rohan

mts @DecagonAI

[rohan @rohan__sanjay]

@kristaletz @mntruell 👀

[rohan @rohan__sanjay]

@kristaletz @tibor_tee @SpaceX @SpaceXAI Congrats krista! amazing working together 🤝

[Cyrus @cyrusasg]

multimodal inference doesn't fit the mold of standard text autoregressive generation. text-to-speech models feature different architectures, different states, and different batch shapes. we rebuilt our tts serving around that and simultaneously reduced our time to first audio while improving throughput by several fold.

[rohan @rohan__sanjay]

@kristaletz @cursor_ai cloud agents are the future

[rohan @rohan__sanjay]

@poteto worktrees are dead

[rohan @rohan__sanjay]

@poteto coming soon 🫡

[rohan @rohan__sanjay]

our cloud agent usage at decagon is up 6x over the past 3 months. cloud agents are quickly becoming the default way our team works because:
- they scale beyond a single laptop: each agent gets its own isolated sandbox (no worktrees) so there’s no local resource contention
- you can spawn swarms from anywhere (web, slack, linear): designers and PMs are shipping
- they can run for hours, spin up all services, and put up PRs with screenshots/videos

but if verification isn't trustworthy, you just multiply slop. the bottleneck isn’t model intelligence, it’s the environment setup and verification. to verify effectively, you need to:
- give it the same tools an engineer has locally (same stack, MCPs, context)
- create a framework (like @poteto's pstack) that encodes how the agent should create/test its changes and ship artifacts you can review without reconstructing the session

the horizon of work you can give agents is much longer - cloud agents can pretty much ship while you sleep. probably the highest leverage thing an eng team can invest in right now.

[Cursor @cursor_ai]

We're continuing to improve cloud agents in Cursor.

They pick up work from events, hold a goal until it's met, and stay on course through long sessions.

[Cynthia Chen @chenxcynthia]

Two years ago, I joined a 12 person startup as my first job out of college.

A lot has changed since then. Here are a few things I learned along the way. https://t.co/ZF7ynR0jkF

[SpaceX @SpaceX]

After approx. 24 days at sea, the SpaceX Recovery team successfully guided Starship to a location just off the coast of Christmas Island. A team of SpaceX engineers is on their way to conduct additional analysis on the vehicle in calmer waters before attempting to return it to Starbase

[rohan @rohan__sanjay]

@cyrusasg don't break cache

[Cyrus @cyrusasg]

tons of hype but it feels like people often conflate "routers", what their purpose should be and where they should reside.

1 - is the router picking between different models, or between different providers of the same model
2- is it embedded in a harness, or fully externalized

openrouter is externalized, but by default just routes between providers of the same model, so it's really more of a gateway. this is fine, but the value add seems pretty narrow. it's not that hard to build your own ai gateway and in an enterprise setting you probably should.

ramp's router actually seeks to change the underlying model, while being external to your harness. this makes even less sense tbh. a model swap mid-trajectory is a full prefill so feels rare that'd actually be worth it. i could only see this for super basic one shot questions, not for any serious use case.

cursor's "routing" makes more sense since it's actually a harness level feature. embedded means 1 - you can easily be cache aware, 2 - you have the actual signal (difficulty, prior failure, tool results), 3 - model-harness co-optimization is a clear trend, so selection belongs where the prompts and formats already live.

although if you're already at the harness level, subagents feel like a better abstraction. a subagent boundary is a fresh prefix anyway, so switching models doesn't evict any cache, and the delegation decision already carries the task scope so you can avoid having a separate classifier bolted on.

overall routing feels pretty overhyped. id rather decompose into an ai-gateway for redundancy and stability, and a harness for efficiency and capability.

[Decagon @DecagonAI]

Modern TTS models can sound great — and still fail badly on pacing, pauses, and prosody.

We adapted DPO + GRPO to flow-matching models to tackle the tail end of TTS behavior: https://t.co/PnE1apjxg6

[DHH @dhh]

For anyone with endless ideas, this agent age is nirvana as those ideas are met with endless execution, endless exploration. I've never had has much fun working with computers as I do right now. What a time to be alive.

[Noam Brown @polynoamial]

An internal version of Astra, @OpenAI’s next major model family, solved 10 major open problems in mathematics, quantum complexity, and theoretical computer science.

We believe it will be a major step for scientific reasoning. https://t.co/iP6cyheZ7i

[Kimi.ai @Kimi_Moonshot]

Releasing the model weights and technical report of Kimi K3.

Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window.

New model architecture: 2.5x the intelligence per unit of compute, not just more params.

Alongside Kimi K3, we're opening up more of the stack behind it — high-performance attention kernels, MoE communication library, and infrastructure for running agent environments at scale.

Model weights: https://t.co/7m7eEg6Y0B
Tech report: https://t.co/yeu6cjpMCT
Tech blog: https://t.co/YTfiMSNM1f

[Thinking Machines @thinkymachines]

Today, we are introducing Inkling.

Inkling reasons efficiently across text, image, and audio modalities. We are making the full weights available.

https://t.co/Ghebq5mG30

Available today for fine-tuning on Tinker. Play with it in the Inkling Playground. 🧵

[Midjourney @midjourney]

Announcing a new division of Midjourney called "Midjourney Medical"

[Tren Griffin @trengriffin]

Coatue's Thomas Laffont on a "Power Law Paradox": a business valued between $100B and $1T (a "Centacorn") has a higher statistical likelihood (31%) of multiplying its value by 10x compared to smaller, earlier-stage unicorns (8%).

[rohan @rohan__sanjay]

@kapursanat The cex process makes you pay for trust you can’t really enforce - bringing this back onchain lets teams define their own incentive structure