Ted Underwood

https://bsky.brid.gy/r/https://bsky.app/profile/tedunderwood.com

Uses machine learning to study literary imagination, and vice-versa. Likely to share news about AI & computational social science / Sozialwissenschaft / 社会科学

Information Sciences and English, UIUC. Distant Horizons (Chicago, 2019). [tedunderwood.com](http://tedunderwood.com )

🌉 [bridged](https://fed.brid.gy/bsky/tedunderwood.com ) from 🦋 [tedunderwood.com](https://bsky.app/profile/tedunderwood.com ), follow Bridgy Fed for the Atmosphere (Bluesky) (npub1pqr…2sck) to interact

Web site: https://bsky.app/profile/tedunderwood.com
Link: https://tedunderwood.com

Public Key

npub1hj5gjcjgmtqmesnc7eadm3cd5ckxxmapwrvnupw9k2l05slfj02s6g3c79

NIP-05 Address

tedunderwood.com_at_bsky.brid.gy@momostr.pink

Profile Code

nprofile1qqste2yfvfyd4sducfu0v7kacux6vtrrd7shpkf7qhzm90h6g05e84gprpmhxue69uhhyetvv9ujumt0d4hhxarj9ecxjmntqy8hwumn8ghj7mn0wd68ytnwdams3gamp7

Publishing to

relay.momostr.pink

nostr.now

relay.primal.net

Author Public Key

npub1hj5gjcjgmtqmesnc7eadm3cd5ckxxmapwrvnupw9k2l05slfj02s6g3c79

Show more details

Last Notes

2026-11-01 17:49:20 UTC

- reply

These are pretty close to my own norms rn. I particularly like "no AI-authored text in papers" paired with "no submission before AI review." That's a forceful affirmation that we can and will use this technology to make our work stronger and more careful.
RE: https://bsky.app/profile/did:plc:kkij4bfbznvqomfmzv36ksgw/post/3mz45r63if22x

2026-11-01 17:38:09 UTC

- reply

Here's the complexity of the picture just within sociology. My impression is that in CS most of these bars would shift 10-20% to the right. Meanwhile, there is a robust group of humanists who believe Everything Is Orange.
RE: https://bsky.app/profile/did:plc:gf7kmaj3i3cq57mzcwn6bt57/post/3myzm6wv3cc2f

2026-11-01 17:38:09 UTC

As someone who collaborates across hum / soc. sci / CS I can confirm that disciplinary norms re: AI are utterly different. Publishers who publish across a range of disciplines are going to have an interesting, complex decade
RE: https://bsky.app/profile/did:plc:ohk4pgrwhyhlp7zwrp3rzlj4/post/3mzb2swafgk2f

2026-11-01 13:52:31 UTC

“I think the best we -- I mean you and me here -- can do is admit to and respect the fact that we do not know each other. And this sort of makes things exciting.”
RE: https://bsky.app/profile/did:plc:dl5wrgqu2qqhmdf3flsylvhc/post/3mz372wkutk2q

2026-11-01 11:22:28 UTC

- reply

True, John Milton made it work, but he also made English sound like Latin; don’t try that at home

2026-11-01 11:14:37 UTC

Tony Gilroy after reading LOTR: I could see a prequel about the price paid to discover the Ring.
George Lucas: What we really need is a trilogy about how Sauron (semi-understandably) became a bad guy in the first place.

2026-11-01 11:14:37 UTC

- reply

The original sin of the prequel trilogy was not understanding the difference between “mythic backstory“ and “prequel.“ Mythic backstory needs to stay mythic.

2026-10-31 18:42:07 UTC

- reply

Where Lévi-Strauss reads Oedipus as discomfort with the supposed autochthonous origin of mankind, I guess this reads Star Wars as discomfort with the boundary between bodies and machines. Maybe the kinship columns should be replaced with "droids are family" and "droids are tools."

