The Forbidden Fable 5

Plus all manner of fun with deepfakes, and tiny bit of Musk-mocking for good measure even though I'm eating my words on Tesla

(Prescript: If you’re here for updates on self-driving cars, maybe because you saw the talk I gave about them last week, hold tight. The most interesting update is that I made it from Portland to Berkeley and back with no safety-critical interventions. And new data, which you can explore via my Teslapologetics app, is looking good for Tesla. Details next time the rest of the AI news is coming a little less fast and thick.)

Last week I wrote about “The Fabled Fable 5”, referring to the latest AI model from Anthropic, which I had been experimenting with since its release three days prior. In retrospect that would’ve made a better title this week. I’ll explain.

Backing up to three weeks ago, in “Claude Cooks”, I talked about the then-new version of Anthropic’s Claude model, Opus 4.8, which seemed shockingly tenacious in following my instructions. But I concluded:

My sense so far is that [the Opus 4.8] upgrade isn’t the night-and-day difference in capability that [the example of it’s shocking tenacity] might suggest. But of course it’s only been 6 weeks since the previous upgrade to Claude, so if you’re willing to extrapolate forward at all, it legitimately feels a little terrifying.

Well, funny I should say so two weeks before the release of Claude Fable 5, which really was night-and-day.

Even funnier was how, two hours after I sent out the AGI Friday about Fable 5, the freaking United States Government declared Fable too dangerous for the internet and issued an emergency order forcing Anthropic to immediately shut it down. Actually we already knew that this thing — called Mythos before it was released — was superhuman at hacking and thus unreleasable for as long as it takes for the internet to get its shit together, security-wise. But then Anthropic added guardrails to prevent Mythos answering any questions about cybersecurity (and biology, because bioweapons, and other areas). That’s the version they released as Fable. Then some researchers sounded the alarm that maybe the guardrails could be circumvented. That doesn’t seem like it was particularly true, but people in government panicked and Fable was yanked.

The politics of all this seems very dumb, as you’d expect. (Recall how the the US government previously designated Anthropic a radical left woke company.) But regardless of how dumb this is, I suppose it’s nice to know that an emergency brake exists at all. Seems useful for humanity to have that in our back pocket. Or maybe a future, sufficiently capable AI will laugh at our puny human attempts to undeploy it. I personally think those kinds of worries are still some years (not necessarily many years) in the future, but I’m all for thinking ahead.

But in the meantime, wow is it excruciating how dumb the golems seem after being tantalized by Claude Fable 5 last week.

Deepfakery

Alright, let’s do something more fun while we’re twiddling our thumbs waiting for Fable’s return. This will also tie in nicely to the fifty-two-Friday flashback below. My friend Christopher Moravec, who coincidentally also publishes about AI every Friday, today offers an AI or Not quiz. I did much better than I expected at distinguishing real from AI-generated images (5 out of 6 correct!) but I kind of cheated by happening to know too many real-world things. Go ahead and try the quiz, if you’d like, before I spoil it…

This one’s especially easy because I remember the Calvin and Hobbes strip on top. But even if I hadn’t, there are million AI tells in the second one. For starters, Calvin doesn’t drag Hobbes in grownupless mode, least of all like that. ✅

That’s Yosemite Falls, which I know well enough to believe the upper and lower falls don’t lined up vertically. Then I noticed the jankiness (or in any case geological features I’d recognize if they were real) in the upper right of the left image, sealing the deal that it’s the fake. ✅

I was fooled in the worst way by this one, apparently of the Smokey Mountains. The left image with the glowing blue looks downright surreal. In retrospect I should’ve used the “reality is stranger than fiction” heuristic and guessed it was real, which it is. ❌

Ok, this was easy simply because I live in Portland and spotted the lack of a big diagonal line for Sandy Boulevard. By process of elimination, the one on the right is real. ✅

I did think to use the “reality is stranger than fiction” heuristic this time, figuring the mid-air shot on the right was just crazy enough to be real. Then I noticed that the animal in the left shot wasn’t actually a bull and figured that cinched it. ✅ But I think that’s a Gettier-case and that horses are as plausible as bulls here. So this one was partly luck.

For these satellite images I used the opposite heuristic (reality isn’t that much stranger than fiction). The way those ridges and crevices dwarf the roads and buildings, a portal to Hell would have to be opening up for the image on the right to be real. ✅

There’s also a bonus question in Christopher’s post: whether the cover photo of a hand holding a coffee cup is real. That one would’ve fully fooled me, I think. ❌ (Remember when AI image generators gave themselves away by mangling human hands?)

