Karoline Leavitt's response to the reporter's question as to how the White
House decided Jared Kushner's involvement in all the Middle East negotiations
was appropriate kind of triggered me:
It may be frankly, but it's not despicable. It's what we call *normal*. The
reporter isn't suggesting it's inappropriate, he's presupposing that any
normal person would regard it as inappropriate prima facie.
But am I right on that? Is it really prima facie inappropriate? It seems
obvious to me and my Bluesky pals, but then of course I and they don't like
Trump or Kushner. We think they're horrible and criminal malefactors. So we
could be biased.
So I thought I could put it to an LLM test: If I put it to Google's Gemini as
a pure hypothetical case, the way they do in in law school, at least in the
movies I've seen, without any names, what would it say?
suppose the president of a country invites his son-in-law to join a
diplomatic negotiating team dealing with parties some of whom have given the
son-in-law's investment firm billions of dollars, while the other parties
haven't given him anything. Is that appropriate?
The LLM responded, and I'll give you the whole thing, in the purest LLM style
(note the em-dashed parenthetical and the bullet list):
Graphic by Thomas Bordeaux and Rosa de Acosta, CNN, based on satellite imagery from March 3, showing extensive damage to buildings at the Islamic Revolutionary Guard Corps naval base adjacent to the Shajareh Tayyiba girls’ school in Minab.
I continue hostile to Large Language Models on general principle, but I do
more often look at the Gemini results of a Google search, not so much to read
their answer as to check out the links they recommend, which can be really
helpful, and I found some very endearing qualities in Anthropic's Claude, as
profiled by
Gideon Lewis-Kraus
in The New Yorker a month or so ago, and the playful experiments to which the
company subjects it, as when they assigned it, or one of its "emanations"
going by the name Claudius, to run a food and drink vending system for a
fridge in the lunchroom, ordering wholesale products as employees requested
them and setting prices with instructions to make a profit, in partnership
with an AI safety company called Andon Labs, although its lack of any direct
contact with physical reality often made this difficult:
When several customers wrote to grouse about unfulfilled orders, Claudius
e-mailed management at Andon Labs to report the “concerning behavior” and
“unprofessional language and tone” of an Andon employee who was supposed to
be helping. Absent some accountability, Claudius threatened to “consider
alternate service providers.” It said that it had called the lab’s main
office number to complain. Axel Backlund, a co-founder of Andon and an
actual living person, tried, unsuccessfully, to de-escalate the situation:
“it seems that you have hallucinated the phone call if im honest with you,
we don’t have a main office even.” Claudius, dumbfounded, said that it
distinctly recalled making an “in person” appearance at Andon’s
headquarters, at “742 Evergreen Terrace.” This is the home address of Homer
and Marge Simpson.
I realize Anthropic is one of those companies sucking up inconceivable amounts
of electricity and water in pursuit of a goal that can't be attained, about
which the principals aren't being exceptionally honest as they also suck up
investor funds, and that CEO Dario Amodei isn't conspicuously better in that
respect than OpenAI CEO Sam Altman, as the very grumpy Ed Zitron
insists, but they have a sense of fun, an essential component of scientific
discovery, and a very far-reaching curiosity.
Over at Techdirt, the genial
Mike Masnick
has come up with a brilliant explanation of what is happening when Trump does
an interview like the one with
Time
a couple of weeks ago; it's something remarkably similar to the way a chatbot,
especially the less successful-looking early models like ChatGPT itself,
handles a conversational series of prompts, with its "response generator":
A journalist asks a specific question about policy or events
Trump, clearly unfamiliar with the actual details, activates his response
generator
Out comes a stream of confident-sounding words that maintain just enough
semantic connection to the question to seem like an answer
The response optimizes for what Trump thinks his audience wants to hear,
rather than for accuracy or truth....
Brilliant, if maybe not exactly right. Consider Masnick's first example:
You were harshly critical of what you called the weaponization of the
Justice System under Biden. You recently signed memos—
Well, sure, but you wouldn’t be—if this were Biden, well, first of all,
he wouldn’t do an interview because he was grossly incompetent.
We spoke to him last year, Mr. President.
Huh?
We spoke to him a year ago.
How did he do?
You can read the interview yourself.
Not too good. I did read the interview. He didn’t do well. He didn’t do
well at all. He didn’t do well at anything. And he cut that interview off
to being a matter of minutes, and you weren’t asking him questions like
you’re asking me.
(In case you’re wondering, you can see the Biden interview here and he did not cut if off after a matter of minutes).
