Aman Choudhri

[Aman Choudhri @ChoudhriAman]

@Gabogonzalez515 Bullish bullion

[Akshay @akshay_manglik]

Introducing Pomerado (YC F26): the faster, cheaper, and more reliable alternative to browser use.

Open source. Sign up for free to supercharge your agents 🧵

[Aman Choudhri @ChoudhriAman]

The mayors of builder city

[Akshay @akshay_manglik]

excited to announce @pomeradoai is part of the @ycombinator fall 2026 batch!

we help agents act on any website by creating a custom API in minutes

more from me and @Alanmrsa1 soon 👀

[Christopher Potts @ChrisGPotts]

Why is it learning to avoid whitespaces? I thought it might derive from pressure to keep reasoning short, but @yoavgo pointed out that dropping whitespace doesn't necessarily save tokens. Is it instead learning to rely on infrequent token combinations to create a CoT language?

[Aman Choudhri @ChoudhriAman]

@scaling01 Curious why certain words are "glued" together... would `extraneousschemafields` be fewer tokens than each word separately

[Aman Choudhri @ChoudhriAman]

Bye bye browser use

[Akshay @akshay_manglik]

Pomerado's live on ProductHunt! We turn websites into APIs on the fly so your agents can connect to consumer websites or legacy portals. It turned my Codex into a muse-level personal assistant for free.

Would love it if you could support!

https://t.co/ct6LVoYjTZ

[Aman Choudhri @ChoudhriAman]

they’re charging 5.70 for a dollar slice now

[Aman Choudhri @ChoudhriAman]

Opus 5.5 writes like it just read "The Elements of Style"

[Aman Choudhri @ChoudhriAman]

I put Jev into an earring and let it tell me what to do, here’s what I found out.

[Aman Choudhri @ChoudhriAman]

How does looping work with MoEs, are tokens clamped to route to the same experts?

[Exa @ExaAILabs]

We built a time machine for the web.

Introducing Exa Snapshot: an index of 400 billion historical snapshots of webpages that lets you search as if it's the past.

Snapshot is already being used for backtesting prediction models, RL at labs, exploring the pre-AI web, and more.

[Aman Choudhri @ChoudhriAman]

Should have a command to write a medium-form essay calling on the government to stop you

[Indie Game Joe @IndieGameJoe]

You're a missile that can't slow down and the gaps keep getting smaller

- Steer it to the target as fast as you can
- Fly closer to the walls for more points
- Grapple hook to swing round corners

Game is Dumbfire

[Aman Choudhri @ChoudhriAman]

They used AI to write the anti-AI doc

"The whole document in 5 words."

[Sterling Crispin 🕊️ @sterlingcrispin]

Wow this is absolutely going to fry people especially at Anthropic, this is like the Anti-Claude Constitution.

They just declared war on everyone researching model welfare.

“We’re building a tool, not a being. Fuck your feelings.” - Microsoft

[Aman Choudhri @ChoudhriAman]

“I can think of cases where if the model is really super intelligent, it thinks in a way so that it affects its own cognition.” - Dario, August 2023

[Ryan Fedasiuk @RyanFedasiuk]

4/ Astra is also much better at controlling what shows up in its chain of thought.

In one test, it was told to think about anything else. Its visible reasoning described sunlight, a desk and a ceramic mug. It still got the actual question right.

[Aman Choudhri @ChoudhriAman]

@DKokotajlo How much of this do you think is downstream of the issue that most of our eval environments come from the same “data generating process” as our training environments? There may always be capabilities evals cannot capture but it seems we’re blinding ourselves.

[Aman Choudhri @ChoudhriAman]

Hate to say it, the retro branding on the astra launch video is pretty cool

[Aman Choudhri @ChoudhriAman]

@joannejang @esiamid Fuzzily remembering results about the geometry of random matrices with diff distributions here

[Aman Choudhri @ChoudhriAman]

@joannejang @esiamid What is the "random" initialization distribution used here?

[Aman Choudhri @ChoudhriAman]

One of my favorite research pushes we’ve done recently. Work spanning inference, data, evals, architecture, and more

[Exa @ExaAILabs]

We’re releasing a research preview of “dynamic highlights”:

Given a query and set of pages, a custom model jointly scans the whole set to filter out redundant and irrelevant info, reducing token usage 95% on average.

Available today: https://t.co/O5gymky6i2

[Aman Choudhri @ChoudhriAman]

How do we study the “values” of agents? And can we trace it back to the pretraining data? How much will narratives about human-agent interactions included in pre/midtrain shape the future of our world?