Category: Web/Tech

A virtual rodent predicts the structure of neural activity across behaviors

by  Tyler Cowen in

Animals have exquisite control of their bodies, allowing them to perform a diverse range of behaviors. How such control is implemented by the brain, however, remains unclear. Advancing our understanding requires models that can relate principles of control to the structure of neural activity in behaving animals. To facilitate this, we built a ‘virtual rodent’, in which an artificial neural network actuates a biomechanically realistic model of the rat in a physics simulator. We used deep reinforcement learning to train the virtual agent to imitate the behavior of freely-moving rats, thus allowing us to compare neural activity recorded in real rats to the network activity of a virtual rodent mimicking their behavior. We found that neural activity in the sensorimotor striatum and motor cortex was better predicted by the virtual rodent’s network activity than by any features of the real rat’s movements, consistent with both regions implementing inverse dynamics. Furthermore, the network’s latent variability predicted the structure of neural variability across behaviors and afforded robustness in a way consistent with the minimal intervention principle of optimal feedback control. These results demonstrate how physical simulation of biomechanically realistic virtual animals can help interpret the structure of neural activity across behavior and relate it to theoretical principles of motor control.

Here is the new Nature article by Diego Aldarnado, et.al.  Via @sebkrier.

Is Apple paying OpenAI, or vice versa?

by  Tyler Cowen in

I do not know.  Am I wrong to think that in optimal contracts theory there should be payments going in both directions, as there is a dual agency relationship?  Of course optimal contracts theory often fails.

Ben Thompson has a hypothesis (gated):

This sounds like a play to acquire users and mindshare, with the potential of upselling those users to a subscription, i.e. the exact same model that OpenAI has on their website and apps. Moreover, if this partnership entails Apple not paying, it also explains why OpenAI is the only option to start: Google, for example, probably wanted to be paid for Gemini, or Anthropic for Claude, and I can imagine (1) Apple holding the line on not paying, particularly if (2) OpenAI is making an aggressive move to build out its consumer business and be a durable brand and winner in the consumer space. In short, my updated current thinking is that both Apple and OpenAI are making the bet that very large language models are becoming increasingly commoditized, which means that Apple doesn’t have to pay to get access to one, and OpenAI sees scale and consumer mindshare as the best route to a sustainable business.

If I had to guess, I would think the main payment goes from Apple to OpenAI?  We know Apple is very, very good at extracting revenue from customers.  OpenAI, no matter how rapid its growth path over the next few years, simply cannot have a comparable record.  There is thus one model where Apple pays for AI access rights, and later on then charges more for iPhones.  You still will be able to buy cheaper phones without these services, if enough people demand that.  OpenAI of course will be charging you as well, but only if you go over quota, as they face the “extra subscription fatigue issue” in a way that Apple does not.

Just a hypothesis.  The fact that we do not know should tell you something about the limits of economic reasoning.

Terence Tao on AI and mathematics

by  Tyler Cowen in

With formalization projects, what we’ve noticed is that you can collaborate with people who don’t understand the entire mathematics of the entire project, but they understand one tiny little piece. It’s like any modern device. No single person can build a computer on their own, mine all the metals and refine them, and then create the hardware and the software. We have all these specialists, and we have a big logistics supply chain, and eventually we can create a smartphone or whatever. Right now, in a mathematical collaboration, everyone has to know pretty much all the mathematics, and that is a stumbling block, as [Scholze] mentioned. But with these formalizations, it is possible to compartmentalize and contribute to a project only knowing a piece of it. I think also we should start formalizing textbooks. If a textbook is formalized, you can create these very interactive textbooks, where you could describe the proof of a result in a very high-level sense, assuming lots of knowledge. But if there are steps that you don’t understand, you can expand them and go into details—all the way down the axioms if you want to. No one does this right now for textbooks because it’s too much work. But if you’re already formalizing it, the computer can create these interactive textbooks for you. It will make it easier for a mathematician in one field to start contributing to another because you can precisely specify subtasks of a big task that don’t require understanding everything.

The entire interview is worth reading.  As Adam Smith once said…

AI and Truth Evasion

by  Alex Tabarrok in

A good insight from Eliot Higgins, the head of the intelligence service Bellingcat.