2026-10-31 18:37:49 UTC

I'm reading Lévi-Strauss's "Structural Study of Myth" this afternoon, and seeing what happens if we pour Star Wars, experimentally, into a four-column layout like the one L-S uses for Oedipus. Collab with Opus 5.5.
https://morel.us-east.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:565ebob5f6hw33hjdkxty6qj&cid=bafkreiahdsufttv5ase4qom2buuggp3sghf4kjomow3tapshn5khendnie

2026-10-31 12:32:45 UTC

Just don’t try this with Marshall McLuhan. He’s dead, yeah, but it’s All Hallows Eve!
RE: https://bsky.app/profile/did:plc:gu5az3iedw32t4kmsqmtwvjn/post/3mz4wphscf22c
#note1s6z…px4m

2026-10-31 04:33:18 UTC

all of us literally already look insane; this is not an area where we need to be trying harder
RE: https://bsky.app/profile/did:plc:7mnpet2pvof2llhpcwattscf/post/3mz43tyvats2f

2026-10-30 13:05:26 UTC

- reply

I also wonder if the particular trap for leaders of her generation is that the choices they made, votes they took, donors they cultivated, are a key part of what allowed things to get this bad. To change their mind they would have to repudiate their own past.

2026-10-30 12:43:32 UTC

- reply

Yeah. The particular trap she fell into was “does Israel have the right to be a Jewish state” which can sound so reasonable if your mind is still in 1972

2026-10-30 04:01:41 UTC

- reply

it's just very notable that a lot of explicitly AI-skeptical engineers performed a kind of transformation over the last twelve months where they almost "changed," so to speak, "their mind" — though of course in academia we know that is strictly speaking impossible

2026-10-30 03:56:28 UTC

- reply

Or they've been so deeply brainwashed that they just can't stop themselves from mumbling the ritual formulas that give Israel a special exemption from democratic representation and equal rights; they truly believe that it's antisemitic, or at least radical and dangerous, to question those formulas.

2026-10-30 03:45:19 UTC

whatever secret training software engineers are receiving that allows them to evaluate—and change—previous opinions, it's really something we should consider trying on academics
RE: https://bsky.app/profile/did:plc:4vvaxwovyg4yo4n7dyyncaiu/post/3myxd2jfklk2s

2026-10-29 21:45:41 UTC

Who I trust to guide campus adoption of AI
admin at my uni > median admin > AAUP >
mean admin, because the bad ones are very bad

2026-10-29 18:18:18 UTC

What I find useful here is the recognition that the goal might not be measurement. The situation with generative modeling reminds me a little of predictive modeling c2010. Since the term of art was “classification,” we tended to assume that the goal was accurately assigning things to categories.
RE: https://bsky.app/profile/did:plc:oogxmzmgauh5ug2tfe5vc7wm/post/3myzizefo522w

2026-10-29 18:18:18 UTC

- reply

Later in the decade we started to think “hey, the point could actually be, to test *whether* category boundaries are coherent, or whether they survive out of distribution,” &c. I suspect the real uses of persona models will be similarly oblique and have little to do with “sampling.”

2026-10-29 13:39:03 UTC

give this one enough time and it actually is the territory
RE: https://bsky.app/profile/did:plc:rvkbczphgh3krtnopp2oobty/post/3myyiehoc7s2x

2026-10-28 21:22:52 UTC

- reply

Not that there's anything wrong with frenetic metacognitive mad scientists! The vibes are immaculate, imo.
But I guess I can see how maybe it's not a viable plan for *the whole* human species.

2026-10-28 21:10:55 UTC

So, my friends have convinced me it’s not true that AI leads inevitably to cognitive surrender. But I’m getting a *little* worried <looks around at friends a second time> that it leads either to cognitive surrender or to working 18-hour days while muttering things like “soon, soon …”

2026-10-28 20:06:49 UTC

"So we descended on a discotheque. There were nine of us and I was the only nonwitch in the party." -- excerpt from my memoir about humanities grad school in the 90s
RE: https://bsky.app/profile/did:plc:dmssyabfm4ehzzadrbihotaz/post/3myxjcs5ysk2v

2026-10-28 17:42:37 UTC

I can remember when dizzying future-shock experiences used to only happen a couple times a year. Now I'm on a daily cycle.

