David Khachaturov (@davidobot_) header
David Khachaturov (@davidobot_) profile picture

founder at @EgoistMachines (YC S26) bye-fellow at catz @Cambridge_Uni also dr. of cs

San Francisco
Joined October 2012
Photo by David Khachaturov (@davidobot_)Photo by David Khachaturov (@davidobot_)Photo by David Khachaturov (@davidobot_)Photo by David Khachaturov (@davidobot_)Photo by David Khachaturov (@davidobot_)Photo by David Khachaturov (@davidobot_)Photo by David Khachaturov (@davidobot_)Photo by David Khachaturov (@davidobot_)Photo by David Khachaturov (@davidobot_)Photo by David Khachaturov (@davidobot_)
i don't know and i'm scared
who's this? wrong answers only
155
or if you want it without incoherent but human-passing text, use ink.ego.ist
Everyone’s talking about how reliably Pangram can detect AI-generated text. But you can easily get Pangram to label a 100% AI-written text as 100% human. Here’s how:
1
1
191
we're edging ai interp understandability
Train the same AI 3 times. Same behavior, different circuits. So is a circuit the algorithm—or just one implementation of it? Meet ACE: a new way to understand AI. It extracts what different wirings share—to explain why AIs grok and to steer LLMs up to 32B 🧵 #NeurIPS2026
2
249
17hrs later we're at 80 :o
we're at 70 customers (excluding enterprise) just two weeks after launch, come see what all the fuss is about
3
14
540
what i meant to reply is
Replying to @erinmerylstudy
tried and it works. How are you doing this? Fine tuned model or some different technique?
1
220
wait that's crazy, who runs this account
we're at 70 customers (excluding enterprise) just two weeks after launch, come see what all the fuss is about
2
1
11
505
we're so close to solving taste-as-a-service it's insane
introducing academic mode for ink say things in the highest-signal-to-noise ratio possible
2
1
10
556
everything is going to be ok
Excellent essay, hard to summarize. Humans are going to be fine without work. Work was just a good way to build reputation and friendship, we will find lots of other ways.
2
269
everything is pointing towards an inevitable sharded internet - the current setup is very reminiscent of peak globalism circa 2016 what this will look like, we can only speculate; networks disconnected from one another across country borders? or perhaps we'll pull through
I don't "get" AI alignment. We very clearly are not aligned ourselves and don't even know how our internal stuff works. All behavior is a very general proxy for something going on inside. All cultures align towards different things. It seems like a very naive field, but I know close to nothing about its technicalities. I guess researchers are reaching for a very minimal "just don't kill us please" state of the models.
2
5
369
"Ut quod ali cibus est aliis fuat acre venenum." > What is food to one, is to others bitter poison. De Rerum Natura, Book IV, line 637
44
we, as people, are fundamentally constrained by creativity, drive, and novel ideas people in europe and uk are the same as people in sf; the latter just run on the spot a lot more to look busy and raise money faster bc that's the culture and there is more money i would argue that there have been as many, if not more, society-changing companies out of europe than out of the us in the last 20 years
I have this weird feeling that everything we're doing in tech right now might just be... busy work. It hit me recently with the slew of vibecoded Rust GUI frameworks. I proselytized this for years. Gave talks. Etched the code by hand, for years. Carefully reviewed hundreds of thousands of lines of intentional, careful, Rust. Hotpatching, WASM bundle splitting, layout, paint, input, state management, the whole deal. Now you just say "make it faster, use arenas, think about re-renders, maybe we can use templates?" and get the same thing in a few hours. We are not approaching the singularity. We are in the singularity. Every day slips by faster and faster. My productivity up and up. I'm sure if I truly took my hands off the wheel, I could be 10x more productive, or more. I'm not upset about it. A little bittersweet perhaps. If you're a builder, this shit is like candy. Anything you dreamed, any idea you had, in your hands instantly. But it also takes the fun out of it. Why build new side projects or embark on new ideas if it's one prompt away for anyone? There's no shared foundation anymore. Just forks and vibecoded personal software. It's fun to be part of the singularity, I guess.
1
8
747
a scene from Miyazaki's The Wind Rises comes to mind if Japan post-WWII is anything to go by, we're in for a wild ride
The coexistence of RSI with a fundamentally bottlenecky world will give rise to a late 2020s that feels like this: the rich strangeness of hyperfuture living alongside large swaths of unchanged or even regressing social practice, a temporal redshift all over your perception
351
someone leaked the secret sauce behind our sota model ink.ego.ist
LLM's weren't great at writing code either until they were put in a harness. Software Engineers don't know how to build a writing harness because they aren't good writers. Writers don't know how to build a writing harness because they aren't good software engineers. Writing isn't magic. It's going to be reducible to a collection of rules and loops like code. With the added need to reflect the zeitgeist, rooted in an opinionated point of view. This is beyond feasible.
1
8
523
what's better - British or US-style ads?
Introducing AI Passport, the (free) simple and secure way of keeping your private information, private. get yours: ego.ist
2
6
436
delaware is about to allow ai-run companies but are we prepared for even more slop automated outreach
SITUATION DETECTED: Delaware is drafting a proposal for a new form of corporate entity called an Artificial Intelligence Company (AIC). An AI managing an AIC would be able to provide services, sign contracts, hold property, and even file or defend lawsuits in the company’s name.
1
2
352
natural-sounding writing is the whole moat imo now as a chatgpt plugin so your agent writes beautifully chatgpt.com/plugins/ink
197
deslop without losing the plot models are very good at maths and not very good at sounding human - we just fixed that
4
2
4
31
35,977
introducing ink - ink.ego.ist it's early days and we're not done with training yet but it's already pretty good - the produced text doesn't get flagged by AI detectors
2
6
661
would love some feedback on the model and style - if you’re unconvinced or want more words to test with, dm me
2
3
358