The Swarm Edition

Editor’s Note:

I’m going to assume that you are seeing plenty of AI content (news and papers, not AI slop, though I’m sure you’re getting plenty of that too) from other sources, so I’m trying to think hard about the marginal value that the faiV can provide. As far as I’ve gotten so far: a) make sure that directly AI-related content does not take up more than 2 of the 5 items, and b) try to focus specifically on the cybersecurity issue for financial inclusion. Beyond that, I’m open to feedback on the value you get from the faiV in relation to AI to help guide us in the future. On a different note, I’ll be making a brief cameo appearance at the e-MFP Inclusive Finance Conference to moderate the opening plenary. If you haven’t registered, do so and I hope to see you there!

- Tim Ogden

1. AI AND DDOS

In the last faiV, I described my main p(Doom) scenario/concern not as extinction of humanity but of a massive distributed denial of service (DDoS) “attack” causing all sorts of systems and institutions to break down. Since then agent swarms have been making the news plenty. Here’s a story about how Muse (Meta’s agent) has been overwhelming lots of systems, with the notable phrase “agentic flooding.” Here’s a higher-level view of what is happening and what needs to happen to keep agent swarms on the rails.

But back to the specific DDoS implications: in the near-term we’re going to have to quickly deal with creating the rules, norms and systems. Patrick Collison has some thoughts worth reading about how AI agents change a lot of the rules and expectations for commerce. He’s far more sanguine than I am about the core question of who controls the agents (see also item 4 below) and what is the least costly adaptive mechanism. I think specifically of two things: 1) how we learned that manufacturing uniform demand was cheaper than the “mass customization” that many thought would be the outcome of internet commerce, and 2) how the people thinking about regulating and planning for these changes are the most likely to remain in control of their AI and agents (cf. Patrick McKenzie on how the designers and regulators of low-income financial services are never users of those services). But we also have to worry about things like courts, access to justice and settled rules and standards (“AI: the cause of, and solution to, all life’s problems”), not to mention AI-generated video victim statements, AI driving up costs in disputes between companies, and indications that AI use reduces people’s willingness to admit that they “don’t know.” Overconfidence of humans using AI is as much a DDoS mode as agentic flooding. 

2. CYBERSECUITY

The future is already here, it’s just unevenly distributed. The case of South Korean banks. And French grocery stores. I can’t wrap my head around anyone who doesn’t think that many small-scale banks and financial services providers are already compromised. Of specific note, the person who perpetrated that attack on South Korean banks was apparently a total amateur who didn’t know how to take even the most elementary steps to cover his tracks. Maybe the view of someone on the inside of OpenAI will help change minds and sharpen focus? Probably not.

Anthropic notes that one of the new Chinese open-weight (i.e., no meaningful guardrails) models has cybersecurity capabilities not far behind the most advanced Anthropic models. And the recent AI math advances that have been publicly shared don’t include anything directly related to cryptography. That’s probably not because there are no advances there. To connect the dots, if you have received a notice from some company that some data of yours they held was stolen, you should assume that the hackers also got a hold of more sensitive information that was encrypted but won’t be for long. 

Even Satya Nadella thinks that companies are so far from being able to control the advanced AI models they have already deployed that the stance should be to treat them as an “insider risk," and consider them compromised from the start. This is fine. 

3. DEVELOPMENT

Back to the wheelhouse I suppose. There are new Nobel laureates in economics but I have nothing to say about them. I did have something to say about an existing Nobel laureate last week in the new magazine In Development (you should subscribe!). In recognition of Michael Kremer’s appointment as the new Chief Economist of the World Bank, they reprinted my interview with Michael, conducted about a decade ago for Experimental Conversations, along with some brief reflections after reading the interview again in current circumstances. As I said elsewhere, if you’re really, really lucky, something you did in the before times might still be relevant today (as long as it involves a superstar). 

In Development has another piece worth reading about the “current circumstances:” Lant Pritchett’s view of the “political suicide” committed by aid agencies. There’s a bit of motte (aid requires political support from the foreign policy establishment) and bailey (pay no attention to the aid-enabled permanent kleptocracy that is useful to the foreign policy establishment). I think everyone can agree that a one-legged stool is unstable, but so is a two-legged one. 

Apropos of Kremer and Pritchett, I do worry about reigniting the RCT wars of the past. Asterisk has a piece critical of the ascendance of RCTs in the evidence pantheon. There’s a new Science paper critiquing the use of one RCT to set policy, which I think everyone would agree is bad. Unbelievably, the example featured in the article is a supposed reliance on an RCT of microcredit impact to justify scaling microcredit. I expect most readers of the faiV to know that no such RCT existed before microcredit scaled, and the lack of credible causal evidence on microcredit was a major driver of the RCT revolution but apparently the editors and peer reviewers at Science don’t. If only they’d read Experimental Conversations! Just as frustrating to me is the appeal to clinical medicine as a standard—anyone with a passing familiarity with clinical practice knows that development economics is light-years ahead on issues like external validity, selection and ascertainment bias, and LATEs. 

4. GAMBLING

As with cybersecurity, this is apparently a drum I’m going to keep banging, because I perceive a lack of urgency in response to foreseeable harms. At least half of the volume of prediction markets like Kalshi and Polymarket appears to be amateurs placing money on long-shot bets that lose 98% of the time (e.g., LeBron winning the next presidential election), essentially just transferring money from their bank accounts to those of professionals. It’s all fun and games, the industry protests. Except that the new gambling enterprises are looking for soon-to-be problem gamblers and facilitating their descent or helping existing ones relapse. This is not a US-only phenomenon. “Africa’s CEOs Warn Online Gambling Is Draining Customer Wallets.”

5. SMALL FIRMS

I do still spend a lot of time thinking about small firms in low-income areas, even though I’m constantly distracted by worries about cybersecurity and unregulated gambling. I got to visit Muhammad Meki, Simon Quinn and Chris Woodruff in Oxford recently to talk about small-firm labor markets and the direction of causality between labor demand and market demand, and the interplay with returns (or lack thereof) to managerial practices and skills. We didn’t come to any conclusions, but we’re not the only ones struggling to make sense of it all. Alvaro Gonzalez has a piece at CGD on the perpetual reorganization of the World Bank’s Private Sector Development work because of lack of clarity on goals. He proposes a three-part test: diagnose a binding market failure, address the distortion, and change a market outcome. Hey, that’s what we’re trying to do with the Small Firm Diaries!

You might think that I think that AI is all bad, but that’s not the case. I’m genuinely open to the idea that AI applied to helping people make decisions when information is hard to come by or process could be really great, for instance this AI agent for microenterprises in Honduras (we’re supporting similar ideas in Colombia via ANIF). The idea that such interventions can truly be helpful is bolstered by this new QJE paper on “retrieval failures”—people have a hard time remembering and planning around small, volatile expenditures—and when you help them remember it leads to positive investment outcomes. Emma Riley and I will have something to say about this in the very near future (at NEUDC, give me a shout if you’ll be there) from an experiment we ran as part of the Small Firm Diaries.

VIDEO OF THE WEEK

@jomatech: Introducing ChatGPT for Dishwashing