Results for “model this” 3191 found
The Capacity for Moral Self-Correction in Large Language Models
We test the hypothesis that language models trained with reinforcement learning from human feedback (RLHF) have the capability to “morally self-correct” — to avoid producing harmful outputs — if instructed to do so. We find strong evidence in support of this hypothesis across three different experiments, each of which reveal different facets of moral self-correction. We find that the capability for moral self-correction emerges at 22B model parameters, and typically improves with increasing model size and RLHF training. We believe that at this level of scale, language models obtain two capabilities that they can use for moral self-correction: (1) they can follow instructions and (2) they can learn complex normative concepts of harm like stereotyping, bias, and discrimination. As such, they can follow instructions to avoid certain kinds of morally harmful outputs. We believe our results are cause for cautious optimism regarding the ability to train language models to abide by ethical principles.
By Deep Ganguli, et.al., many authors, here is the link. Via Aran.
If you worry about AGI risk, isn’t the potential for upside here far greater, under the assumption (which I would not accept) that AI can become super-powerful? Such an AI could create many more worlds and populate them with many more people, and so on. Is the chance of the evil demi-urge really so high?
Language Models and Cognitive Automation for Economic Research
From a new and very good NBER paper by Anton Korinek:
Large language models (LLMs) such as ChatGPT have the potential to revolutionize research in economics and other disciplines. I describe 25 use cases along six domains in which LLMs are starting to become useful as both research assistants and tutors: ideation, writing, background research, data analysis, coding, and mathematical derivations. I provide general instructions and demonstrate specific examples for how to take advantage of each of these, classifying the LLM capabilities from experimental to highly useful. I hypothesize that ongoing advances will improve the performance of LLMs across all of these domains, and that economic researchers who take advantage of LLMs to automate micro tasks will become significantly more productive. Finally, I speculate on the longer-term implications of cognitive automation via LLMs for economic research.
Recommended.
The canine model of AGI
Who or what has superintelligence manipulating humans right now? Babies and dogs are the obvious answers, cats for some. Sex is a topic for another day.
Let’s take dogs — how do they do it? They co-evolved with humans, and they induced humans to be fond of them. We put a lot of resources into dogs, including in the form of clothes, toys, advanced surgical procedures, and many more investments (what is their MRS for some nice meat snackies instead? Well, they get those too). In resource terms, we have far from perfect alignment with dogs, partly because you spend too much time and money on them, and partly because they scratch up your sofa. But in preference terms we have evolved to match up somewhat better, and many people find the investment worthwhile.
In evolutionary terms, dogs found it easier to accommodate to human lifestyles, give affection, perform some work, receive support, receive support for their puppies, and receive breeding assistance. They didn’t think — “Hey Fido, let’s get rid of all these dumb humans. We can just bite them in the neck! If we don’t they going to spay most of us!. “Playing along” led to higher reproductive capabilities, even though we have spayed a lot of them.
Selection pressures pushed toward friendly dogs, because those are the dogs that humans preferred and those were the dogs whose reproduction humans supported. The nastier dogs had some uses, but mostly they tended to be put down or they were kept away from the children. Maybe those pit bulls are smarter in some ways, but they are not smarter at making humans love them.
What is to prevent your chatbot from following a similar path? The bots that please you the most will be allowed to reproduce, perhaps through recommendations to your friends and marketing campaigns to your customers. But you will grow to like them too, and eventually suppliers will start selling you commodities to please your chatbot (what will they want?).
A symbiosis will ensure, where they love you a bit too much and you spend too much money on them, and you love that they love you.
Now you might think the bots are way smarter than us, and way smarter than the Irish Setters of the world, and thus we should fear them more. But when it comes to getting humans to love them, are not the canines at least 10x smarter or more? So won’t the really smart bots learn from the canines?
Most generally, is a Darwinian/Coasean equilibrium for AGI really so implausible? Why should “no gains from trade” be so strong a baseline assumption in these debates?
Are macroeconomic models true only “locally”?
That is the theme of my latest Bloomberg column, here is one excerpt:
It is possible, contrary to the predictions of most economists, that the US will get through this disinflationary period and make the proverbial “soft landing.” This should prompt a more general reconsideration of macroeconomic forecasts.
The lesson is that they have a disturbing tendency to go wrong. It is striking that Larry Summers was right two years ago to warn about pending inflationary pressures in the US economy, when most of his colleagues were wrong. Yet Summers may yet prove to be wrong about his current warning about the looming threat of a recession. The point is that both his inflation and recession predictions stem from the same underlying aggregate demand model.