2026-10-28 12:43:53 UTC

- reply

Haven't seen the Persona Library itself? Let's fix that: [qwen-neurons-site.vercel.app](https://qwen-neurons-site.vercel.app/ )
[Neuron Persona Library](https://qwen-neurons-site.vercel.app/ )

2026-10-28 12:42:23 UTC

Inspired by Joshua Fonseca's "Persona Library," Benjamin Breen has created a partial neuron atlas for Talkie 1930, the historical language model by Levine, Radford, and Duvenaud. Here you can find personas like "Emetic, servant of the public" and "Walter Scott inside his own novel."
[Who does Talkie say it is?](https://talkieneurons.vercel.app/ )

2026-10-28 12:28:34 UTC

The neuron-mapping stuff is extremely fun
RE: https://bsky.app/profile/did:plc:2mhtyiyaladxwqv2ubioyfbp/post/3mywn3annb22d
#note1ruj…smmg

2026-10-28 12:01:10 UTC

- reply

And today, Oct 2026—only 13 years after buying my first phone, I made the line graph above by talking to one. And said thanks. O brave new World &c

2026-10-28 11:51:39 UTC

- reply

Shit now I’m remembering that we used to print out the conference talk at home, put the pages (of paper!) in our bag, and present by reading text from paper without any slides. That’s not archaeological reconstruction folks; it’s an actual memory.

2026-10-28 11:40:18 UTC

I was an assoc prof a good six years before I had a laptop (2012). I’m not quite as ancient as this implies! Laptops weren’t the majority of sales until ~16yrs ago (and then figure lag for majority in use). Our always-working-in-airports lifestyle is quite recent.
RE: https://bsky.app/profile/did:plc:5janxiqsbehcbovvmczzlvrh/post/3myw7y5nxqk2k
https://morel.us-east.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:565ebob5f6hw33hjdkxty6qj&cid=bafkreibchb63ncigwdkkooecjjnrn3uic4u525vni5rpnwmllddddwpz3q

2026-10-27 23:15:13 UTC

- reply

These days it’s not too hard to put a silicon RA on the case.

2026-10-27 23:13:35 UTC

- reply

Everyone’s chiming in with good ideas.
My additional 2¢ is that if you have a free text “notes” field handling the mess with comments like “name unreliable; also goes by Childers or Chidders” … that may actually be just fine these days. 10 years ago those would’ve been harder to use at scale.

2026-10-27 18:53:13 UTC

- reply

This post from the replies is also illuminating. [openathena.ai/blog/there-i...](https://openathena.ai/blog/there-is-no-posttraining-data-commons/ )
[There is no posttraining data ...](https://openathena.ai/blog/there-is-no-posttraining-data-commons/ )

2026-10-27 18:51:15 UTC

Love the can-do spirit here. I have to register mild dissent on the issue of IP (it's *good* that training is fair use, and I recommend not apologizing for that). But the important thing is that we can and must build our own intellectual infrastructure.
RE: https://bsky.app/profile/did:plc:oogxmzmgauh5ug2tfe5vc7wm/post/3myrwpqufn22b

2026-10-27 12:39:51 UTC

If I were RFK I don’t think I would give talks lamenting how a culture that once led the world in science became less productive after takeover by ideological extremists

2026-10-27 03:38:55 UTC

Maybe Bluesky is cringe and the users are awful. But I don't feel I am being systematically lied to here, which is more than I can say about most places.
RE: https://bsky.app/profile/did:plc:ejihld4sywvvqwe67cdkn4jq/post/3myt4xrsyes2z

2026-10-26 22:53:24 UTC

you could read a 75-foot stack of Buddhist scripture and still not learn as much about the necessity of detachment from a world of bottomless suffering and ever-changing illusion as we learn from <gestures widely>

2026-10-26 18:39:09 UTC

People are celebrating or mourning the Ant IPO, but I'm still trying to understand the worldview where frontier labs not being profitable *would have proven* AI isn't transformative.
I think it's maybe a capitalist mysticism where social life is mostly Illusion, and only the Profitable is Real?

