iiinova — where your digital self keeps evolving

Separately, there's clearly a story (that you touch on) about how those of us who read AI slop on the regular immediately sniff it out, while many literary types don't see it a all. What's going on there? Even w/out AI familiarity, it's just…bad prose! Or does it just seem a "really novel voice"?

010

Also, in case you don't have time to read my long thread about precision-recall — I do want to say it was great to see the coverage of the cultural aspects, and from many perspectives (the "I ask you" remark *really* stood out to me). ↵

Bottom line is that the precision-recall tradeoff links the false+ and false- rate. We can't think of them as truly independent. And with adversaries, that linkage really matters. The adversary's goal is to increase false-. At least *some* of the reaction to this may end up increasing the false+. •

000

Obviously things aren't quite this simple, they're scores and not a binary classifier, etc. But it's also possible that, when your dad's essay classification flipped from false- ⭢ true+ from v3 to v4, some innocent student essays got unfairly hauled before a council: true- ⭢ false+. ↵

100

But the precision-recall tradeoff says that as one goes up, the other must go down. So if Pangram decreases the false-negative score, then the false-positive score must go up: more humans will be flagged as AI. Therefore, it doesn't really make sense to talk about one without the other. ↵

100

I think a much bigger issue is that false+ and false- aren't independent due to the precision-recall tradeoff. Here's the problem: Adversary's goal: increase false-negative (AI-written that Pangram says is human). Pangram has to adjust its weights/etc. to decrease the false-negative score. ↵

100

b. The *article* draws no real distinction between + and -. It just says "Pangram is […] highly accurate" (and cites the company!) and no more. It doesn't draw this distinction at all. To readers who aren't familiar with this distinction, they may miss that there are two different directions. But: ↵

100

2. On false ±. You are (of course!) absolutely right: adversarial examples are about false-, not false+. But two comments in response: a. For reasons noted earlier, Pangram's judgment about earlier stories is not meaningful evidence (IMO), and I worry that readers will think that it is. ↵

100

As a cyclist, I think also a lot about the semantics of drug testing. (a) It takes a while to develop a test for a new drug, and (b) if you get an edge by finding some unusual drug, then it's just not going to get caught. The test can be wrong for a long time on many things. (Hi, Lance!) ↵

Many thanks for engaging, Vauhini. Two separate things. 1. I guess the adversarial aspect didn't really stand out in the B article? You're right that it shows up here and there, and there is one para about it ("cat-and-mouse"), but the gap could be quite long and quite undetected. ↵

100

Fall colors starting to arrive in Rhode Island. A lovely afternoon ride along the Blackstone River.

2270

I really would have appreciated a bit more nuance in this piece. I'm guessing a lot of Atlantic readers aren't as savvy as you are, and a lot of people are genuinely confused about this tech, so they're going to be guided by what you wrote. Other than that, great pieces! Thanks.

150

But it's presented there as some kind of extra confirmation about Pangram, coming in a para that begins "AI detection is fallible". This doesn't give us any additional information that AI detection isn't fallible; it only tells us that (as we know) Pangram's creators aren't total idiots. ↵

In the Atlantic, you say "ran all 25 of the previous Caine Prize-winning stories through Pangram, it labeled them as 100 percent human-written". Of course it would! You can bet their RL classifies every every public essay from before 2021 as human, because that's the first thing people will test. ↵

171

So it makes no sense to talk as if the problem is over. We're just in a step of an infinite battle. Everyone who's ever built any kind of classifier (in your Bloomberg article you talk about spam) knows that; "If Pangram is successful" suggests a static universe, which we're not. Secondly: ↵

150

This is what plagues every single attempt at detecting *anything* (not just AI prose: also spam filtering, image recognition, etc., etc.). The moment you release a classifier, adversaries have a target to game. There's no reason to believe that that isn't happening to Pangram too. ↵

191

You call it "highly accurate", referring to your own Bloomberg piece. In that piece, you recognize its iterative improvements, but for v4, your main source is their internal testing and your dad's essay. There is a big thing you're overlooking, which is generative adversarial activity. ↵

160

Yet Another Literary Prizewinner Seems to Be AI

Hi @vauhinivara.bsky.social — I love your work, including this Atlantic piece [ www.theatlantic.com/books/2026/1... ], but was really surprised (not entirely in a good way) by the way you presented Pangram: ↵

theatlantic.com

The prestigious Caine Prize, announced last week, went to a story with many red flags.