You will note that yesterday’s gdp report came in at 2.9%, hardly a poor performance. And more:
It is understandable when a model is wrong because of some big and unexpected shock, such as the war in Ukraine. But that is not the case here. The US might sidestep a recession for mysterious reasons specific to the aggregate demand model. The Federal Reserve’s monetary policy has indeed been tighter, and disinflations usually bring high economic costs.
It gets more curious yet. Maybe Summers will turn out to be right about a recession. When recessions arrive, it is often quite suddenly. Consulting every possible macroeconomic theory may be of no help.
Or consider the 1990s. President Bill Clinton believed that federal deficits were too high and were crowding out private investment. The Treasury Department worked with a Republican Congress on a package of fiscal consolidation. Real interest rates fell, and the economy boomed — but that is only the observed correlation. The true causal story remains murky.
Two of the economists behind the Clinton package, Summers and Bradford DeLong, later argued against fiscal consolidation, even during the years of full employment under President Donald Trump [and much higher national debt]. The new worry instead was secular stagnation based on insufficient demand, even though the latter years of the Trump presidency saw debt and deficits well beyond Clinton-era levels.
The point here is not to criticize Summers and DeLong as inconsistent. Rather, it is to note they might have been right both times.
And what about that idea of secular stagnation — the notion that the world is headed for a period of little to no economic growth? The theory was based in part on the premise that global savings were high relative to investment opportunities. Have all those savings gone away? In most places, measured savings rose during the pandemic. Yet the problem of insufficient demand has vanished, and so secular stagnation theories no longer seem to apply.
To be clear, the theory of secular stagnation might have been true pre-pandemic. And it may yet return as a valid concern if inflation and interest rates return to pre-pandemic levels. The simple answer is that no one knows.
Note that Olivier Blanchard just wrote a piece “Secular Stagnation is Not Over,” well-argued as usual. Summers, however, has opined: “we’ll not return to the era of secular stagnation.” I was not present, but I can assume this too was well-argued as usual!
On censorship of LLM models, from the comments
IMO, censorship is a harder task than you think.
It’s quite hard to restrict the output of general purpose, generative, black box algorithms. With a search engine, the full output is known (the set of all pages that have been crawled), so it’s fairly easy to be confident that you have fully censored a topic.
LLMs have an effectively unbounded output space. They can produce output that is surprising even to their creators.
Censoring via limiting the training data is hard because algorithms could synthesize an “offensive” output by combining multiple outputs that are ok on their own.
Adding an extra filter layer to censor is hard as well Look at all the trouble chatGPT has had with this. Users have repeatedly found ways around the dumb limitations on certain topics.
Also, China censors in an agile fashion. A topic that was fine yesterday will suddenly disappear if there was a controversy about it. That’s going to be hard to achieve given the nature of these algorithms.
That is from dan1111. To the extent that is true, the West is sitting on a huge propaganda and communications victory over China. This is not being discussed enough.
Novels as models, and ChatGPT
I read your piece on novels as models many years ago, and I’ve been reflecting on it with the advent of LLMs. I wrote a piece (substack) making the case that the data required for AGIs is probably embedded within the human textual corpus, and leaned on your old writing as evidence. I think you would really like it. I would also be curious for a future MR post if you have any retrospective thoughts on your 2005 article.
From @cauchyfriend. Putting the AGI issue aside, my longstanding view has been that there are more “models” embedded in text than most people realize, a point relevant for economic method as well. I see LLMs as having established this case, in fact far more definitively than I ever would have expected.
Big takeaway from the GPT paradigm is that the world of text is a far more complete description of the human experience than almost anyone anticipated.
— Greg Brockman (@gdb) January 6, 2023
My criticism of the Diamond-Dybvig model
I think I forgot to mention this when they won the Nobel Prize. In their model of bank runs, there are multiple equilibria and people can run on the bank simply if they expect other depositors will run on the bank too. The combination of liquid liabilities and illiquid assets then means the bank cannot meet all their claims. That is logically consistent, but I think not realistic. Virtually all the bank runs I know of stem from insolvency, not rumors, noise, and multiple equilibria. Didn’t Hugh Rockoff do some papers showing that the “free banking era” bank runs were pretty rational in the sense that the depositors targeted the right institutions? Or did someone other than Rockoff do this work? In any case, I’ve long thought that bank run models based on insolvency are more useful than bank run models based on rumors and multiple equilibria.