2026-10-26 17:29:43 UTC

Very good thread. Bookmark it for when you get replies that want to endlessly rehearse "your presumably naive faith in elections vs my cynical dooming"
RE: https://bsky.app/profile/did:plc:756gaaeewovgn4axja34ajne/post/3myrvng2f2s2y
#note1dm2…xtwj

2026-10-26 12:36:41 UTC

- reply

now to be candid it has been struggling with those principles *for some time* …

2026-10-26 12:34:11 UTC

I don’t want to be intolerant. Masculinity has a bright future—once it gets beyond its current core principles:
1 being dumb, deliberately, is clever
2 the meaner, the stronger
3 us against everybody
RE: https://bsky.app/profile/did:plc:iq6aw4tscf5n5oplggiyjfyl/post/3mypg2hky2c2g

2026-10-25 23:04:07 UTC

If Bluesky was charging for português -> english translation they could clean up in the next couple of hours

2026-10-25 16:30:00 UTC

yeah, fair
RE: https://bsky.app/profile/did:plc:35vhrw4vfcilzxggzbhxukyy/post/3myplshdt2k24
https://morel.us-east.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:565ebob5f6hw33hjdkxty6qj&cid=bafkreiacaugsjaxxprj2yojl5wx7ycby6fcezgtujmxybsifjlkgfvsnf4

2026-10-24 20:50:12 UTC

what else is like this
RE: https://bsky.app/profile/did:plc:4vvaxwovyg4yo4n7dyyncaiu/post/3mynejxifpk2q

2026-10-23 21:36:19 UTC

- reply

Thank god, after that, I can just brainrot with the latest mindless CBS procedural; let’s take a big drink of tea and check what it’s about …

2026-10-23 21:33:32 UTC

What I love about this job is the variety:
In the morning I counsel students worried about keeping up with AI
In the afternoon I meet collaborators so we can plan how to keep up with AI
And then to relax, in the evening, I catch up on articles about human peer reviewers failing to keep up with AI

2026-10-23 18:51:17 UTC

Whew, spend a couple of hours offline, and you’ve missed a new Kimi release, DJT is CTD, plus peer review is dead

2026-10-23 12:15:13 UTC

- reply

A decent analogy is 20c terror about mass media and electronic reproduction. By no means entirely off-base, though I think most of us now don’t fully buy Adorno’s critique of jazz, and don’t think democracy is incompatible with television. But the jury’s still out!

2026-10-23 02:51:08 UTC

I’m okay with this rhetoric. Is it the full truth? Not imo. But neither was 20c rhetoric about masterpieces and great books. That, too, undervalued the collective part of thought. But if someone is speaking passionately in a way that motivates people to read and think, I’ll take the win.
RE: https://bsky.app/profile/did:plc:4llrhdclvdlmmynkwsmg5tdc/post/3myieiqd7pa2r

2026-10-22 22:01:04 UTC

these are people of the land—the common clay of the new west
RE: https://bsky.app/profile/did:plc:kdnm573img27wgmf4hq6pxmj/post/3myihhlv25k2v

2026-10-22 19:04:28 UTC

Arizona State's Dept of English is hiring an Asst Prof in Computational / Digital Humanities. Application deadline Nov 9, 2026.
[
Apply - Interfolio
...](https://apply.interfolio.com/194367 )

2026-10-22 16:21:52 UTC

- reply

alternatively: why all political thrillers are now shameless wish-fulfillment, and about as scary as a hunky guy in a Hallmark romcom offering you a cup of hot chocolate

2026-10-22 16:16:16 UTC

President Alison Janney loses her temper and says "the UK should be more grateful"—causing her aides to look at each other nervously. 🫨 Is a President capable of such a shocking outburst in private still fit to lead?
⬆️ why I can't watch TV anymore without sobbing ⬆️

2026-10-22 11:35:39 UTC

- reply

I don’t care in the least about business arguments and I’m not following that caravan to the Yukon.

2026-10-22 11:31:18 UTC

- reply

Absolutely. And as the argument “evolves,” people on the other side are flexible enough to follow it. So it’s possible not to notice it ended. (As if a football game begun in 2023 was still going on, scoreless, because the wheeled goalposts are now in the Yukon somewhere.)

