Arjun Narayan

experimental theologist @AmplifyPartners.

[Arjun Narayan @narayanarjun]

@davidcrawshaw quicksandbox

[Dan @pizzaboy]

PewDiePie just shipped his free AI Workspace product 12 minutes ago btw

[Arjun Narayan @narayanarjun]

@teej_m People won't say this out loud but all the Indian food in Britain is extremely mid, you get better Indian food in any random town in the south bay.

[Arjun Narayan @narayanarjun]

@AvgDatabaseCEO Spencer Kimball and its not even close. Just look at the CockroachDB early commit history.

[Arjun Narayan @narayanarjun]

1. IP structure innovation so you could build a Bell Labs style research lab that can actually internalize the gains from basic research, rather than relying on a benevolent corporate or academic patron. the goal should be to securitize and thus pull in 10x-100x the current dollar spend on basic research.
2. build for profit mass transit (internalizing the connectivity gains through real estate options/purchases a la japan)
3. katerra that actually works/fix the 50 year construction productivity flatline
4. alpha school style college that focuses on 10xing outcomes at undergraduate level education

[Ben @Financialtailor]

I just checked and there are literally no historic examples of memory stocks crashing due to oversupply.

[Arjun Narayan @narayanarjun]

SpaceX drops the hypiest IPO filing and says "we believe we have identified the largest TAM in human history", and you're like, my interplanetary east India company, what delectable spice have you decided to ship across the stars, and it's like... 22 trillion dollars of b2b saas

[Arjun Narayan @narayanarjun]

@matt_slotnick This is a good summary
https://t.co/t6IfA22Qqn

[Lenny Pruss @lennypruss]

We’re watching Git melt down in front of our eyes as agents write most of our code.

This is why many people, present company especially, are excited about Jujutsu (jj).

Some observations on Git’s history, why it’s creating problems for an agent-first SDLC, and how jj may help:

A brief history of version control systems:

In the early days of computing, version control didn't exist because it didn't need to. Programs were small enough that a single developer could hold one in their head. As teams grew, early systems like RCS tried to solve coordination with file locks that serialized work one person at a time, but locks created as many problems as they solved.

Centralized systems like SVN and Perforce reframed the question entirely, reconciling simultaneous work through merging rather than preventing it. Then Git arrived in 2005, built by Linus Torvalds in a matter of days and described in its first commit as "the information manager from hell."

What it added was decentralization: every developer got a full local copy of the repo, which made branching trivial and gave the whole system a resilience no centralized tool could match. Even so, Git didn't win on technical merit alone; GitHub understood that version control systems are fundamentally collaboration platforms, and that insight, more than anything else, is what made Git the standard.

Why Git is suboptimal for an agent-first world

Git's well documented (and memed) problems fall into two categories: data structures and UX.

At scale, operations like finding merge bases require walking the entire commit history, and Git's reconciliation algorithms max out at roughly 2-3 merges to main per second. That’s fine for startups, but a hard ceiling for hyperscalers.

The UX problems are more universally felt. Git exposes its internal state model to the user, forcing developers to reason about whether a file is untracked, staged, committed, or stashed, when most people just want their work recorded. Merge conflicts are incredibly frustrating because Git treats them as invalid states that must be resolved before you can continue, with no way to set one aside and come back later.

The maintainers know all this, but Git is load-bearing infrastructure with 20 years of backwards-compatibility requirements, and the core team is severely constrained in what they can change.

Agents make all of it exponentially worse, amplifying Git's limitations by generating code at high velocity, turning hyperscaler edge cases into mid-sized-team problems, and the same state machine complexity that frustrates developers becomes reasoning overhead for agents.

Jujutsu (jj) and why people are excited about it

jj started in 2019 as an experiment towards a next-generation VCS by Martin von Zweigbergk at Google. The core difference is that jj eliminates the staging area entirely; every change is always committed, just like Dropbox, and all the little Git incantations are abstracted away.

Several design choices follow naturally. Undo is a primitive: jj maintains an operation log of every action you've taken, so undoing a rebase is one command, whereas Git users have all lost data hunting through the reflog. Conflicts are first-class, represented in the underlying data structures rather than treated as invalid states, which means you can have a conflicted commit, move it around, and come back later without interrupting your work.

And jj builds its own data structures on top of a storage backend, which today is Git, so you could start using it at work without anyone knowing.

For those nodding along, check out @erscio. ERSC was started by JJ’s core committers and they’re building the version control system and forge that will take software development into the next 20 years, for humans and agents.

[Arjun Narayan @narayanarjun]

@ericneyman You got me. But then again, everything is also in fact The Word...

[Arjun Narayan @narayanarjun]

What did you think "Amo Dei" meant? Vibes? Essays?

[Ilya Sukhar @ilyasu]

The solution to all of this is to have kids.

[Arjun Narayan @narayanarjun]

Cerebras popped 90% today. Is there anyone at Benchmark with opinions about IPO pricing? Curious to hear all perspectives!

[Arjun Narayan @narayanarjun]

Pray Mr. Suskever, if I ask the agent the wrong question, will it still give me the right answer?

[Arjun Narayan @narayanarjun]

Funnily enough you're being too charitable to Rawls because he's even worse! Rawls' answer to the question of economic growth is that it's *bad* and we shouldn't do it because intertemporally, investment is redistribution from the poor (the past) to the rich (the future), so not only should everyone be equally poor, they should be equally poor *forever*!

Can't believe there are people who take Rawls seriously.

[Arjun Narayan @narayanarjun]

@tszzl why do the models seem to need Jesus
https://t.co/t3sBfCCEas

[Arjun Narayan @narayanarjun]

@tszzl @lulumeservey insane to launch a board coup without first reading "coup d'etat: a practical handbook" and just following its precepts

[Arjun Narayan @narayanarjun]

@richardartoul "if the breakthrough was legit they should show benchmarks that are 100m context not 12m"

[Arjun Narayan @narayanarjun]

@richardartoul yes. i agree with you.

i agree i wouldn't want to share the details, but they should be able to demo that its meaningfully larger by 10x-100x on context window, and they haven't done that.

[Arjun Narayan @narayanarjun]

the people who turn off fsync dont advertise it as such. if i told you i built a new s3-only database with 1ms write latencies, you'd be mad and immediately know what shenanigans were going on. my main pushback is youre mad that theyre mad. but ai researchers are rightly annoyed by this attention grabbing without details.

thats why youre seeing what youre seeing from ai researchers. sparse attention is THE thing everyones chased for years. all the obvious things have been tried. they all have pitfalls. there are no details provided here.

besides, if you could do subquadratic attention, youd demo a 100m or 1b context window at known long context benchmarks and silence all doubters.