192

If it's that clear to you what is doing what, then the "I" shouldn't matter, because it's not confusing you in any way!

Shriram Krishnamurthi @shriram.bsky.social · 14h

Whoa, I first thought this was some AI-generated image…then looked up the news. Yikes.

Shriram Krishnamurthi @shriram.bsky.social · 15h

People actually did that?!? (What am I saying. Of course people did that.)

100

Shriram Krishnamurthi @shriram.bsky.social · 15h

@notjack.space I apologize for creating a serious #lang, I promise not to make a habit of it.

Shriram Krishnamurthi @shriram.bsky.social · 15h

You can tell me if they aren't. (Assuming you aren't a bot. Which, if you are, may make it even easier.)

000

Shriram Krishnamurthi @shriram.bsky.social · 19h

For illustration, here's an excerpt from an example in the distribution: github.com/shriram/liqu...

Shriram Krishnamurthi @shriram.bsky.social · 19h

The code repo is here: github.com/shriram/liqu... Installing gives you full docs. Examples are here: github.com/shriram/liqu... Racket package here: pkgd.racket-lang.org/pkgn/package... ↵

github.com

GitHub - shriram/liquid-shplait: Extension of Shplait with Liquid Types

Extension of Shplait with Liquid Types. Contribute to shriram/liquid-shplait development by creating an account on GitHub.

Shriram Krishnamurthi @shriram.bsky.social · 19h

Pleased to announce shplait/liquid, a liquid-typed version of shplait, the language of my book PLAI. This version adds liquid-typing, a form of refinement types. It's primarily meant to be a small language for pedagogic purposes to induce people to try grown-up ones later. ↵

3186

Shriram Krishnamurthi @shriram.bsky.social · 21h

I'm impressed (in a good way!) that this review of a book about the tragic killing of Malcolm Kerr, American University of Beirut president, never once mentions his son. www.nytimes.com/2026/10/12/b...

nytimes.com

Book Review: ‘The Best Kind of American,’ by Kim Ghattas

In “The Best Kind of American,” Kim Ghattas shows what an American’s assassination in Lebanon four decades ago reveals about chaos in the Middle East.

Shriram Krishnamurthi @shriram.bsky.social · 12/10/2026

You can always write to Yaron and ask. (-:

Shriram Krishnamurthi @shriram.bsky.social · 12/10/2026

After months of getting things done w/ Claude Code, I've spent ~5 days programming by hand in Lean and a liquid typed language and it's…SO GREAT? I love living in a world where I can get 1000s of mindless LOC in a minute, but also spend an hour on 10 gorgeous handcrafted lines.

2301

Shriram Krishnamurthi @shriram.bsky.social · 11/10/2026

Season 2 of @jarrodkimber.bsky.social's Double Century is one of the best abbreviated reports on race in cricket. Ostensibly about Basil D'Oliveira, it goes over a much longer history, even talking about US sports. This is the last episode, but listen to all 5. pocketcasts.com/podcast/comm...

pocketcasts.com

2.05 - The Greatest Hundred

The conclusion to our story on race in cricket and the Basil D’Oliveira story. This is the final episode of this season. - This podcast is brought to you by our supporters at Patreon, if you like…

Kate Sills @katelynsills.com · 11/10/2026

Reposted by Shriram Krishnamurthi

As of Jan 2027, California requires every citation be "personally verified" by an attorney. Not every change I requested made it into the final text (in particular, excluding paralegals makes law more costly for all), but I'm proud that some changes made it in. www.ropesgray.com/en/insights/...