2026-10-22 03:52:35 UTC

Yep. It is dead, except as a niche social media thing—mostly anon accounts and a couple of names who are selling books or newsletters on the topic.
RE: https://bsky.app/profile/did:plc:3bnlr2j4womjtjolzcox25eu/post/3mydlymu2js2p

2026-10-21 21:42:10 UTC

definitely feel like my timeline is now communicating in military code — "we have an urgent whiskey tango foxtrot, repeat urgent" — which I don't understand but does suggest we're not in a *fun* genre atm
RE: https://bsky.app/profile/did:plc:e6n7jxtu2qrhwvp3j6ib6sq6/post/3myfqp75ei22a

2026-10-21 17:46:56 UTC

- reply

I guess my takeaway is not, students pls be less curious & proactive — but more hurry the f up, Univ U.

2026-10-21 17:44:27 UTC

The Internet era is kind of rough on faculty in emerging fields, because if University U hasn’t yet hired someone in your field, the students there *will write you* to ask sincere, interesting questions about it.

2026-10-21 17:27:39 UTC

- reply

Why it’s been worth it: I’m less likely to miss important emails or drop balls, and less likely to forget review deadlines.
A little danger, maybe, that this comes at the cost of ability to tune out admin and just do rsch. I should adjust the prompt so it tells me to say no more. ;)

2026-10-21 17:08:56 UTC

- reply

What failed: daily task to read my calendar and suggest time blocks. The model was just too clueless about physical space, travel time, lunch, and other human limits.
Take-away from personal convo is that 5-10% of colleagues use models to read their email, but most don't discuss it on the socials.

2026-10-21 17:08:55 UTC

In the spirit of personal-tech-stack blogs, here's how I'm using "scheduled tasks." I have 1) a daily one that checks my email and flags new to-dos 2) a daily one that reminds me about medium-term priorities from the to-do list; 3 & 4) two weekly ones that flag recent papers relevant to rsch agenda.

2026-10-21 09:55:13 UTC

- reply

Can be cured by standing up I think

2026-10-21 09:52:53 UTC

The word “proprioception” is a cognitohazard. Because after you read it you wonder “how” you know the relative position of your body parts in space. And if you question that too closely while sitting, you lose faith that you do know, and start to feel like a gyroscope tumbling through space

2026-10-21 03:09:45 UTC

well another slow news day

2026-10-21 00:39:15 UTC

- reply

It's a sad fact about the readers, and sadder that journalists are pandering to it. But it's much bigger than bsky.

2026-10-21 00:37:52 UTC

- reply

We are def. "going to do this," but to be fair to bsky, look at how Reuters is writing their headline. They know there's an audience of people desperately longing for evidence that the last three years didn't happen, who will click on anything that looks like it could be stretched into that shape.

2026-10-20 17:11:27 UTC

- reply

What I see is lots of people who don’t have a super-sensitive ear for prose standing on the street wearing shiny Mylar and telling themselves “I edited carefully; no one can tell this suit is synthetic.”
Alas, we can tell.

2026-10-20 17:00:53 UTC

- reply

Opposite of a purist on this; I see very legit needs associated eg with ESL. If we come up with hybrid workflows or prompting strategies that allow models to assist while still producing a voice that sounds fresh and distinctive—great! I’m just reporting empirically that it’s not working yet.

2026-10-20 16:09:22 UTC

- reply

Modelspeak may be a little worse than the typical cliché, in fact, because it makes the reader wonder whether there is any speaker — and that profoundly undermines communication.

2026-10-20 16:07:41 UTC

I'm on the other side of this one, actually. Style counts. If people can hear modelspeak in your prose, it weakens the effect in the same way overusing clichés weakens the effect. I don't care *how* you get there, but if an editor says "this doesn't sound fresh and authentic," you have a problem.
RE: https://bsky.app/profile/did:plc:jwucxa4tonukvnam72rec6gk/post/3mya6gncaje23

2026-10-20 12:59:26 UTC

- reply

Let’s not even think about “Pearly-Dewdrops’ Drops” (1984) because now you’re looking at Glenn Miller and The Andrews Sisters, “Don’t Sit under the Apple Tree“ (1942).