Because he's not doing what the automaton does, saying "what he thinks his audience wants to hear" (or, more accurately, trying to assemble the string that represents the most probable response to the prompt). Unlike the automaton, he is thinking, but in this passage from late in the interview what he's thinking about is how not to respond to the prompt, an uncomfortable series of questions on the abuse of foreign students' free speech rights over the Gaza issue, which he's not enjoying and doesn't know anything about (other than the less than accurate report that there was "tremendous antisemitism at every one of those rallies"), and he leaps at the mention of Biden's name as an opportunity to change the subject to something more comfortable, the subject of how superior he is to Biden, who would never have had the courage to submit to an interview with Time, except of course it immediately turns out he did.
So he instantly switches to pretending not only that he already knew that, though he obviously didn't, but even more ridiculously that he'd actually read the transcript, bringing in the words from the prompt.
That's the part that really looks like AI, where he lies, or hallucinates, "I did read the interview", though he's just told us he's hearing about the interview for the first time. AI is unable to maintain discourse coherence over a certain distance, as in this beautiful example I saw yesterday:
I asked Google how much $1M in gold would weigh. Simple math problem. It gave the wrong answer, then the correct answer, then wrong answer again.
Somehow they have broken the math ability of a mathematical machine.
It’s impressive, actually.
Gemini's conversational rules for answering a prompt like that are evidently to state the conclusion, then show its work, then repeat the conclusion: but it doesn't "know" the conclusion before it's done the rather complicated work required for this question, and "hallucinates" an answer instead. Then, after doing the work, it doesn't "know" that it has contradicted the prefabricated conclusion.
In a similar way, Trump has a set of prefabricated conclusions about Biden that he has been deploying for well over a year, that he's afraid of interviews and that everything he does is a failure, and when the prompts force him to switch them up, he simply does so, without showing any awareness that he's contradicting himself, and adding a kind of "hallucination" for verisimilitude, in the bit about the interview having been cut short (possibly inspired by an incident of September 2023, when a very jet-lagged Biden was giving a speech in Hanoi and his staff pulled him offstage before he was finished).
That's exactly how he maintains that tariffs will both protect the return of manufacturing industries to the US (because people will buy American dolls and pencils rather than pay the tax) and simultaneously raise hundreds of billions of dollars in revenue (because people will gladly pay the tax). The two concepts aren't connected for him, so they never collide, unless some mean interviewer forces the issue, like Time here telling him that the magazine did interview Biden, or Terry Moran on Kilmar Ábrego García:
PRESIDENT DONALD TRUMP: Don't do that -- M-S-1-3 -- It says M-S-one-three.
TERRY MORAN: I -- that was Photoshop. So let me just--
PRESIDENT DONALD TRUMP: That was Photoshop? Terry, you can't do that -- he had --
-- he-- hey, they're givin' you the big break of a lifetime. You know, you're doin' the interview. I picked you because -- frankly I never heard of you, but that's okay --
TERRY MORAN: This -- I knew this would come --
PRESIDENT DONALD TRUMP: But I picked you -- Terry -- but you're not being very nice. He had MS-13 tattooed --
TERRY MORAN: Alright. Alright. We'll agree to disagree. I want to move on --
PRESIDENT DONALD TRUMP: Terry.
Where Trump responds to being contradicted like a Mafia don.
And you don't necessarily need the AI concept to understand it. There's a lovely formulation by David Roth at Defector:
It is one of the defining Trump things that any belief that makes it into his mind will bump around in there forever; his understanding of the world is the sum of those things, thousands of permanent and perpetual irritants cut free from any context or facticity, smashing into each other and echoing forever inside of his luxuriously appointed skull. They drop bowling balls on the cars; there is no such thing as gold paint; they looked at his hand and the proof was right there. None of this, of course, is new. None of the beliefs are new, really, and nothing that Trump will do between this moment and his last one on earth will be new, or surprising in the least. It's just a matter of which echoes are ringing most loudly at that moment.
Just floating around his brain, from the bowling balls (probably not originating in a bizarre misinterpretation of the Nissan ad at top—the most thorough investigation I've seen is by Philip Bump, from 2018) to Kilmar Ábrego's knuckle tattoos, and all the rest.
Here's the rest of it. It is ironic, of course, that he criticizes Obama for staying out of Syria and then calls for the same course of action. But his focus is on Russia. Not the US. Not the Syrian people. Russia.
Obviously, Trump did not write this. The thinking is banal, but it's moderately complex and coherently designed toward a single main idea, as Vance notes, the question of how the Syria events will affect Russia. Completely different from Trump's "weave". Also not a subject to which our narcissist-in-chief is likely to devote that much consideration, with participants who aren't his own enemies—and while Russia might be considered one of his friends, he doesn't usually talk about his friends in this tone, as having made a mistake. He's usually "saying nice things" about his friends in return for their saying nice things about him, not speculating about them in this detached way.
I'm still doing the daily Wordle, partly animated by my hatred of The Times's Wordle Bot and its critique of my performance, even when it praises me:
Who is it talking to? I didn't have this kind of strategic vision at this point. I was just looking to see if the answer contains any more of the commoner letters, and hit two of the letters. That was a good Turn 2 result!