When a lot of people think about AI, they think, “Oh, it’s going to fool people into believing stuff that’s not true.” But what it’s really doing is giving people permission to not believe stuff that is true. Because they can say, “Oh, that’s an AI-generated image. AI can generate anything now: video, audio, the entire war zone re-created.” They will use it as an excuse.

From an extensive interview in Wired.

Addendum: A case in point.

Leopold Aschenbrenner on AGI, and security matters

by  Tyler Cowen in

Here is the 165 pp. pdf piece, here is html, here is a Twitter short TOC preview.

And now Leopold on Dwarkesh.  I haven’t heard it, but a friend sends along this excerpt:

Why didn’t I ultimately pursue econ academia? There were several reasons, one of them being Tyler Cowen. He took me aside and said, “I think you’re one of the top young economists I’ve ever met, but you should probably not go to grad school.”
Dwarkesh Patel 02:19:50
Oh, interesting. Really? I didn’t realize that.
Leopold Aschenbrenner 02:19:53
Yeah, it was good because he kind of introduced me to the Twitter weirdos. I think the takeaway from that was that I have to move out west one more time.
Dwarkesh Patel 02:20:03
Wait Tyler introduced you to the Twitter weirdos?
Leopold Aschenbrenner 02:20:05
A little bit. Or just kind of the broader culture?
Dwarkesh Patel 02:20:08
A 60-year-old economist introduced you to Twitter?

TaskRabbit for AI Hires

by  Alex Tabarrok in

Many people are interested in knowing, which AI is the closest to achieving AGI? That’s an important question for philosophers and computer scientists but more and more I am seeing firms arise to answer a different question, Which AI should I hire?

EquiStamp, for example, rates dozens of AIs based on multiple evaluations, including custom evaluations tailored to specific business tasks. For instance, one firm may want to hire an AI to handle customer queries, another to sort packages, another to summarize internal technical documents. The best AI for each task might differ from the AI that scores highest on general reasoning power.  In addition, businesses care not just about performance but also about speed and cost. No reason to hire AI-Einstein to sort the mail. AIs are also continually being re-trained so their performance can fluctuate. Businesses, therefore, may want to continually test their AIs and quickly hire and fire AIs as needed.

In short, a spot-market for hiring AIs is developing. 

Is broadband good for you?

by  Tyler Cowen in

Kathryn R. Johnson and Claudia Persico have a new NBER working paper on exactly that topic:

Between 2000 and 2008, access to high-speed, broadband internet grew significantly in the United States, but there is debate on whether access to high-speed internet improves or harms wellbeing. We find that a ten percent increase in the proportion of county residents with access to broadband internet leads to a 1.01 percent reduction in the number of suicides in a county, as well as improvements in self-reported mental and physical health. We further find that this reduction in suicide deaths is likely due to economic improvements in counties that have access to broadband internet. Counties with increased access to broadband internet see reductions in poverty rate and unemployment rate. In addition, zip codes that gain access to broadband internet see increases in the numbers of employees and establishments. In addition, heterogeneity analysis indicates that the positive effects are concentrated in the working age population, those between 25 and 64 years old. This pattern is precisely what is predicted by the literature linking economic conditions to suicide risk.

It seems broadband is indeed (was indeed?) good for you.

Why some additional regulation would help crypto

by  Tyler Cowen in

That is the topic of my latest Bloomberg column, here is one part of the argument, which focuses on the bill that recently passed the House but may stall in the Senate:

As for the policy details: Is this a good bill? Mostly, yes. Without a coherent regulatory framework, the US crypto sector won’t be competitive with those of other nations. That damages the potential for American innovation, encourages some entrepreneurs to take their businesses abroad, and could eventually limit the integration of crypto with mainstream financial infrastructures, which would put the US financial sector at a disadvantage.