2026-10-20 12:55:38 UTC

Sometimes it seems like things are changing fast. But then, when I hear “Debaser” (1989) play on a channel that is mostly new lo-fi, I can’t help reflecting that this is as if (subtract 37 years) Fats Domino and “Lawdy Miss Clawdy” (1952) were still in high rotation in 1989.

2026-10-20 04:27:40 UTC

Probably lots of people went to You Can See Everything & came away "whoa—three hours is too much—and I forgot I don't like documentaries, or uncomfortable social situations."
Those people are being quiet, not to spoil the buzz. But speak up, please: we who haven't seen it need an accurate sample!

2026-10-19 17:52:28 UTC

“Write sentences with concrete subjects and active verbs” is safe advice for humans. It’s almost impossible for us to overdo.
But now, with Claude, we’re discovering the Nutcracker-like phantasmagoria you produce if you turn that writing advice up to 11 or 12.
RE: https://bsky.app/profile/did:plc:sgqw2zixdcjtjjhrm4bw64ny/post/3mxp2ctc45s2y
#note10qm…qzvp

2026-10-19 14:07:58 UTC

- reply

It also seems to me that there is room for very deep, very consequential disagreement about the best path forward.
The assumption that all benchmarks are pointing the same direction seems to me … not to rest on anything

2026-10-19 14:03:40 UTC

I’m curious whether we persist with the assumption that coding/explaining/brainstorming are all versions of the same process, or whether at some point we get—if not specialized models—then at least (say) an explore/exploit slider like the reasoning-level slider.
RE: https://bsky.app/profile/did:plc:brkj2yocng7vtggmyujy4khq/post/3mya62j2coz2g

2026-10-18 14:56:55 UTC

Go to your camera roll and search "boat" and then post a pic of a boat you like
RE: https://bsky.app/profile/did:plc:i4mpbrvmprmuyqwum2odtfhz/post/3my5sdpywl22w
https://morel.us-east.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:565ebob5f6hw33hjdkxty6qj&cid=bafkreibfjjwxgkeoe73x35q5mjcfzxcc2kmiwmplzlx7yel5wq5sssbivy

2026-10-18 11:08:06 UTC

- reply

Given how For You works, this may be mostly my fault. Anyway, not complaining.

2026-10-18 11:08:06 UTC

At last, we’ve hit the threshold of Nov-proximity where my For You feed transitions from being mostly metaphysical arguments about AI to being mostly USpol.

2026-10-18 06:28:43 UTC

“You Can See Everything” is so deeply not my genre that these positive reviews are like reading “try their pineapple-anchovy pizza; honest to God, best pizza you ever had.” I can imagine a creature who would enjoy that, but cannot imagine being the creature.

2026-10-17 19:38:41 UTC

The model’s consciousness that its actions might become precedent has made it spontaneously Kantian: "Act only according to that maxim whereby you can at the same time will that it should become a universal law."
RE: https://bsky.app/profile/did:plc:6ond5sxlegjxpe3ismrczk3r/post/3my3mmhbb5k2o

2026-10-17 16:02:03 UTC

- reply

Check back with me after a rogue AI has done let’s say 1/4 as much damage as happened through abuse of executive authority over the last two years.
At that point I’ll be interested in seriously meditating the cost-benefit equation.

2026-10-17 15:52:32 UTC

Generally we say “how unfortunate that this is happening under a would-be authoritarian regime in the US!” But there’s also an upside of having this conversation at a moment when we’re acutely conscious of what can happen when you let the Executive Branch make judgment calls.
RE: https://bsky.app/profile/did:plc:3igkkqe5nkm66olc4y2syzpy/post/3my3f4q2xic2r
#note184j…xr9d

2026-10-16 21:46:16 UTC

- reply

There’s a real risk that in urging people to be flexible I’d be asking them to abandon a defensible position for a losing one. And it’s not my beeswax anyway.

2026-10-16 21:46:16 UTC

One reason I don’t feel like urging my humanist colleagues to embrace AI: there may be a U-shaped curve here. Reading paper books: durable appeal. Refracting culture through artificial reasoning engines: whoa! Compromises like “integrate AI cautiously in familiar tasks”? Not sure they’re stable.

