AI link dump for September 17th 2024
Researchers use outdated AI tech; Terence Tao & o1-preview; applying AI to industrial processes; AI agents & stablecoins threaten traditional payment rails
Researchers use outdated AI tech
Some researchers have determined that large language models can help persuade people not to believe in conspiracy theories. The abstract to their paper reads:
Conspiracy theory beliefs are notoriously persistent. Influential hypotheses propose that they fulfill important psychological needs, thus resisting counterevidence. Yet previous failures in correcting conspiracy beliefs may be due to counterevidence being insufficiently compelling and tailored. To evaluate this possibility, we leveraged developments in generative artificial intelligence and engaged 2190 conspiracy believers in personalized evidence-based dialogues with GPT-4 Turbo. The intervention reduced conspiracy belief by ~20%. The effect remained 2 months later, generalized across a wide range of conspiracy theories, and occurred even among participants with deeply entrenched beliefs. Although the dialogues focused on a single conspiracy, they nonetheless diminished belief in unrelated conspiracies and shifted conspiracy-related behavioral intentions. These findings suggest that many conspiracy theory believers can revise their views if presented with sufficiently compelling evidence.
The abstract highlights an interesting issue with academic research, given the rate at which AI tech is improving. The research was done using one of OpenAI’s older LLMs, GPT-4 Turbo (since that was state of the art, we have seen GPT-4o and o1-preview come out). There is a lag between when research using AI tech is done and when it is published, and that lag means that researchers appear to use outdated tech to reach their conclusions. The obvious answer would be to reduce the lag between when research is done and when research is published, but I suspect that all kinds of institutional imperatives militate against that.
Terence Tao uses o1-preview and has some thoughts
I see a lot of professors complain about their students and AI. No students, they claim, want to think because the students all reach for ChatGPT. While I understand the motivation behind these complains, I can’t help but think that some of their ire arises due to pedagogical authority devolving away from the professor and accruing to AI. If an AI can help a student grasp a concept, what use is the professor? Maybe there is a place for the professor to serve as a complement to the AI. But I don’t think many professors want to view themselves as a complement to anything.
Anyway, given all of this, it is refreshing to see professors who actually wrestle with AI tech, and try to use it towards productive ends. Here’s an interesting post from mathematician Terence Tao about his experience working with o1-preview:
In https://chatgpt.com/share/94152e76-7511-4943-9d99-1118267f4b2b I gave the new model a challenging complex analysis problem (which I had previously asked GPT4 to assist in writing up a proof of in https://chatgpt.com/share/63c5774a-d58a-47c2-9149-362b05e268b4 ). Here the results were better than previous models, but still slightly disappointing: the new model could work its way to a correct (and well-written) solution *if* provided a lot of hints and prodding, but did not generate the key conceptual ideas on its own, and did make some non-trivial mistakes. The experience seemed roughly on par with trying to advise a mediocre, but not completely incompetent, graduate student. However, this was an improvement over previous models, whose capability was closer to an actually incompetent graduate student. It may only take one or two further iterations of improved capability (and integration with other tools, such as computer algebra packages and proof assistants) until the level of "competent graduate student" is reached, at which point I could see this tool being of significant use in research level tasks.
This seems like a pretty important point. Tao claims that it will not be too long before AI tech is “of significant use in research level tasks.” My bet is that the professors who approach AI with his curiosity will fare better than the professors who complain about their students offloading their cognition to silicon. (And, I suspect that this pattern will repeat itself across professions: those who adopt AI tech will fare better than those who simply complain about it.)
Enterprise adoption of AI will be slower than many expect
When it comes to AI, Silicon Valley has become besotted, in the way that only it can, with what I call the Everything, Everywhere, All at Once fallacy. A lot of VCs seem to think that advanced AI will change everything, everywhere, all at once. While I understand the impetus behind their thinking—it would be great for their portfolios if their AI dreams came to fruition—the world is, as ever, more complicated than they’re willing to consider.
And one place where this ought to be obvious is in applying AI to industrial processes. A lot of the people who think that AI will greatly increase productivity, and therefore economic growth, point to its vast potential to optimize legacy industries’ operations. And there’s some logic to their arguments. Many companies generate vast amounts of data about their operations, and then do nothing with that data. Layer an LLM on top of the company’s data, the thinking goes, and watch productivity skyrocket.
