The most recent issue of the FT Weekend Magazine is about games of all kinds, but the highlight is a massive article about the tragedy of Disco Elysium. It is depressing throughout, with a glimmer of hope buried near the end:

Kurvitz is making his next game at a new studio, Red Info, with Aleksander Rostov, Helen Hindpere and Chris Avellone, lead writer of the 1999 video game Planescape: Torment, a huge influence on Disco. “[Kurvitz] felt that Disco was the project in his head, and once he was cut off from the franchise, he was worried he didn’t have any other ideas in him,” Avellone told me. “I felt that was bullshit . . . Robert’s too creative to simply ‘not’ create something or rely on a single world idea in his head.”

Creators of Planescape: Torment and Disco Elysium working together on a new game? Be still, my heart.

• From the FT Editorial Board: The risk of letting AI do your thinking (the only nit I have to pick here is that by “AI” the esteemed Board means “LLMs” or, if you want to be kind and stretch the definition of intelligence, “generative AI”)
• From Dave Winer: AI should behave like a computer (see previous note)
• From a person on the Internet: Cognitive Hygiene: Why You Need to Make Thinking Hard Again (file under YouTube videos that should have been blog posts)
• From Michael Lopp: Every Single Human. Like. Always. (subtitle: The robots… They did the thing.)

And with these four links I hereby declare a moratorium on LLM-related matters on this blog, until further notice.

For a contrarian take on LLMs as intelligent machines, here is Alexey Guzey saying that:

• ChatGPT understands what it reads
• LLMs are creative
• Compression is intelligence and ChatGPT compresses really well
• The idea of AGI is stupid
• It doesn’t matter if AGI is real or not

I remain dubious.

An important note from Dave Winer:

I say ChatGPT instead of “AI” because I’m not comfortable characterizing it as intelligence. Deeper you get into it you learn that these beings whatever they are have serious character flaws that are counter-intelligent.

Exactly. LLMs are closer in intelligence to a screwdriver than to a human.

Flighty does not seem to be as up-to-date traveling internationally as it is on domestic flights. The IST airport departures board had our flight listed as delayed as soon as we got there, yet the app thought everything was fine. Trust no one.

Casey Handmer on LLMs:

Every time one of the labs releases an updated model I give it a thorough shakedown on physics, in the style of the oral examination that is still used in Europe and a few other places. Claude, Grok, Gemini, and GPT are all advancing by leaps and bounds on a wide variety of evals, some of which include rather advanced or technical questions in both math and science, including Physics Olympiad-style problems, or grad school qualifying exams.

And yet, none of these models would be able to pass the physicist Turing test. It’s not even a matter of knowledge, I know of reasonably talented middle schoolers with no specialized physics training who could reason and infer on some of these basic questions in a much more fluent and intuitive way.

Alexander the Great had Aristotle, some poor kid will have a brain-dead version of Wheatley.

(Casey’s post is deeper than simple LLM-trashing for he gives the actual 8-step process of reasoning through physics problems, so please do read the whole thing.)

In what feels like a troll but is in all likelihood completely serious, some parents are deciding to have their children fully immersed in AI LLMs:

We’re declaring bankruptcy on humans. Bring on the AI. In addition to integrating AI into as many facets of our lives as possible (our health, our work, our entertainment, and our personal lives), we’re designing an AI-integrated childhood for our kids—all while feeling like we’re helping them dodge a major bullet.

Did CS Lewis suspect, when he wrote The Abolition of Man, that the anti-human sentiment would be expressed as freely and overtly as the first sentence of this intellectually bankrupt paragraph? A paragraph that would be horrifying even if the AI it touts was actual intelligence, an AGI, but what these families are actually immersing themselves in is industrial-grade bullshit. As useful as bullshit can be — I hear it makes for great fertilizer! — one should not drink it as one would do Kool-Aid.

📚 Finished reading: In the Beginning… Was the Command Line by Neal Stephenson, almost thirty years old and more relevant than ever. Download it for free here, and if you think you don’t have time for all 65 pages Chapter 12 about the Hole Hawg should motivate you to read the entire essay.

