic theories, the development of critical thinking skills, and opportunities to engage in policy discussions. The author also provides some advice for prospective LL.M in Economics students, including the importance of prior preparation, the need to develop time-management skills, and the value of networking opportunities."

The metal axe still needs improvement but it will get there with time.

Reply to RHM

GitHub-flavored Markdown & a sane subset of HTML is supported.