The bill has the critical provision of requiring crypto infrastructures to be sufficiently decentralized, at least if that infrastructure is to fall under the jurisdiction of the CFTC rather than the SEC. These decentralized crypto infrastructures, which would include Bitcoin and the Ethereum network, are considered to be “digital commodities,” and are granted greater freedom. Such assets are not like shares of Apple stock, where the buyer expects a very particular kind of corporate responsibility and predictable financial reporting. So the bill stipulates that, for many crypto assets, initial issuance must involve tighter disclosure and regulation, with a role for the SEC. Over time, as the blockchain for that asset became more mature and well-established, regulation would relax.

Any particular proposal for such rules will necessarily involve some ambiguities and be susceptible to being gamed (by companies) or abused (by regulators). What is considered “adequate” decentralization, for example, is ultimately a subjective question. Still, this bill seems like a reasonable place to start for crypto regulation.

The alternative to such regulations of course is complete regulatory discretion, which is currently part of the status quo.  And here is Alex’s recent post on crypto regulation.

AI passes the restaurant review Turing test

by  Tyler Cowen in

Surprisingly easy, it turned out. In a series of experiments for a new study, Kovács found that a panel of human testers was unable to distinguish between reviews written by humans and those written by GPT-4, the LLM powering the latest iteration of ChatGPT. In fact, they were more confident about the authenticity of AI-written reviews than they were about human-written reviews.

Here is the full story, via Sarah Jenislawski.

My Conversation with the excellent Michael Nielsen

by  Tyler Cowen in

Here is the audio, video, and transcript.  Here is the episode summary:

Michael Nielsen is scientist who helped pioneer quantum computing and the modern open science movement. He’s worked at Y Combinator, co-authored on scientific progress with Patrick Collison, and is a prolific writer, reader, commentator, and mentor. 

He joined Tyler to discuss why the universe is so beautiful to human eyes (but not ears), how to find good collaborators, the influence of Simone Weil, where Olaf Stapledon’s understand of the social word went wrong, potential applications of quantum computing, the (rising) status of linear algebra, what makes for physicists who age well, finding young mentors, why some scientific fields have pre-print platforms and others don’t, how so many crummy journals survive, the threat of cheap nukes, the many unknowns of Mars colonization, techniques for paying closer attention, what you learn when visiting the USS Midway, why he changed his mind about Emergent Ventures, why he didn’t join OpenAI in 2015, what he’ll learn next, and more. 

And here is one excerpt:

COWEN: Now, you’ve written that in the first half of your life, you typically were the youngest person in your circle and that in the second half of your life, which is probably now, you’re typically the eldest person in your circle. How would you model that as a claim about you?

NIELSEN: I hope I’m in the first 5 percent of my life, but it’s sadly unlikely.

COWEN: Let’s say you’re 50 now, and you live to 100, which is plausible —

NIELSEN: Which is plausible.

COWEN: — and you would now be in the second half of your life.

NIELSEN: Yes. I can give shallow reasons. I can’t give good reasons. The good reason in the first half was, so much of the work I was doing was kind of new fields of science, and those tend to be dominated essentially, for almost sunk-cost reasons — people who don’t have any sunk costs tend to be younger. They go into these fields. These early days of quantum computing, early days of open science — they were dominated by people in their 20s. Then they’d go off and become faculty members. They’d be the youngest person on the faculty.

Now, maybe it’s just because I found San Francisco, and it’s such an interesting cultural institution or achievement of civilization. We’ve got this amplifier for 25-year-olds that lets them make dreams in the world. That’s, for me, anyway, for a person with my personality, very attractive for many of the same reasons.

COWEN: Let’s say you had a theory of your collaborators, and other than, yes, they’re smart; they work hard; but trying to pin down in as few dimensions as possible, who’s likely to become a collaborator of yours after taking into account the obvious? What’s your theory of your own collaborators?

NIELSEN: They’re all extremely open to experience. They’re all extremely curious. They’re all extremely parasocial. They’re all extremely ambitious. They’re all extremely imaginative.

Self-recommending throughout.

