Chat-GPT: What Is It Good For?, & BRIEFLY NOTED
For 2023-01-11 We: I think my assessment is going to be something like this: an occasionally hallucinatory conventional-wisdom generator. After all, what else could it possibly be?...
CONDITION: I Am Significantly Dismayed by þe Harvard Administration’s Denial of a Fellowship to Kenneth Roth of HRW:
Using Ken White’s terminology, Kenneth Roth’s speech is not illegal or unrightful. It does not tend to erode the free speech community. And while some might call him indecent in using hyperbole in calling Israel today an apartheid state, I certainly would not—although I would say that that hyperbole is not likely to be a constructive contribution. So I see absolutely zero case for fellowship denial. (And I thought the administrative denial of a fellowship to Chelses Manning was extremely dismaying as well.
That said, I cannot sign on to the Open letter to Harvard Kennedy School Dean Douglas Elmendorf, Harvard President Lawrence Bacow, and Incoming Harvard President Claudine Gay. To claim that “welcoming Ehud Barak and Tzipi Livni” is part of “a long history of anti-Palestinian bias, discrimination, and racism at the Harvard Kennedy School” is not a statement that anyone who wants to contribute to a free-speech or an academic culture can make.
But do not take the absence of my signature from the letter as evidence that I am not significantly dismayed.
FOCUS: Chat-GPT: What Is It Good For?:
I was playing with Chat-GPT this morning, instead of doing real work:
I think I am converging towards an assessment. I think my assessment is going to be something like this: an occasionally hallucinatory conventional-wisdom generator.
After all, what else could it possibly be? Chat-GPT looks back at up to 4000 tokens (maybe 3000 words?) in the prompt(s), and then, probabilistically, it selects the next token to output. It was trained on half a trillion words, and it has 175 billion connection weights between nodes. And yet how it is wired together is not obvious to me, and it is not obvious why its wiring is the right wiring for us to be trying to build. It seems to me to be too word-cloudy to be likely to be successful. But, then, I also think that most oral and written communication is too word-cloudy to be well-designed for its purposes.
The worrisome thing is that what has made it viral is that it speaks with much too much authority, it is too convincing at conveying the false illusion that there is actually a human mind back there. And that is likely to mislead us.
For example, I fed it the quick brown fox jumped over the lazy. Its response was not “dog”, period, followed by something else. Its response was:
The phrase "the quick brown fox jumped over the lazy" is a well-known sentence that is used to test the quality of typefaces. The sentence uses every letter of the English alphabet, making it a good sentence to test the completeness of a typeface. The sentence is also commonly used as a typing test and as a sentence that can be used to test the autocomplete function of a word processing program.
This is wrong.
“The quick brown fox jumped over the lazy” is not a sentence.
“The quick brown fox jumped over the lazy” is not well-known.
“The quick brown fox jumped over the lazy” is not used to test the quality of typefaces.
All of those are true not of the prompt, but rather of “The quick brown fox jumped over the lazy dog.”
What is right is that “The quick brown fox jumped over the lazy” can be used to test the autocomplete function of a world processing program.
Do note that, as autocomplete, Chat-GPT has just flunked.
It is pretty clear to me what Chat-GPT did here: It thought:
“The quick brown fox jumped over the lazy” is almost the same as “The quick brown fox jumped over the lazy dog”.
So I will say the things that usually follow ““The quick brown fox jumped over the lazy dog”.
There are some occurrences of “The quick brown fox jumped over the lazy” which are then soon followed by “test the autocomplete function” and “word processing program”.
So I will output that as well.
No, the issue is complicated by the fact that a lot of the time—95%?—hen we speak or write, we are stochastic parrots giving the conventional or conventional wisdom tokens in response to a prompt. But there is that very meaningful extra 5%. That is what GPT-Chat lacks.
Perhaps what we are learning is (a) how much of language is social, reinforcement, blue rather than information, transmission, and (b) how much of language it is information transmission is, indeed, the transmission of useful and relatively straightforward conventional wisdom that Chat-GPT can grok by essentially averaging the corpus of human language instances? Perhaps not. Perhaps I am off base here…
MUST-READ: Sell þe Bonds!