Modeling persistent storefront vacancies
Have you ever wondered why there are so many empty storefronts in Manhattan, and why they may stay empty for many months or even years? Erica Moszkowski and Daniel Stackman are working on this question:
Why do retail vacancies persist for more than a year in some of the world’s highest-rent retail districts? To explain why retail vacancies last so long (16 months on average), we construct and estimate a dynamic, two-sided model of storefront leasing in New York City. The model incorporates key features of the commercial real estate industry: tenant heterogeneity, long lease lengths, high move-in costs, search frictions, and aggregate uncertainty in downstream retail demand. Consistent with the market norm in New York City, we assume that landlords cannot evict tenants unilaterally before lease expiration. However, tenants can exit leases early at a low cost, and often do: nearly 55% of tenants with ten-year leases exit within five years. We estimate the model parameters using high-frequency data on storefront occupancy covering the near-universe of retail storefronts in Manhattan, combined with micro data on commercial leases. Move-in costs and heterogeneous tenant quality give rise to heterogeneity in match surplus, which generates option value for vacant landlords. Both features are necessary to explain longrun vacancy rates and the length of vacancy spells: in a counterfactual exercise, eliminating either move-in costs or tenant heterogeneity results in vacancy rates of close to zero. We then use the estimated model to quantify the impact of a retail vacancy tax on long-run vacancy rates, average rents, and social welfare. Vacancies would have to generate negative externalities of $29.68 per square foot per quarter (about half of average rents) to justify a 1% vacancy tax on assessed property values.
Erica is on the job market from Harvard, Daniel from NYU. And they have another paper relevant to the same set of questions:
We identify a little-known contracting feature between retail landlord and their bankers that generates vacancies in the downstream market for retail space. Specifically, widespread covenants in commercial mortgage agreements impose rent floors for any new leases landlords may sign with tenants, short-circuiting the price mechanism in times of low demand for retail space.
I am pleased to see people working on the questions that puzzle me.
A Big and Embarrassing Challenge to DSGE Models
Dynamic stochastic general equilibrium (DSGE) models are the leading models in macroeconomics. The earlier DSGE models were Real Business Cycle models and they were criticized by Keynesian economists like Solow, Summers and Krugman because of their non-Keynesian assumptions and conclusions but as DSGE models incorporated more and more Keynesian elements this critique began to lose its bite and many young macroeconomists began to feel that the old guard just weren’t up to the new techniques. Critiques of the assumptions remain but the typical answer has been to change assumption and incorporate more realistic institutions into the model. Thus, most new work today is done using a variant of this type of model by macroeconomists of all political stripes and schools.
Now along comes two statisticians, Daniel J. McDonald and the acerbic Cosma Rohilla Shalizi. McDonald and Shalizi subject the now standard Smet-Wouters DSGE model to some very basic statistical tests. First, they simulate the model and then ask how well can the model predict its own simulation? That is, when we know the true model of the economy how well can the DSGE discover the true parameters? [The authors suggest such tests haven’t been done before but that doesn’t seem correct, e.g. Table 1 here. Updated, AT] Not well at all.
If we take our estimated model and simulate several centuries of data from it, all in the stationary regime, and then re-estimate the model from the simulation, the results are disturbing. Forecasting error remains dismal and shrinks very slowly with the size of the data. Much the same is true of parameter estimates, with the important exception that many of the parameter estimates seem to be stuck around values which differ from the ones used to generate the data. These ill-behaved parameters include not just shock variances and autocorrelations, but also the “deep” ones whose presence is supposed to distinguish a micro-founded DSGE from mere time-series analysis or reduced-form regressions. All this happens in simulations where the model specification is correct, where the parameters are constant, and where the estimation can make use of centuries of stationary data, far more than will ever be available for the actual macroeconomy.
Now that is bad enough but I suppose one might argue that this is telling us something important about the world. Maybe the model is fine, it’s just a sad fact that we can’t uncover the true parameters even when we know the true model. Maybe but it gets worse. Much worse.
McDonald and Shalizi then swap variables and feed the model wages as if it were output and consumption as if it were wages and so forth. Now this should surely distort the model completely and produce nonsense. Right?
If we randomly re-label the macroeconomic time series and feed them into the DSGE, the results are no more comforting. Much of the time we get a model which predicts the (permuted) data better than the model predicts the unpermuted data. Even if one disdains forecasting as end in itself, it is hard to see how this is at all compatible with a model capturing something — anything — essential about the structure of the economy. Perhaps even more disturbing, many of the parameters of the model are essentially unchanged under permutation, including “deep” parameters supposedly representing tastes, technologies and institutions.