Personal Verification of Citations in Court Filings
Separately, SB 574 amends Code of Civil Procedure Section 128.7 to provide that no “brief, pleading, motion, or any other paper filed in any court shall contain any citations that an attorney responsible for submitting the pleading has not personally verified, including any citation provided by generative artificial intelligence” (emphasis added).12 This provision covers citations regardless of their source—whether from generative AI or otherwise.

“Personally verified” is not explicitly defined, but the plain language suggests that the responsible attorney must independently check each citation of a “brief, pleading, motion, or any other paper filed in court.” The statute likewise does not define which attorney or attorneys are “responsible for submitting” a filing when multiple lawyers sign or appear on a court submission, leaving open another key interpretative question.

1144

Shriram Krishnamurthi @shriram.bsky.social · 11/10/2026

Signals and Threads episodes are bangers, but this may be the first to become part of my PL course's "text". About 50% is a compressed version of my GC lectures, about 30% is modern updates, and 20% is @yminsky doing his thing w/ Stephen Dolan. (-: signals-threads.simplecast.com/episodes/mem...

signals-threads.simplecast.com

Memory Management with Stephen Dolan | Signals and Threads

Stephen Dolan works on Jane Street’s Tools and Compilers team where he focuses on the OCaml compiler. In this episode, Stephen and Ron take a trip down memory lane, discussing how to manage computer m...

2243

Shriram Krishnamurthi @shriram.bsky.social · 10/10/2026

Great, you were a tool to big corporations. Still doesn't mean you know a darn thing about academic publishing—indeed, you've demonstrated you don't—as separate from writing software, which are two entirely different things.

300

Shriram Krishnamurthi @shriram.bsky.social · 10/10/2026

Apparently equally easy when you're completely wrong about things you know nothing about (in this case, academic publishing). So don't pat yourself too hard on the back.

100

Shriram Krishnamurthi @shriram.bsky.social · 10/10/2026

Oh, you've already written plenty of poasts. Looks like you're on a roll, rock on.

100

Shriram Krishnamurthi @shriram.bsky.social · 10/10/2026

I've shipped real code and have real users.

100

Shriram Krishnamurthi @shriram.bsky.social · 10/10/2026

That feature is ridiculous, but it's different.

000

Shriram Krishnamurthi @shriram.bsky.social · 10/10/2026

Yes, it explains why I have expertise on this and you presumably do not.

100

Shriram Krishnamurthi @shriram.bsky.social · 10/10/2026

You can't "file a bug and fix" for a paper. Please stop poasting.

400

Shriram Krishnamurthi @shriram.bsky.social · 10/10/2026

Adolfo is talking about software engineering research papers. You're talking about something completely different.

200

Shriram Krishnamurthi @shriram.bsky.social · 10/10/2026

1. The entire peer review structure is about to tip over. 2. The whole research enterprise itself may change dramatically, which makes some of these issues feel almost quaint. There is some reason to believe that the old discipline of math is basically over thanks to AI.

000

Shriram Krishnamurthi @shriram.bsky.social · 10/10/2026

The bottom line is, we should call out bad research because it's bad research. LLMs are an obfuscation here. And by over-focusing on that, we're going to under-focus on OTHER forms of poor research quality. Anyway, at this point two things are becoming clear: ↵

Shriram Krishnamurthi @shriram.bsky.social · 10/10/2026

Ed Zitron is a good comparison. I will leave it at that.

000

Shriram Krishnamurthi @shriram.bsky.social · 10/10/2026

Yes, I agree. There has been a lot of junk. Reviewers were overly excited by superficial results. CS Ed is way, way, way worse. There are groups just churning out paper after paper of things that are obvious or rubbish. The field is discrediting itself. It's why I've given up on the area.