Selling bonds for more than par is less stupid and less goofy than “disinvesting” the Social Security trust fund and then claiming that the trust fund balance is not part of the debt subject to limit. And that has been semi-routinely done for 28 years now:
Matt Yglesias: A new plan to get around the debt ceiling hostage: ‘If the platinum coin is too weird, meet high-yield bonds…. It’s debt ceiling time again…. Instead of selling a $100 bond that has a low interest rate for $100 dollars, they can offer a $100 bond that has a high interest rate and see how much money people will give them for it…. The Treasury could raise more than $100 while only issuing $100 worth of debt… stupid, but… less goofy than a platinum coin. And… stupid in a manner that is perfectly suitable to the underlying stupidity of the statutory debt ceiling, which purports to limit the face value of debt that the Treasury can issue separate from all the legislation that specifies how much money the Treasury has to spend…
ONE IMAGE: A little inflation is a much better problem to have þan an economy depressed for a decade:
ONE VIDEO: Four Pillars of AI:
BRIEFLY NOTED:
Very Briefly Noted:
Ethan Mollick: How to... use Chat-GPT to boost your writing: ‘The key to using generative AI successfully is prompt-crafting… very elaborate… abstract… almost poetic…. More elaborate and specific prompts work better… appears to me that the conversational interface is leading people to use the tool badly.
David Dayen: Regulators Prevented a Crypto-Fueled Economic Downturn: ‘The most important regulation has already been done. Keeping crypto out of the broader financial system was the most important regulatory action of the past decade…. If we manage to get out of this cycle without a recession, we will have the banking regulators, primarily Gary Gensler at the Securities and Exchange Commission, to thank… A 100% correct analysis from David Dayan/
Simon Wren-Lewis: Did 2010 austerity permanently reduce UK output?: ‘ There is evidence that austerity, in creating an unusually protracted recovery in aggregate demand from the GFC recession, did have a negative impact on productivity growth and therefore a persistent negative impact on output supply… There is very, very strong evidence.
Ian Buruma: China’s Autocracy in Crisis: ‘Top-down decisions, often implemented by competent technocrats, have enabled China to build high-quality infrastructure and achieve rapid economic growth. But President Xi Jinping’s handling of the COVID-19 crisis and his erratic decision-making have highlighted the flaws in China’s one-party regime… Was the PRC always imperial, with a mask of collective leadership in the same way that the Julio-Claudians preserved the mask of the republic and of the authority of the Senate? Right now it looks like we had three good emperors, and now have one very erratic one.
Tyler Cowen: China & the Risk of Great Power War: ‘If something has not happened for a long time, most people simply forget about it…. I think a major war between great powers falls into that category… But I have no idea what a war between great powers would look like today. "War" covers an extremely broad range of levels of violence.
Jürgen Osterhammel: The Transformation of the World: A Global History of the Nineteenth Century… A great great book.
K.P.: Turning the Page on 2022: ‘SLOUCHING TOWARDS UTOPIA: An Economic History of the Twentieth Century, by J. Bradford DeLong. A masterful analysis of the extended 20th century (1870-2020) U.S. economy, by the former Treasury official now at the University of California, Berkeley… “Masterful”. I like that. Very much.
Mark Galeotti: ‘Putin's decision to appoint Chief of the General Staff Gerasimov as new overall commander of Ukraine op is significant…. For Gerasimov… it is… the most poisoned of chalices. It's now on him, and I suspect Putin has unrealistic expectations again…. There will be spring offensives… demonstrating to the West that Russia is in this for the long haul, and hoping that we will lose the will and unity to continue to support Kyiv… "Supporting Kyiv" is not a heavy burden for the NATO alliance. And what with North American shale and LNG tankers, the costs for the European civilian economy will be low. The primary economic burden of continued war will fall upon the poor of Nigeria and Egypt. This makes no sense for Russia.
Dave Guarino: ‘One of the most dangerous feedback loops is narrative resonance. Much of the opportunity operating in complex systems is understanding the gap between 1. what those fed info via summary (synoptic lenses) understand the problems to be, and 2. ground truth…. Adjusting conclusions in response to what resonates with the former will tend towards convergence back to the (partial) epistemic model which is already a source of disconnect/problems… This is why directors and such should spend a month in the trenches. Of course, it is important that they spend the month in a not atypical trench.
Chris Hayes: ‘FIRE really is one of the only consistent voices in the academic free speech discourse, with a genuinely principled commitment that spans different ideological controversies… I confess I have probably been undervaluing FIRE.
¶s:
What to do when a person I know to be smart, careful, and well-informed talks things that seem to me to be, well, bonkers? I can read the Fed’s credibility with the bond market off of the numbers. The Fed has immense credibility as an inflation fighter with the bond market. And the consequences for being at the zero lower bound, and thus having the market rate of interest well-above the neutral rate are immense—not just economic consequences, but eroding democracy, and aiding rising fascism as well. So where does all this come from?