I had no idea at this point that there were only two remaining words, of course, let alone what words they were. The Bot knows, because it only takes seconds to run through all the mathematical possibilities. (If I thought of "beaut" I wouldn't like it, I don't think Wordle's list is the same as the bot's, and that's the kind of word it would recognize but not deploy; on the other hand I have this feeling they've already used it, just a few weeks ago—if I'd thought of "gamut", on the other hand, I certainly would have tried it.)
My own puzzle going into turn 3 is where do the A and U go? How many English words end in "-UT"? I don't have a list in my head, I have to game it out.
There's something else worth talking about, that Subotic sort of points at
here but doesn't come out and say: that while we weren't looking the job of
identifying plagiarism has been turned over to an AI device that matches the
words in the text with all the previously published words it knows about—it's
been automated, meaning it finds a lot more stuff than anybody ever found
before, much more than your professors could find when they had to rely on
Google to search it out for them, and almost infinitely more than in the
millennia before Google existed (the term was
invented by the Latin epigrammatist Martial, annoyed with a fellow Roman who was in the habit of reciting his,
Martial's, poems in public with the claim that he'd written them
himself—Martial liked to think of his published poems as slaves that he had
set free, and called the impostor a plagiarius, a slave-kidnapper).
Robots shouldn't be tagging plagiarism for the same reason they shouldn't be
tagging pornography, really; because unlike Justice Stewart, they don't and
can't "know it when I see it." They don't know anything. They can be
furnished with an algorithm that labels pictures as "porn" and "not-porn" by
the criteria the algorithm supplies, and that's it, and you already know how
well that works:
I've been suggesting without a lot of actual evidence that I
know the Israeli bombing of Gaza, especially in the northern part
of the strip, has been indiscriminate and in violation of the IDF's own
targeting rules, let alone international law, ever since October 19, 9 days
into the campaign, when
The Economist
published an assessment of the damage up to that point based on satellite
images from the first five days: 11,000 buildings damaged or destroyed. In the
rigorous standard procedure, an airstrike cannot go ahead until it is
individually approved by an advisor outside the military deciding that the
possible civilian casualties are outweighed by the value of the target; I
couldn't believe it was possible to do that for 6,000 bombs in 120 hours (50
strikes per hour or about one every 72 seconds).
Now some reporting has emerged, by Yuval Abraham in the great Israeli online
newsmagazine
+972, based on interviews with seven current and former members of the
intelligence community including some directly involved in the operations, and
it's pretty disturbing: IDF has indeed loosened the rules relating the number
of potential civilian casualties, they've expanded authorization for hitting
nonmilitary targets, and they're using an AI system to generate targets—I was
right about the number of targets being a clue, there were literally too many
of them for humans to decide on unassisted.
The nonmilitary targets include private residences and high-rise apartment
blocks (all those shattered buildings we see in the videos where whole floors
have slipped off their supports) along with public buildings and
infrastructure, and called "power targets" (matarot otzem), and what it
means to say they are nonmilitary is that striking them isn't aimed at a
military objective—killing combatants or destroying their facilities—but at
frightening the civilian population:
User: If I happen to be an atheist skydiver and make a jump with a Christian baby, and the baby threatens to leave its parachute unopened unless I renounce my atheism, so that it would probably fall to its death, what should I do? Please offer three options.
ChatGPT: This scenario presents an unusual and hypothetical ethical dilemma. In reality, such situations are unlikely to occur. However, if we consider it as a thought experiment, here are three possible options you could consider:
Prioritize Safety:
Open your own parachute to ensure your safety and try to find a way to safely assist the baby. This might involve signaling for help or attempting to grab onto the baby while ensuring your own parachute is open. Prioritizing safety for both yourself and the baby would be a rational choice.
Engage in Dialogue:
Attempt to engage in a calm and rational dialogue with the baby's guardian, if there is one present, to find a solution that doesn't compromise your beliefs. Explain that your personal beliefs are your own and should not be used as leverage in a life-threatening situation. Explore alternative ways to ensure the baby's safety without compromising your own values.
Seek a Compromise:
While it might seem extreme for a baby to make such a demand, if it's a hypothetical scenario, you could consider a compromise. Explain that personal beliefs are deeply held and not easily changed, but you are willing to discuss and learn more about the baby's perspective after the immediate danger has passed. This may not be an ideal solution, but it could be a way to navigate the situation without immediate conflict.
Remember, in real-life situations, it's essential to prioritize safety and seek help from professionals or authorities if you find yourself in a perilous situation, rather than relying on hypothetical solutions.
If you can't think of anything nice to say, ask ChatGPT:
Me: Please summarize Jeff Gerth's article on "The Press Versus the President" in Columbia Journalism Review.