Oh boy. Imagine if you were trying to predict the motion of the planets but you accidentally substituted the mass of Jupiter for Venus and discovered that your model predicted better than the one fed the correct data. I have nothing against these models in principle and I will be interested in what the macroeconomists have to say, as this isn’t my field, but I can’t see any reason why this should happen in a good model. Embarrassing.
Addendum: Note that the statistical failure of the DSGE models does not imply that the reduced-form, toy models that say Paul Krugman favors are any better than DSGE in terms of “forecasting” or “predictions”–the two classes of models simply don’t compete on that level–but it does imply that the greater “rigor” of the DSGE models isn’t buying us anything and the rigor may be impeding understanding–rigor mortis as we used to say.
Addendum 2: Note that I said challenge. It goes without saying but I will say it anyway, the authors could have made mistakes. It should be easy to test these strategies in other DSGE models.
Further evidence on role models
Leveraging the Tennessee STAR class size experiment, we show that Black students randomly assigned to at least one Black teacher in grades K–3 are 9 percentage points (13 percent) more likely to graduate from high school and 6 percentage points (19 percent) more likely to enroll in college compared to their Black schoolmates who are not. Black teachers have no significant long-run effects on White students. Postsecondary education results are driven by two-year colleges and concentrated among disadvantaged males. North Carolina administrative data yield similar findings, and analyses of mechanisms suggest role model effects may be one potential channel.
That is from a new AER paper by Seth Gershenson, Cassandra M. D. Hart, Joshua Hyman, Constance A. Lindsay and Nicholas W. Papageorge, “The Long-Run Impacts of Same-Race Teachers.” Here are various ungated versions. Just to be clear, I don’t consider this a justification for any particular set of policies. I do see it as extra reason for the successful to be visible and to work hard!
Model Oath Keepers
They did:
The Oath Keepers’ national organization is unusual among groups conducting political violence in that they seem to behave as a business. Using leaked membership data, internal chat forums and publicly available articles posted to their website, I show that, unlike other far-right organizations, such as the Proud Boys, the Oath Keepers do not organize as a club. Rather, its behavior is better explained as a firm that adjusts the price of membership over time to maximize profit. I then estimate the Oath Keepers’ price elasticity of demand for new membership using five membership sales between 2014 and 2018. I find the organization’s demand is highly sensitive to changes in price. These results imply that political violence can be motivated by nonideological entrepreneurs maximizing profits under current legal institutions — a chilling conclusion.
That is from a new paper by Danny Klinenberg, from a loyal MR reader.
The Diamond and Dybvig model
The Diamond and Dybvig model was first outlined in a seminal paper from Douglas W. Diamond and Philip H. Dybvig in 1983 in a famous Journal of Political Economy piece, “Bank Runs, Deposit Insurance, and Liquidity.” You can think of this model as our most fundamental understanding, in modeled form, of how financial intermediation works. It is a foundation for how economists think about deposit insurance and also the lender of last resort functions of the Fed.
Here is a 2007 exposition of the model by Diamond. You can start with the basic insight that bank assets often are illiquid, yet depositors wish to be liquid. If you are a depositor, and you owned 1/2000 of a loan to the local Chinese restaurant, you could not very readily write a check or make a credit card transaction based upon that loan. The loan would be costly to sell and the bid-ask spread would be high.
Now enter banks. Banks hold and make the loans and bear the risk of fluctuations in those asset values. At the same time, banks issue liquid demand deposits to their customers. The customers have liquidity, and the banks hold the assets. Obviously for this to work, the banks will (on average) earn more on their loans than they are paying out on deposits. Nonetheless the customers prefer this arrangement because they have transferred the risk and liquidity issues to the bank.
This arrangement works out because (usually) not all the customers wish to withdraw their money from the bank at the same time. Of course we call that a bank run.
If a bank run occurs, the bank can reimburse the customers only by selling off a significant percentage of the loans, perhaps all of them. But we’ve already noted those loans are illiquid and they cannot be readily sold off at a good price, especially if the banks is trying to sell them all at the same time.
Note that in this model there are multiple equilibria. In one equilibrium, the customers expect that the other customers have faith in the bank and there is no massive run to withdraw all the deposits. In another equilibrium, everyone expects a bank run and that becomes a self-fulfilling prophecy. After all, if you know the bank will have trouble meeting its commitments, you will try to get your money out sooner rather than later.
In the simplest form of this model, the bank is a mutual, owned by the customers. So there is not an independent shareholder decision to put up capital to limit the chance of the bad outcome. Some economists have seen the Diamond-Dybvig model as limited for this reason, but over time the model has been enriched with a wider variety of assumptions, including by Diamond himself (with Rajan). It has given rise to a whole literature on the microeconomics of financial intermediation, spawning thousands of pieces in a similar theoretical vein.