The theory is conceptually simple. But the problem is that it is facile. Companies are complex organizations. Sarah Constantin has a great post about this:
Manufacturing process engineers, for nearly a hundred years, have been using data to inform how a factory operates, generally using a framework known as statistical process control. However, in practice, much more data is generated and collected than is actually used. Only a few process variables get tracked, optimized, and/or used as inputs to adjust production processes; the rest are “data exhaust,” to be ignored and maybe deleted. In principle, the “excess” data may be relevant to the facility’s performance, but nobody knows how, and they’re not equipped to find out.
This is why manufacturing/industrial companies will often be skeptical about proposals to “use AI” to optimize their operations. To “use AI”, you need to build a model around a big dataset. And they don’t have that dataset.
And, even when the company retains its data, and hires your company to implement an AI solution, sometimes you can’t get access to the data:
in the semiconductor industry, everyone is justifiably paranoid about industrial espionage. They are not putting their factory data “on the cloud.” They may have fully airgapped facilities where nothing is connected to the open internet. They do not want images of in-progress chips, or details of their production processes, getting into “the wrong hands.”
…
Sometimes complying with security requirements for data sharing is simple; but the larger the company, the more likely you are to encounter multiple IT policies from different departments….
Sometimes people are worried that a “big data” tool will replace their jobs, or make thier performance look bad, and they’re trying to obstruct the process.
She goes on this vein for quite a bit—the post is long, but well worth reading. Far too many Silicon Valley executives and venture capitalists have spent no time working at legacy companies, and so just assume that when they show up with fancy new tech, companies will readily adopt it. But the play that works in B2B SaaS for selling, say, marketing tech to consumer goods companies, doesn’t work for manufacturing or industrial companies.
So, while the argument that AI will rapidly grow the economy sounds reasonable in principle, the boots-on-the-ground reality is much different. AI will be adopted by enterprises, and enterprises will benefit, significantly, from AI, but the adoption curve will be much slower than the febrile VC class expects. And, though this observation may be bad news for the VCs’ portfolios, it is, over the longer term, very good news for the US economy.
Nonetheless, the point remains: far too many people see the rapid improvement in AI tech over the past several years, and mistakenly assume that enterprise adoption will be similarly fast. That assumption just doesn’t hold, for a variety of fundamental, and poorly understood, reasons that Sarah explains in great detail.
AI agents and stablecoins will kill traditional payments rails
This last item isn’t a link, but rather my thoughts on the threat that autonomous AI agents using crypto as a payment rail poses to traditional payments companies like VISA and Mastercard. I previously wrote that, since AI agents can’t open bank accounts, they will use crypto channels to facilitate payment for transactions. (I also wrote a separate post about counterarguments to this claim, but I don’t think the counterarguments are very good.)
The biggest threat to VISA and Mastercard is that AI agents will use stablecoins like USDC to facilitate payment for transactions. Most Wall St. executives, to say nothing of the executives who work at these companies, have no clue about this. So let’s look at why AI agents using stablecoins poses such a big threat to the traditional payment processors.
Disintermediation. Stablecoins, when integrated with AI agents, allow one to bypass traditional intermediaries like Visa and Mastercard. Once businesses catch on to the lower costs, they will pursue this aggressively.
Instant settlement. Stablecoins enable near-instant settlement, compared to traditional payment systems. AI agents, particularly those which execute high-volume, high-frequency transactions, such as in the IoT space, would benefit from this speed.
Cross-border payments: Stablecoins allow for seamless cross-border payments without the friction of currency conversions and the associated fees. AI agents operating globally would have a distinct advantage using stablecoins, since they would bypass the complex web of currency regulations and restrictions that traditional networks face.1
Programmability. AI agents will use smart contracts to enforce programmable payments, such as recurring micropayments or conditional payments based on predefined outcomes. This kind of transactional complexity is not easily supported by traditional payment rails. And, to the extent that recurring payments are supported by traditional payment rails, the payment processors extract fees every time the fee is paid.
Lower fees. The fees associated with stablecoin transactions, especially on Layer 2 solutions, are minuscule compared to the fees charged by credit card processors. As AI agents optimize for efficiency and cost, they’ll use stablecoins.
Your average Wall St executive isn’t focused on this. They’re still tied to traditional financial systems, and they don’t grasp how rapidly AI and crypto are evolving together.
One might reasonably argue that though stablecoins don’t yet face the onerous regulations that the traditional currency markets face, they will in the future. And that’s true! At the same time, regulators tend to respond reactively, not proactively, and both decentralized finance and artificial intelligence technologies are rapidly improving and diffusing throughout the world. At some point, these systems become so pervasive that regulators’ only choice is to impose rules, rather than outright bans.