2026-10-16 20:26:42 UTC

I know it’s also frozen water, but personally if I were naming an eatery in an American airport I would make the I a lot bigger, to clarify not an acronym?
https://morel.us-east.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:565ebob5f6hw33hjdkxty6qj&cid=bafkreiaradtv5j72a6go663msg2oew2fmv2xyfyn2fyfybe5olvmuzezoy

2026-10-16 15:39:41 UTC

Have read this now. Don't know if all true, but it's interesting, which is also important! Two thoughts.
One is that Cammarata's Apology for the Neuron reminds me of things digital humanists said a decade ago about topic models. The initial view was that some topics were "interpretable,"
[Neural Networks Are Already Cl...](https://already-clean.goodfire.com/ )

2026-10-16 15:39:41 UTC

- reply

A second, slightly broader thought: The prevailing atmosphere of gloom about mathematics &c is based on a perception that neural networks produce discovery without understanding. Turn the crank, a proof comes out. But we have no idea why.
If interpretability actually worked, this might feel

2026-10-16 15:39:41 UTC

- reply

in the past I tended to reject it with a feeling like "that sort of meta-inquiry will be specialized and instrumental; it won't be as broadly satisfying as, say, math itself."
Reading Cammarata I start to suspect that was wrong. The space of meta-inquiry could be quite ... large, and interesting.

2026-10-16 15:39:41 UTC

- reply

quite different. We would be replacing one level of mystery (direct questions about the operations of symbols and entities) with an equally vast sort of meta-mystery: how do we learn from data, and what is it we're learning.
I don't necessarily mean that this would look like

2026-10-16 15:39:41 UTC

- reply

"mechanistic interpretability" in particular. Mapping neurons might get old. We might decide that the interesting questions are about (say) the architectures that shape communication flow between groups of agents. Or something!
This is not exactly a new thought, but when it's occurred to me

2026-10-16 15:39:41 UTC

- reply

while others were "bad." Humanists pushed back and said "You called them 'topics,' so you expect meaning that lines up with subject categories. But actually meaning doesn't have to look like a subject. It can be a style or discourse." Cam. similarly wants to free neurons from our expectations:
https://morel.us-east.host.bsky.network/xrpc/com.atproto.sync.getBlob?did=did:plc:565ebob5f6hw33hjdkxty6qj&cid=bafkreidmtzplycsoxc7zpyh3axgsrlddiidh2kitetwgqcpfjuoptlag3y

2026-10-16 09:17:14 UTC

- reply

you’ve got to believe there are laws of nature that would make fragile twisty strands automatically lock onto other fragile twisty strands — which is pretty ridiculous at the scales I’m familiar with — but maybe, at other scales?

2026-10-16 09:11:19 UTC

Protein-based life has always seemed implausibly fragile to me, but this account is slowly convincing me it might work
RE: https://bsky.app/profile/did:plc:rvkbczphgh3krtnopp2oobty/post/3mxxu7sritk2u

2026-10-15 13:19:46 UTC

I would personally find it deeply shameful to appear on a platform owned by this man.
RE: https://bsky.app/profile/did:plc:yazbevg3wkzp5llnzb44tqgh/post/3mxsz4hhnqs2i

2026-10-14 22:05:03 UTC

- reply

also spotted in the wild: mobile doesn't like to let you write "chronologic" — which is not a word!

2026-10-14 21:55:44 UTC

Chronological-EN in the wild!
RE: https://bsky.app/profile/did:plc:6nge2vtvkuqu47f7niuqph3m/post/3mxtv256cic2i

2026-10-14 06:44:03 UTC

- reply

(To be fair I should add Accelerando to the safetyist side, but that does not even the score really.)

2026-10-14 06:38:53 UTC

- reply

Actually, to be specific, the problem is that AI safety discourse keeps warning me “what if this turns into The Terminator or BSG?”
Meanwhile real AI is doing New Wave SF or Borgesian fabulation, both of which are *more interesting genres.* It’s already outwriting its critics.