If you’re wondering what, exactly, you might gain access to if you subscribe to Sparky’s List, here’s the newsletter that went out this week with an advance peek at the Large Language Muddle cartoon. This one’s probably wonkier than usual, but it gives you a general idea of how I try to provide background notes and a look at my work process, and also just throw in random stuff of interest.
* * *
Okay friends, buckle up! This week’s newsletter is mostly a deep dive into the LLM weeds. If that stuff bores you, please be sure to scroll down to the bottom, where there are some breaking updates from the world of Tomorrow.
I’ve been told that one of my skills as a cartoonist is the ability to condense complicated subjects into a coherent, comic-sized summary. I’d like to believe it’s true, but boy did this week’s cartoon put that to the test. It took me almost three full days to write the text for this cartoon (I can usually do it in one, two at the most), going through endless drafts, figuring out the bare minimum of necessary exposition, and trying to make sure I was properly understanding the analysis of people far more informed on the topic than myself. (Specifically, Cory Doctorow, Ed Zitron, and Cal Newport.)
The core idea that I wanted to convey is that, despite the breathless reports you have seen in the news (including on MS-NOW, from anchors who really should know better), the “rogue a.i. agents” did not, in fact, wake up one morning and decide on their own to launch a cyberattack on the Hugging Face server. Nor did the chatbot direct the attack, because *LLMs are not conscious*. This gets complicated, but as as best I understand it: OpenAI was running an unsupervised cybersecurity test on new large language models, using something called a harness program, which repeatedly prompted the chatbot to find ways to solve a hacking challenge. Which it did, using, at least as one source, its training on decades of transcripts of hacker “capture the flag” contests. The harness then utilized the data the chatbot provided to coordinate the attacks and solve the challenge, though apparently more successfully than anyone expected. (Why anyone thought any of this was a good idea in the first place, I cannot tell you.) Remember, too, that these programs are backed by enormous amounts of computing power (or as the techies call it, “compute,” a coinage that seems as unnatural as using “creative” as a noun, but nobody asked me). This is also how the programs circumvented their internet containment (the “sandbox”). And as Cory explains:
Designing autonomous, malicious software is generally considered irresponsible and dangerous. If you showed up at Defcon and gave a talk about how your autonomous malware did something unexpected and damaged someone else’s computers, the first question from the audience would be “Why are you so shit at making secure sandboxes?” It wouldn’t be “How are you so awesome at making hacking tools?”
One of the methods in the LLM training data, an old hacker tradition, was to use comments sections on random obscure sites to coordinate information, which the so called rogue agents emulated, while also imitating the hyperbolic language of the teenage hackers on which they had been trained. Again, I would urge you to read the posts linked above, especially Cory’s, for much more informed explanations of all of this, but the short story is, the singularity— the Skynet moment when the machines become self aware — is not here yet. The capabilities of autonomous software *are* developing at an alarming rate, and that’s a real concern. But they aren’t little conscious beings who “decided” to do things. They were following their programming and using their training data.
Basically, I ended up condensing all of that into the middle panels, and I hope I did it at least some amount of justice — the fourth panel, and in particular the line “mimicry is not sentience” were written with the misleading “rogue agents” narrative in mind — but of course any misrepresentations here are entirely my own, and should not reflect on the work of Doctorow, Zitron and Newport. Also, special thanks to Cory, who was kind enough to look over early drafts of this cartoon and reassure me that I had a handle on the subject.
This cartoon was also informed, in a more general way, by the book Empire of A.I., which a kind supporter sent from the wish list recently. Another problem with trying to summarize the a.i. industry is that it’s full of wild contradictions. There are people within the labs who genuinely believe that their work is very important but will also lead to the destruction of humanity (the extent to which an individual falls into the “doomer” camp is defined by their self-assigned “p(doom)” number, referring to their estimate of the probability of the robot apocalypse), and others who discount the possibility and insist on charging full speed ahead (the “boomer” faction), so as not to lose a competitive advantage. There is also a lot of speculation about when, not if, self-aware a.i. will emerge. Their goal is “a.g.i.,” or artificial general intelligence (or as I describe it in the cartoon, “a new god which will someday rule over us all!”). In interviews given just before the Hugging Face incident, Altman expressed his belief that the singularity has already occurred, or more accurately, that it is in the process of occurring.
Anyway, I’m looking forward to the pushback this cartoon will receive from the pro-a.i. crowd! That will be a lot of fun. Speaking of feedback, the final panel was initially going to close after Altman’s final word balloon, but then I realized that when you make an “either this or that” statement on the internet, the rest of your week will be consumed with “why not both” replies. So I decided to pre-empt all of that and just include the damn line myself.
*****
And that’s probably enough for this week! At least I got another rare break from writing about you-know-who, though I’m still reading Regime Change for “relaxation.” I don’t think any of this is great for my state of mind, but happily I’ve got a trip to Italy coming up in a few weeks, to take part in an arts and journalism festival held by the newspaper il Manifesto, where some friends of mine work, and which occasionally publishes my photojournalism work. More on that later.
Also, for whatever it’s worth, it looks like one of my big outlets is shutting down at the end of the month, or at least pivoting into something that doesn’t really make sense for me to be part of, so I’m going to be compensating by leaning into my DIY roots and updating thismodernworld.com more often. And if you *wanted* to write places like The Nation or The American Prospect and mention that you miss seeing my fabulous, award-winning work there, I wouldn’t actively *discourage* you. This also holds true if you have a surviving local altweekly that’s not running TMW.
Wish list thanks to what appears to be an intentionally anonymous supporter. And finally, here’s a couple of pictures I took inside the cathedral of St. John the Divine at a memorial service for Sonny Rollins. The globe is a temporary, unrelated installation, and it’s really quite striking — it’s huge and glowing, suspended in mid-air by an almost invisible wire, and slowly rotating.
Until next time,
Dan (very tired, but what else is new)
Wish list
The new Gumroad store, where you can purchase ebooks of TMW25 and OLNN!
Topataco store— t-shirts, signed prints, physical copies of OLNN, plush Sparkies, and more
Threadless store — more ephemeral designs, as well as prints, coffee mugs, other print-on-demand stuff
Personal website (more active than it used to be): www.thismodernworld.com
Instagram (mostly inactive): @realtomtomorrow
Bluesky (excessively active): @tomtomorrow.bsky.
Newsletter (for those of you enjoying a forwarded copy — please remember that this newsletter is intended for paid subscribers who keep my work afloat! Please consider getting your own subscription!!)
Paypal donations