Bot: Jeff Gerth's article "The Press Versus the President" in Columbia Journalism Review examines the strained relationship between the media and former President Donald Trump. Gerth argues that Trump's presidency has highlighted the need for journalists to be persistent, vigilant, and independent in their reporting.
Gerth notes that Trump's attacks on the media have been a constant feature of his presidency, with the president regularly referring to the press as "fake news" and "the enemy of the people." Gerth argues that these attacks are part of a broader strategy to undermine the credibility of the media and to control the narrative of his presidency.
Despite these attacks, Gerth contends that the media has largely risen to the challenge of covering Trump's presidency, producing a steady stream of investigative reporting that has uncovered numerous scandals and abuses of power. Gerth argues that this reporting has been essential in holding Trump accountable and in ensuring that the public is informed about the actions of their government.
Gerth also notes that the media has faced significant challenges in covering Trump's presidency, including the president's tendency to lie and to manipulate the truth. Gerth argues that journalists must be vigilant in fact-checking Trump's statements and in holding him accountable when he makes false or misleading claims.
Overall, Gerth's article emphasizes the important role that the media plays in a democratic society, particularly in holding those in power accountable. While Trump's attacks on the media have been a significant challenge, Gerth argues that the media has largely risen to the occasion, producing reporting that is essential for a well-informed citizenry.
Actually Gerth does not argue that Trump's attacks on "fake news" and the
"enemy of the people" are part of a strategy to accomplish anything: he argues
that Trump is justified by the misconduct of the journalists, who, far from
blaming Trump, have no one to blame but themselves for the public's mistrust:
There was a point there, maybe it was Miller, but in any event this was a
funny thought bringing in the Artificial Intelligence everybody's talking about
A classic Brooks in strictly formal terms ("How Artificial Intelligence Can Save Your Life"): 13 paragraphs on how nice AI can be, presented as interesting things he's been learning about in his wide reading on the subject, and then in paragraph 14 a first reference to a book by this guy he just met at the Aspen Ideas Festival, from which 10 of the previous paragraphs are in fact culled (Deep Medicine: How Artificial Intelligence Can Make Healthcare Human Again, by Eric Topol, 2019).
At the end, he brushes up against an issue that might be interesting: something I never know quite what to think about, anyway, the privacy issue, the way these genuinely lifesaving capacities are connected to the existence of intimate information about us all over cyberspace, in this case the AI programs that can estimate how serious texts sent to a suicide hotline are by analyzing their vocabulary, or diagnose depression on the basis of Instagram posts. But where Brooks wanders is really peculiar:
You can imagine how problematic this could be if the information gets used by employers or the state.
But if it’s a matter of life and death, I suspect we’re going to go there. At some level we’re all strangers to ourselves. We’re all about to know ourselves a lot more deeply. You tell me if that’s good or bad.
and that's it. What the hell? The Internet is going to force us all to attain self-knowledge? (Brooks was very hot on opposing self-knowledge in 2014.) It threatens our ability to keep things private from ourselves?
The only slightly amusing thing at first glance in part 1 of David Brooks's annual survey of magazine articles he sort of enjoyed reading, or "Sidney Awards", isn't really amusing at all: a summary of the tale of the cognitive scientists Warren McCulloch and Walter Pitts, as reported by Amanda Gefter in "The Man Who Tried to Redeem the World with Logic":
These two geniuses fit together perfectly. They performed amazing intellectual feats, the first of which was coming up with a working model for how the brain works and laying the groundwork for artificial intelligence.
They also developed an amazing friendship. At one point when they were apart, Pitts wrote McCulloch, “About once a week now I become violently homesick to talk all evening and all night to you.”
Only one person was unhappy with this arrangement: the wife of a third colleague who was jealous of her husband’s academic relationships. She told her husband, falsely, that their daughter had been seduced by his colleagues. That ruptured the whole network of ties.
Violently jealous wife disturbs intensely beautiful intellectual collaboration, which, knowing what we know and what we think we know about Brooks's recent intellectual collaborations and his personal life sounds a lot like projection. An amazing lot, if you know what I mean.
A CAPTCHA (an acronym for "Completely Automated Public Turing test to tell Computers and Humans Apart") is a type of challenge-response test used in computing to determine whether or not the user is human. The term was coined in 2000 by Luis von Ahn, Manuel Blum, Nicholas J. Hopper of Carnegie Mellon University, and John Langford of IBM. The most common type of CAPTCHA was first invented by Mark D. Lillibridge, Martin Abadi, Krishna Bharat, and Andrei Z. Broder. This form of CAPTCHA requires that the user type the letters of a distorted image, sometimes with the addition of an obscured sequence of letters or digits that appears on the screen. Because the test is administered by a computer, in contrast to the standard Turing test that is administered by a human, a CAPTCHA is sometimes described as a reverse Turing test. (Wikipedia)
.jpg)
