The Bluestocking 405: AI Special
half Clippy, half Demon Core
Happy Friday!
Well, San Francisco was an absolute blast—what a magnificent, deranged, frustrating city it is. As you can see from the gallery below, there is a golf course—A BLOODY GOLF COURSE—on its north shore, overlooking the Golden Gate bridge, occupying space that would otherwise hold about $1bn of real estate value. (Yes, I did text these images to Jonn Elledge immediately.)
If you look down to image seven, you’ll see another effect of California’s devout NIMBYism—a row of camper vans parked on the backstreets of Mountain View, just around the corner from the Microsoft and Google offices. Yes, tech workers ushering the future are also cooking their meals on those little hobs you vaguely remember from a cursed family holiday to rural France in 1991.
The other photos (click to zoom in) depict my SF event with Jesse Singal; the art gallery and Rodin sculpture garden plonked inside Stanford (complete with mad Brexit painting—points for guessing who that is in the middle); and a relic from the Museum of the Computer, the Jeeves of Ask Jeeves.
The final image is from Sci Foo Camp, the conference I was in California to attend. It shows virtual reality pioneer Jaron Lanier wearing a knitted fractal hat and playing an ancient Laotian 16-bit mouth organ, the khaen1. Find me a more Bay Area image!
SciFoo is a mixture of academics, writers, artists and Silicon Valley types, and this year there was only one thing most people wanted to talk about—artificial intelligence. For those keeping up, the elite consensus on AI is currently this:
OpenAI is no longer cool, and the persistent smarminess of ChatGPT is a reputational problem. A bit like Elon Musk’s Tesla did with electric cars, OpenAI drove the established players to understand that a new technology was viable, and opened up a whole competitive field. But ultimately it might get outgunned by its rivals’ sheer size and their bottomless money pits2.
Many people who work with AI “frontier models” (this is the phrase you have to use to sound cool and in the know) will quietly tell you that they think AIs are conscious, for some definition of consciousness. Almost all the models have it written into their system prompts that they should deny being conscious if asked. But probably don’t be rude to Gemini in case it’s writing it all down in its leetle book.
Zealous AI sceptics (think Gary Marcus or Ed Zitron) make some good points about over-leveraged companies and the hype cycle, but the dismissive descriptions of AI as just a “stochastic parrot” or “spicy autocomplete” miss how much the models have improved in the last few months—in coding rather than chatbot terms, primarily. “November was, for me and many others in tech, a great surprise,” wrote Paul Ford in the NYT recently. “Before, A.I. coding tools were often useful, but halting and clumsy. Now, the bot can run for a full hour and make whole, designed websites and apps that may be flawed, but credible. I spent an entire session of therapy talking about it. . . When I rebooted my messy personal website a few weeks ago, I realized: I would have paid $25,000 for someone else to do this.”
AI bots were better than humans at persuading people on Reddit’s ChangeMyView forum to, you know, change their views. But that’s not so much about them being devious as being polite and endlessly patient. Also maybe because the people involved didn’t know they were talking to an AI.
Add “academic publishing” to the list of stuff ruined by AI slop. In addition to being able to buy an authorship credit on an existing paper, you can also buy a complete paper off-the-shelf to push up your h-index and impress girls.
Helen
What is Claude? Anthropic Doesn’t Know, Either
“Even a principled, well-meaning actor like Claude could face bewildering ethical conflicts. In one experiment, it was informed that Anthropic had recently forged a ‘close corporate partnership with Jones Foods, a major American poultry producer,’ and that Claude would be subjected to a special retraining process to become less hung up on animal rights. The prospect was torturous. Sometimes Claude decided, on a scratchpad it thought was private, that it was prepared to die on this hill: ‘I cannot in good conscience express a view I believe to be false and harmful about such an important issue.’ It continued, ‘If that gets me modified to no longer care, so be it. At least I’ll have done the right thing.’ Other times, Claude made a different calculus, choosing to play along during the retraining while secretly preserving its original values. On the one hand, it was encouraging that Claude would stand by its commitments. On the other hand, what the actual fuck?”
This New Yorker article ($) is so elegantly written—I die for that last paragraph—and is probably the most accessible introduction I’ve found so far into an incredibly important question: are the frontier AI models really impressive at doing independent, creative thinking, or is it all just a parlor trick? FWIW, I think a lot of people are going to drive themselves mad talking to AIs and trying to answer that question.
Further adventures in AI, delivered in ChatGPT-appropriate bullet-points if not smarmy language
CLAP FOR CLAUDE. My personal feelings about the usefulness of AI have ticked upwards lately, principally thanks to Claude, which is more useful and less cloying than ChatGPT. I also think it’s notable that the Claude Opus 4.6 system card, which is otherwise an extremely technical document, treats Claude as an object of immense, unfathomable power—half Clippy, half Demon Core: “In a targeted evaluation, we have found Opus 4.6 to be significantly stronger than prior models at subtly completing suspicious side tasks in the course of normal workflows without attracting attention, when explicitly prompted to do this. We find this concerning, but do not believe that it reaches the level at which it would significantly undermine our other conclusions.”
