iiinova — where your digital self keeps evolving

david barsky @davidbarsky.com · 16/09/2026

i found the post. they’re thinking, there’s a chain of thought, but they’re *not sentient*

010

david barsky @davidbarsky.com · 15/09/2026

respectfully, at no point did will say they were sentient, i’m not sure where that entered the conversation.

141

david barsky @davidbarsky.com · 14/09/2026

bsky.app/profile/davi... i refer to this post

010

david barsky @davidbarsky.com · 14/09/2026

this guy is simply too much of a handsome jock be aware of worktrees or bundles, which, god i envy him

050

david barsky @davidbarsky.com · 14/09/2026

like, google and fb hit this 15+ years ago. agents just means that the market for “scalable source control” goes from 10 companies to “everyone in software”

030

david barsky @davidbarsky.com · 14/09/2026

you may not like how this guy is saying it, but this is fundamentally our bet. git is nice! but it’s designed for linus merging changes from some pals he met at a conference recently

370

david barsky @davidbarsky.com · 11/09/2026

where’s the white monster

010

david barsky @davidbarsky.com · 09/09/2026

no problem! i can’t comment on the theological/consciousness arguments in here, but it is correct that unfolding is possible (just wildly impractical). i also think the comparison to frozen corpses is evocative, and uh… not wrong 😬

010

david barsky @davidbarsky.com · 09/09/2026

…and training this thing starts to approach “heat death of the universe”-levels of absurd timelines

110

david barsky @davidbarsky.com · 09/09/2026

ah, no, not really, they can’t be reduced to a single layer. the interesting interactions are between layers (induction heads), and even if you could, the “width” of the layer would be exponential with respect to squashed layers. there’s no storage medium dense enough to store it…

110

david barsky @davidbarsky.com · 08/09/2026

but yes, dennett would not approve, but i genuinely wonder if he’d revisit his position if he were alive today.

110

david barsky @davidbarsky.com · 08/09/2026

i mean this in good faith: are you invoking a sorites paradox? cause an attention head reading across layers in LLM is how they’re able to *do* stuff; they’re called induction heads. a single layer would be a very expensive bigram/skip-trigram lookup

210

david barsky @davidbarsky.com · 08/09/2026

yeah, but there were like… 2 people other people qualified to do so, and they were working on astral.sh/blog/ty at a separate company (you’ll see my name on that blog posts. we shared the same underlying library and worked together on it for the better part of a year.)

010

david barsky @davidbarsky.com · 08/09/2026

with ai. i did an equivalent task (sans ai) 2 years ago that took roughly 6 months, end to end.

100

david barsky @davidbarsky.com · 08/09/2026

i think there’s a third option, which is “we stumbled into engineering intelligent p-zombies with lots of matmul whose unreasonable effectiveness in software is a function of robust RL environments, not of any innate capability”

110

david barsky @davidbarsky.com · 08/09/2026

i don’t think it does. these tasks would have not been feasible before this year.

100

david barsky @davidbarsky.com · 08/09/2026

it does not actually help. why do you think this is negligence?

000

david barsky @davidbarsky.com · 08/09/2026

it was the case for some proofs, but it was not the case for erdos problem 90 or the unit distance problem done via astra.

020

david barsky @davidbarsky.com · 08/09/2026

let’s say i have a TBI and forgot those classes. what would i learn that would make this leap a bad idea

100

david barsky @davidbarsky.com · 08/09/2026

how about 2 papers from MIT (arxiv.org/abs/2310.02207, proceedings.mlr.press/v235/jin24e....), or one from harvard, MIT, and northeastern (arxiv.org/abs/2210.13382)? same findings. “probability modeling”: of what? the last paper showed the model learned board state from moves alone

010

david barsky @davidbarsky.com · 08/09/2026

the other thing: portia spiders! they’re incredibly intelligent, can plan an indirect path to prey, and only have 100k neurons. i doubt they have qualia, but we *know* they’re not performing mimicry. if portia spiders are doing reasoning, then i don’t know how LLMs aren’t

230

david barsky @davidbarsky.com · 08/09/2026

anthropic put out a paper last year showing how haiku answers “capital of the state containing dallas”. it activates a “texas” representation, then austin. when they replaced “texas” with “california” mid-computation to the same question, it output “sacramento”. that’s casual inference and meaning!