Ragu Rajan: Central banks can’t win when it comes to credibility on inflation: ‘Institutions are scrambling to rebuild their toolkits to deal with the new regime: The kind of credibility needed to escape a regime of overly low inflation, which we had until recently, is different from the kind needed to curb high inflation, which we have now. And by its very nature, credibility does not turn on a dime…. Should the Fed work once more to regain credibility as an inflation hawk? Credibility takes a long time to build, and inflation regimes could switch again…. [But] central banks will probably be most effective if they rebuild their commitment to combating high inflation. And if inflation falls too low, perhaps we should learn to live with it…
How is it that Poland and Malaysia have managed to successfully utilize FDI for growth in ways that other countries have not?
Noah Smith: The Poland/Malaysia model: ‘The richer you get, the harder it is to grow, so by reaching the ~$30,000 range, Poland and Malaysia have accomplished an impressive feat… not quite as impressive as South Korea, but then again, who is?… They relied heavily on foreign direct investment…. Poland and Malaysia are both very export-intensive economies—as much or more so than South Korea…. Ha-Joon Chang and some other industrial policy fans think that FDI is not the basis of a sound development strategy…. Developing the South Korean way, by building a bunch of world-beating high-tech manufacturing companies from scratch, is incredibly hard. An FDI-centric strategy, on the other hand, is simple and straightforward…
A good example of the evil mindset of the Washington press corps back in the day and, indeed, today. Reporters who are flattered as peers by government officials willing to cut corners, and tell lies are than willing to go the extra mile to excuse such a government officials. I assure you that Cap Weinberger did not look at Richard Cohen as a “basic sort of guy” and a sort-of peer, but as a mark:
Richard Cohen (1992): Without Candor: ‘Back when Caspar Weinberger was secretary of defense, he and I used to meet all the time. Our “meetings”… took place in the Georgetown Safeway.... Once... he bought a turkey. I tell you this about the man President Bush just pardoned because it always influenced my opinion of Weinberger.... Based on my Safeway encounters, I came to think of Weinberger as a basic sort of guy, candid and no nonsense—which is the way much of official Washington saw him. It seemed somehow cruel that he should end his career—he's 75—either as a defendant in a criminal case or as a felon. The man deserved better than that.... He may have lost his good sense when he allegedly withheld evidence. That being said, I was pleased when he was pardoned...
Kevin McCarthy plays Jim Fallows. As long as McCarthy has the guys to talk about stolen elections, socialists, and Hunter Biden, he can appear to take the high road and thus store up street cred with which he can gull Jim in the future. That’s my view, at least:
James Fallows: OK, There Was Something Positive in the Speaker-Vote Debacle: ‘What Kevin McCarthy didn't say is worth at least noticing as well…. Part of McCarthy’s speech was origin story… and part of it was standard MAGA platform… swamp… Afghanistan… the origins of COVID… the weaponization of the FBI… Select Committee on China to investigate how to bring back the hundreds of thousands of jobs…. But if you’d listened to the… general GOP rhetoric of the era, you’d notice the things McCarthy did not say… “stolen” election… “socialists”… Hunter Biden… [no] saber-rattl[ing] about raising the debt ceiling… not use “Democrat” as an adjective (rather than “Democratic”), in the routinely insulting way pioneered by Newt Gingrich… did not disparage the man who gave him the gavel, Hakeem Jeffries…. Will this make any difference in terms of votes or policy?… No. But it could have been worse…. It will get worse, but for those few minutes it wasn’t. For now let’s take the win…
The failure of the Republican “moderates” to hold up the passage of the rules package for even a day is a very bad sign:
Robert Hubbell: Backroom deals, lies, and performative grift!: ‘The public version of the House resolution containing the rules omitted a secret supplement that conceals the corrupt promises made by McCarthy to purchase votes of reluctant GOP members who did not trust him…. A secret agreement to “freeze” congressional spending! Sounds like something Americans deserve to know about, doesn’t it? Or how about a “debt-ceiling strategy” that was part of a quid-pro-quo to drag McCarthy over the line in his ego-driven quest to become Speaker?… There are no innocent answers to that question, and we are entitled to presume the worst…
Fallows: ' not use “Democrat” as an adjective (rather than “Democratic”), in the routinely insulting way pioneered by Newt Gingrich'
The pedant strikes again: the phrase "Democrat party" is one I remember from my high school days, which ended in 1959.
I wonder how this compares to the Chinese development in their AI program?