Financial Statement Analysis with Large Language Models

by  Tyler Cowen in

We investigate whether an LLM can successfully perform financial statement analysis in a way similar to a professional human analyst. We provide standardized and anonymous financial statements to GPT4 and instruct the model to analyze them to determine the direction of future earnings. Even without any narrative or industry-specific information, the LLM outperforms financial analysts in its ability to predict earnings changes. The LLM exhibits a relative advantage over human analysts in situations when the analysts tend to struggle. Furthermore, we find that the prediction accuracy of the LLM is on par with the performance of a narrowly trained state-of-the-art ML model. LLM prediction does not stem from its training memory. Instead, we find that the LLM generates useful narrative insights about a company’s future performance. Lastly, our trading strategies based on GPT’s predictions yield a higher Sharpe ratio and alphas than strategies based on other models. Taken together, our results suggest that LLMs may take a central role in decision-making.

That is from a new paper by Alex Kim, Maximilian Muhn, and Valeri V. Nikolaev, all at Chicago Booth.  Via William Allen.

The death of the AI safety movement?

by  Tyler Cowen in

That is the topic of my latest Bloomberg column, and here is one part:

The safety movement probably peaked in March 2023 with a petition for a six-month pause in AI development, signed by many luminaries, including specialists in the AI field. As I argued at the time, it was a bad idea, and got nowhere.

Fast forward to the present. Senate Majority Leader Chuck Schumer and his working group on AI have issued a guidance document for federal policy. The plans involve a lot of federal support for the research and development of AI, and a consistent recognition of the national-security importance of the US maintaining its lead in AI. Lawmakers seem to understand that they would rather face the risks of US-based AI systems than have to contend with Chinese developments without a US counterweight. The early history of Covid, when the Chinese government behaved recklessly and nontransparently, has driven this realization home.

No less important is the behavior of the major tech companies themselves. OpenAI, Anthropic, Google and Meta all released major service upgrades this spring. Their new services are smarter, faster, more flexible and more capable. Competition has heated up, and that will spur further innovation.

Do note this:

The biggest current obstacles to AI development are the hundreds of pending AI regulatory bills in scores of US states. Many of those bills would, intentionally or not, significantly restrict AI development, as for instance one in California that would require pre-approval for advanced systems. In the past, this kind state-level rush has typically led to federal consolidation, so that overall regulation is coordinated and not too onerous. Whether a good final bill results from that process is uncertain, but Schumer’s project suggests that the federal government is more interested in accelerating AI than hindering it.

Safety work of course will continue under many guises, both private and governmental, but “AI safetyism,” as an intellectual movement, has peaked.

This piece was drafted before some of the recent controversies at OpenAI, and the argument does not rely on any particular interpretation of those events, one way or the other.

Arnaud Schenk has some useful clarifications, especially for those who cannot read the full column.

Here is a version of the column, adapted for Don Mclean’s song “American Pie.”

Is the internet bad for you?

by  Tyler Cowen in

A global, 16-year study1 of 2.4 million people has found that Internet use might boost measures of well-being, such as life satisfaction and sense of purpose — challenging the commonly held idea that Internet use has negative effects on people’s welfare.

“It’s an important piece of the puzzle on digital-media use and mental health,” says psychologist Markus Appel at the University of Würzburg in Germany. “If social media and Internet and mobile-phone use is really such a devastating force in our society, we should see it on this bird’s-eye view [study] — but we don’t.” Such concerns are typically related to behaviours linked to social-media use, such as cyberbullying, social-media addiction and body-image issues. But the best studies have so far shown small negative effects, if any2,3, of Internet use on well-being, says Appel.

From separate Gallup polls:

Pryzbylski and his colleagues analysed data on how Internet access was related to eight measures of well-being from the Gallup World Poll, conducted by analytics company Gallup, based in Washington DC. The data were collected annually from 2006 to 2021 from 1,000 people, aged 15 and above, in 168 countries, through phone or in-person interviews. The researchers controlled for factors that might affect Internet use and welfare, including income level, employment status, education level and health problems.

…The team found that, on average, people who had access to the Internet scored 8% higher on measures of life satisfaction, positive experiences and contentment with their social life, compared with people who lacked web access. Online activities can help people to learn new things and make friends, and this could contribute to the beneficial effects, suggests Appel.

Do note that in these latter data sets women ages 15-24 still are worse off from internet access.

Here is the Nature piece, via Clara B. Jones.