When I wrote about formalizing AI “peer” review I meant it as a tongue-in-cheek comment on the shoddy human peer review we are getting anyway. “Wittgenstein’s ruler: Unless you have confidence in the ruler’s reliability, if you use a ruler to measure a table you may also be using the table to measure the ruler. The less you trust a ruler’s reliability (in probability called the prior), the more information you are getting about the ruler and the less about the table.”, Nassim Taleb in Fooled by Randomness. Peer reviewers are the ruler, the articles are the table, and there is zero trust in the ruler’s reliability. It was also (1) a bet that the median AI review would soon be better than the median human review (and remember, the median journal article is not submitted to Nature or Cell but to a journal that’s teetering on being predatory), and (2) a prediction that the median journal is already getting “peer” reviews mostly or totally “written” by LLMs.

Things have progressed since January on both of these fronts. In a textbook example of the left hand not knowing what the right hand is doing, some journals are (unintentionally?) steering their reviewers towards using AI while at the same time prohibiting AI from being used. And some unscrupulous authors are using hidden prompts to steer LLM review their way (↬Andrew Gelman). On the other hand, I have just spent around 4 hours reviewing a paper without using any AI help whatsoever, and it was fun. More generally, despite occasionally writing about how useful LLMs can be, my use of ChatGPT has significantly decreased since I fawned over deep research.

Maybe I should be using it more. Doc Searls just wrote about LLM-driven “Education 3.0”, with some help from a sycophantic ChatGPT which framed eduction 1.0 as “deeply human, slow, and intimate” (think ancient Greeks, the Socratic method and the medieval Universities), 2.0 as “mechanized, fast, and impersonal” (from the industrial revolution until now), and 3.0 as “fast and personal”. Should I then just let my kids use LLMs whenever, unsupervised, like Neal Stephenson’s Primer (“an interactive book that will adapt as the user grows and learns”)? But then would I want my kids hanging out with a professional bullshitter? Helen Beetham has a completely contrarian stance — that AI is the opposite of education — and her argument is more salient, at least if we take AI to mean only LLMs. Hope lies eternal that somebody somewhere is developing actual artificial intelligence which could one day lead to such wonderful things as the “Young Lady’s Illustrated Primer”.

Note the emphasis on speed in the framing of Education 3.0. ⊕ I am less concerned about LLM bullshit outside of education, in a professional setting, since part of becoming a professional is learning how to identify bullshitters in your area of expertise. But bullshit is an obstacle to learning: this is why during medical school in Serbia I opted for reading textbooks in English rather than inept translations to Serbian made by professors with an aptitude for bulshitting around ambiguity. This is, I suppose, the key reason why we need LLMs there in the first place for there is nothing stopping a motivated learner from browsing wikipedia, reading any number of freely available masterworks online, watching university lectures on YouTube, and interacting with professionals and fellow learners via email, social networks, Reddit and what not. But you need to be motivated either way: to be able to wait and learn without immediate feedback in a world without LLMs, or to be able to wade through hallucinations and bullshit that LLMs can generate immediately. Education faces a bootstrapping problem here, for how can you recognize LLM hallucinations in a field you yourself are just learning?

The through-line for all this is motivation. If you review papers in order to check a career development box, to get O1 visa/EB1 green card status, and/or get brownie points from a journal I suspect you would see it as a waste of time and take any possible shortcut. But if you review papers because of a sense of duty, for fun, or to satisfy a sadistic streak — perhaps all three! — why would you want to deprive yourself of the work? Education is the same: if you are learning for the sake of learning, why would you want to speed it up? Do you also listen to podcasts and watch YouTube lectures at 2x? Of course, many people are not into scientia gratia scientiae and are doing it to get somewhere or become something, in which case Education 2.0 should be right up their alley, along with the playback speed throttle.

📚 Finished reading: Thinking With Tinderbox by Mark Bernstein, after starting two months ago. It is broader in scope and less Tinderbox-specific than The Tinderbox Way, his first book about, well, Tinderbox, a lovingly crafted “tool for thinking” that I have been using off and on for the last seven years. This is for the best: The Tinderbox Way was meant to convert the technical language from the official code reference into something us muggles can use, which is a job that ChatGPT can do much better and using the latest version of the app. Thinking With Tinderbox is more strategic than tactical, elaborating on why anyone whose primary job is not programming would want to dabble in code in the first place.