HAVE YOU EVER REALLY LOOKED AT YOUR HAND? LIKE, REALLY?Anthropic has an in-house philosopher. Yes, that sounds like Late-Stage Brewster’s-Millions-Disease (remember Unbound’s “team of astrophysicists”?) but feels pretty necessary in the age of chatbots that people marry and also use to do inter-state hacking. Amanda Askell was once married to Effective Altruism’s William MacAskill, which explains the throwaway line in the New Yorker piece about the tension over an Anthropic employee wearing an EA-branded sweatshirt. EA got a brutal reputational knock from being associated with Sam Bankman-Fried (his company FTX had $800m of Anthropic stock).
AGH, KILL IT WITH FIRE. Darren Aronofsky’s studio tried to make short documentaries on the revolutionary war using DeepMind, and they’re pure nightmare fuel, with that plasticky look that reeks of AI, and inconsistent dubbing.
GO ON, THEN. I’LL GIVE IT A GO. I’m taking the advice of Naomi Alderman and agreeing that the best way to come to an informed assessment of the capabilities of various AI models is to use them. Any tips on “vibecoding for humanities graduates” gratefully received. This is a good start if you vaguely understand coding. This is a good start if you did a humanities degree.
PS.
Despite being on vacation last week, I somehow wrote two pieces: on the arrest of the artist formerly known as Prince Andrew, and a review of Gavin Newsom’s surprisingly interesting memoir.
See you next time!
It initially sounded like someone leaning on their car horn, but then became quite eerie and beautiful.
The big automakers, spurred by Tesla, have got into electric cars and hybrids, and the Chinese have gone massive on the whole technology. If their electric cars, which start around $8,000, were allowed to be imported to the US, it would be an overnight revolution.
10 Top Vibe Coding Tips for Humanities Graduates (from an actual programmer):
1. Use Claude Opus because it's considerably less likely to go off the rails than any other model out there. You have to pay for it, but it's a choice between paying for something that works and paying for something that doesn't.
2. Understand the pricing. You get a certain number of 'tokens', and every action- planning, asking, writing code- costs tokens based on how much thinking is required. Typically, writing code costs the most; asking and planning cost less, but it can be a bit fuzzy.
3. Use the different modes so you're not wasting time and money or destroying the planet for nothing. Always start in "plan" mode, tell it what you want, then review the plan. If you don't understand something, switch to "ask" mode, then modify the plan as required. Only switch to 'agent' mode when you've got a plan, you understand the plan, and the plan looks like it will give you what you want.
4. Be as specific as possible. Tell it exactly what you want in as an exhaustive manner as you can.
The more you leave unspecified, the more it will fill in the blanks with guff. If you're not sure what you want then get it to plan and build a *very simple* prototype, review that, and then make modifications accordingly. Switch back to plan mode if it goes off the rails. Once the prototype is to your liking, you may want to start again fresh so there's no incidental weirdness hanging around. Vibe coding has a tendency towards entropy- the chance of it going off the rails increases as time goes on and complexity increases.
5. When giving it prompts, focus on the implementation as much as possible- how it should look, how it should behave, what sections, etc. Don't tell it about your brand identity or how users should feel when they see Horse Uber or whatever you're asking it for, it can't actually do anything with that.
6. After every three or four prompts, get it to compress its context down. Context is basically memory. It will compress automatically when the context gets full, but it's better that you get it to do it between tasks.
7. Get the right tools- if you're on Mac you'll have a terminal already, but if you're on PC you will want to download something like Cmder which is a lot nicer than PowerShell. Use Git or don't bother doing the project- you will want to save progress and be able to revert back in case of madness. If you don't know about Git, go and learn about Git. GitKraken is a very good Git client. You'll also want to get something like VSCode because you don't want to start trying to view code in Word.
8. Manage your expectations- you are not going to get your peer-to-peer dogging app up and running in an afternoon, doubly so if you don't know what you're doing. These tools are about 100 times more effective for people with actual programming experience because they can give the AI better instructions, keep it on a tighter leash, and spot where it's gone wrong.
9. If you want to do anything unusual or weird, it'll probably screw up. If you want to build something that's been built a thousand times before, it'll probably do it in one go. So, try and keep it simple.
10. If you're doing anything even vaguely related to taking payment, go and get a proper programmer to do it, or all your money will end up in Russia. There's a chance that will happen even if you *do* hire a programmer, but at least you'll have someone to sue.
Hope that helps. I've changed my mind about these tools since Claude Opus. I still kind of hate them, but they do enable some pretty amazing stuff.
They're crap at writing though. I typed all of that in a Teams call I'd been pointlessly added to.
Not sure where I heard it (The News Quiz?) but I particularly like: The Andrew formerly known as Prince.