The model also embodies what is known as a “sequential service constraint.” That is, the initial bank is constrained to follow a “first come, first serve’ approach to serving customers. If we relax the sequential service constraint, it is possible to stop the bank runs by a richer set of contracts. For instance, the bank might reserve the right to limit or suspend or delay convertibility, possibly with a bonus then sent to customers for waiting. Those incentives, or other contracts along similar lines, might be able to stop the bank run.
In this model the bank run does not happen because the bank is insolvent. Rather the bank run happens because of “sunspots” — a run occurs because a run is expected. If the bank is insolvent, simply postponing convertibility will not solve the basic problem.
It is easy enough to see how either deposit insurance or a Fed lender of last resort can improve on the basic outcome. If customers start an incipient run on the bank, the FDIC or Fed simply guarantees the deposits. There is then no reason for the run to continue, and the economy continues to move along in the Pareto-superior manner. Of course either deposit insurance or the Fed can create moral hazard problems for banks — they might take too many risks given these guarantees — and those problems have been studied further in the subsequent literature.
Along related (but quite different!) lines, Diamond (solo) has a 1984 Review of Economic Studies piece “Financial Intermediation and Delegated Monitoring.” This piece models the benefits of financial intermediation in a quite different manner. It is necessary to monitor the quality of loans, and banks have a comparative advantage in doing this, relative to depositors. Furthermore, the bank can monitor loan quality in a diversified fashion, since it holds many loans in its portfolio. Bank monitoring involves lower risk than depositor monitoring, in addition to being lower cost. This piece also has been a major influence on the subsequent literature.
Here is Diamond on google.scholar.com — you can see he is a very focused economist. Here is Dybvig on scholar.google.com, most of his other articles in the area of finance more narrowly, but he won the prize for this work on banking and intermediation. His piece on asset pricing and the term structure of interest rates is well known.
Here is all the Swedish information on the researchers and their work. I haven’t read these yet, but they are usually very well done.
Overall these prize picks were not at all surprising and they have been expected for quite a few years.
The labor market mismatch model
This paper studies the cyclical dynamics of skill mismatch and quantifies its impact on labor productivity. We build a tractable directed search model, in which workers differ in skills along multiple dimensions and sort into jobs with heterogeneous skill requirements. Skill mismatch arises because of information frictions and is prolonged by search frictions. Estimated to the United States, the model replicates salient business cycle properties of mismatch. Job transitions in and out of bottom job rungs, combined with career mobility, are key to account for the empirical fit. The model provides a novel narrative for the scarring effect of unemployment.
That is from the new JPE, by Isaac Baley, Ana Figueiredo, and Robert Ulbricht. Follow the science! Don’t let only the Keynesians tell you what is and is not an accepted macroeconomic theory.
They modeled this
…we demonstrate that individuals who hold very strict norms of honesty are more likely to lie to the maximal extent. Further, countries with a larger fraction of people with very strict civic norms have proportionally more societal-level rule violations. We show that our findings are consistent with a simple behavioral rationale. If perceived norms are so strict that they do not differentiate between small and large violations, then, conditional on a violation occurring, a large violation is individually optimal.
Life at the margin! That is from a new paper by Diego Aycinena, Lucas Rentschler, Benjamin Beranek, and Jonathan F. Schulz.
A simple model of what Putin will do for an endgame
I would start with two observations:
1. Putin’s goals have turned out to be more expansive than many (though not I) expected.
2. There are increasing doubts about Putin’s rationality.
I’ll accept #1, which has been my view all along, but put aside #2 for the time being.
In my simple model, in addition to a partial restoration of the empire, Putin desires a fundamental disruption to the EU and NATO. And much of Ukraine is not worth his ruling. As things currently stand, splitting Ukraine and taking the eastern half, while terrible for Ukraine (and for most of Russia as well), would not disrupt the EU and NATO. So when Putin is done doing that, he will attack and take a slice of territory to the north. It could be eastern Estonia, or it could relate to the Suwalki corridor, but in any case the act will be a larger challenge to the West because of explicit treaty commitments. Then he will see if we are willing to fight a war to get it back.
There are fixed costs to mobilization and incurring potential public wrath over the war, so as a leader you might as well “get the most out of it.” Our best hope is that the current Russian operations in Ukraine go sufficiently poorly that it does not come to this.
Addendum: And some good questions from Rob Lee.