210

david barsky @davidbarsky.com · 08/09/2026

i believe this is what paul is referring to when he talks about “reasoning”, not that they’re literally conscious

110

david barsky @davidbarsky.com · 08/09/2026

please note that when i describe LLMs as intelligent, i am not making any phenomenological claims. i do not think claude has qualia. however, llms clearly has an internal representation of concepts that it can manipulate to arrive at previously unstated conclusions

111

david barsky @davidbarsky.com · 08/09/2026

well, a few things: 1. the new models are roughly as good as i am at writing rust. 15 months ago, they were unusable. 2. idk how else to describe giving a model an ambiguous goal in natural language and have it accomplish it, entirely unaided, hours later. they’re clearly intelligent!

110

david barsky @davidbarsky.com · 08/09/2026

idk what this is, but’s certainly no calculator printing “boobies”, that’s for sure

3263

david barsky @davidbarsky.com · 08/09/2026

wild to see the replies to this post. i just had sol-5.6 write a fuzzer for incremental compilation in rustc almost autonomously. one of my cofounders had claude spend the night profiling/benchmarking our storage system and came back to a stack of changes in the following morning

7642

david barsky @davidbarsky.com · 01/09/2026

nick noble drove this new site. he collaborated with stephen nixon of www.arrowtype.com. we have a bunch of other stuff coming and we can't wait to show it off!

130

david barsky @davidbarsky.com · 01/09/2026

we're ridiculously excited to have martin, and we're *also* super excited to unveil our new site!

1100

Reposted by david barsky

martinvonz.bsky.social @martinvonz.bsky.social · 01/09/2026

I left Google and joined East River Source Control almost two months ago. Our redesigned website is now up, including a blog post about me joining: ersc.io/blog/martin-....

ersc.io

East River Source Control Names Jujutsu Creator Martin von Zweigbergk Chief Technology Officer // ERSC

After 15 years at Google, the creator of JJ joins ERSC as CTO.

315626

david barsky @davidbarsky.com · 24/08/2026

hmm, you’re right. i think the actual statement needs to be weaker: this works if you shard your agents into standalone/separate work. cross-shard transactions/edits aren’t really supported :)

120

david barsky @davidbarsky.com · 24/08/2026

yup! i'll also author a blog post about this, linking to isaac's post!

120

david barsky @davidbarsky.com · 24/08/2026

annoyed by large build directories or cross-worktree artifact sprawl? a #jj-vcs megamerge is a great place to run your builds and tests, because because if the megamerge builds, so do all of your worktrees! plus, your build directories are O(1) with respect to worktrees, not O(worktrees)!

2221

david barsky @davidbarsky.com · 21/08/2026

amusingly (for some definition of amusing…), no government is *really* pushing for this to be addressed! the wording in the EU’s laws basically allows enforcement to not look at code forges

170

david barsky @davidbarsky.com · 20/08/2026

for observability: so much `tracing` in honeycomb. locally: lol good question; everything seems to suck

160

david barsky @davidbarsky.com · 20/08/2026

steve replied partially (bsky.app/profile/stev...), but to answer questions: 1. mixed auth: would be neat! not supported now. i’ll make a ticket. 2. capybara: some ideas, but it’s early days. 3. glean: we want some codenav, but didn’t think about to integrate arbitrary systems

150

david barsky @davidbarsky.com · 19/08/2026

in general, would a relatively short blog post about how we have fine-grained, zanzibar-style permissions, no-long lived secrets, and first-class service account/workload identity (i forget how people refer to each…) be of interest? (in addition to the standard audit logs, ofc)

3130

david barsky @davidbarsky.com · 19/08/2026

god if we’re the blonde of source control/forges, that’d be a treat and an honor

140

david barsky @davidbarsky.com · 19/08/2026

we have no idea what software is going to look like a year from now other than “there will be far more code than anyone expects”. we’re building for *that* world.

130

david barsky @davidbarsky.com · 19/08/2026

as a spoiler: we integrate with your existing CI; we don’t offer one at moment. current integration is buildkite, but it’s very easy for us to add one.

070

Reposted by david barsky

Steve Klabnik @steveklabnik.com · 19/08/2026

Gonna start actually posting to the @ersc.io blog for real soon What do you want to hear about?

281147

david barsky @davidbarsky.com · 18/08/2026

democratic centralist web apps

070

david barsky @davidbarsky.com · 18/08/2026

buy it now at $110; my 13-old year iron rangers are worth their weight in gold

020

david barsky @davidbarsky.com · 18/08/2026

i mean, i’m *also* less confused, but… it feels like there are less moving parts with the former?

010

david barsky @davidbarsky.com · 18/08/2026

i think there’s a lot of stuff that matters less, but monorepos vs. polyrepos aren’t one of them, imo. i feel like my claudes/codexes are far less confused when they work in a monorepo than in a polyrepo setup since there’s just one canonical history they can see

110