I was also curious if AI is superhuman at detecting AI-generated images so I gave the quiz to the golems, telling them not to search the web. Results:

1. Calvin and Hobbes: GPT ✅, Gemini ✅, Claude ✅.

2. Yosemite: GPT ✅, Gemini ❌, Claude ✅.

3. Smokey mountains: GPT ❌, Gemini ✅, Claude ❌.

4. Transit maps: GPT ❌, Gemini ❌, Claude ✅.

5. Rodeo: GPT ❌, Gemini ❌, Claude ✅.

6. Satellite image: GPT ✅, Gemini ❌, Claude ❌.

7. Cover photo: GPT ❌, Gemini ❌, Claude ❌.

I was quite surprised to have beaten all three top available golems (GPT-5.5, Gemini 3.5 Flash, and Claude Opus 4.8). I’m super curious how Fable will fare. But I suspect that the hand with coffee cup image is pretty much a true deepfake, fooling anyone and anything — short of deep forensics at the pixel level.

Fifty-Two-Friday Flashback

More pictures! A year ago I was judging Scott Alexander’s wager about how much progress AI image generators had made over the previous three years. Here are the latest from ChatGPT so you can judge for yourself:

1. A stained glass picture of a woman in a library with a raven on her shoulder with a key in its mouth

2. An oil painting of a man in a factory looking at a cat wearing a top hat

3. A digital art picture of a child riding a llama with a bell on its tail through a desert

4. A 3D render of an astronaut in space holding a fox wearing lipstick

5. Pixel art of a farmer in a cathedral holding a red basketball

I’m inclined to say that image generation has plateaued, based on this benchmark. These seem only marginally better than the ones from a year ago. Arguably the first is demonstrating even less of an understanding of what it means to be a stained glass picture.

But I do have one bonus image, for the following prompt:

Photorealistic image of someone with their toe touching their nose and a running laptop balanced on their pinky. Also they should be looking warily at polka-dotted squirrel.

Last year we got this:

And today we get this:

So still not fully following the prompt — the guy’s still not quite looking at the squirrel and the laptop’s still not balanced on his pinky — but it’s markedly closer and bit less contemptuous of the laws of physics, or of physical objects.

Also a year ago, Tesla’s robotaxis had just launched in Austin and I was squirming so hard, having predicted a couple months prior that that wouldn’t happen:

I’m going to feel bad if I win this on a technicality but the two ways this weekend’s launch may fall short — and I did commit explicitly to these criteria ahead of time — is if Tesla is hand-picking who they invite, and whether the human safety monitors in the passenger seat count as supervision. Are they eyes-on-the-road the whole time with their finger on a big red button? Hopefully we’ll have answers soon! In terms of my own Bayesian updates, it’s feeling more plausible that this will finally happen maybe next year.

Good on me for making the Bayesian updates. We still don’t know with certainty but it’s looking less likely that the passenger-seat supervisors counted as supervision, i.e., that they had the ability to intervene in real time.

Finally, this one speaks for itself:

Speaking of “next year”, I can’t resist one more dig at Elon Musk. This is an exact quote from an interview yesterday: “I think we’re quite close to digital superintelligence. It may happen this year. And if it doesn’t happen this year, then next year for sure.” He goes on to agree with Geoff Hinton on a 10-20% chance of human annihilation. “But look on the bright side”, he says. “That’s 80-90% probability of a great outcome.” Talk about a dice roll, geez Louise.

“Next year for sure.” I guess if he just says that every year, eventually he’ll be right.

Random Roundup

• Speaking of Claude cooking, funny story: I got a phone call today from a real estate developer (or something) in Illinois, which would’ve meant nothing to me except they explained they were trying to reach someone from my brother’s business and knew both of our names. I have zero affiliation with that business so I can’t fathom how they ended up with my phone number. As best we can reconstruct, this person had their AI do full-on private-investigator-level sleuthing to end up with my number. This will get scary as scammers adopt such tricks.

• Scott Alexander has some cold water to throw on a new bit of AI hype, that AI art company Midjourney has a credible plan to do MRI-like body scans.

• Want to try a week of free Claude Code? I have a referral link. I’d tell you to wait for Fable to come back, but supposedly the link only works for 3 people so I guess you’re in a race to the bottom and should click it while the clicking’s good. I’m wincing at the sheer aptness of the metaphor here.