Adam Hevenor

@signulll The routes are the crux according to his book.

@signulll Branson tried.

some rich billionaire, can you please create an airline that will destroy every other airline? - charge fair straight forward premium prices - optimize for comfort, food, & premium experiences. - fuck the rewards, credit cards, & points. just clean beautiful experiences. maybe acquire few airlines to do it for gate access.

A WAL on s3 is the modern db foundation.

@thdxr If you plan to do this recommend go over js

we've been experimenting with getting rid of the bash tool agents can write js fine which can do what bash can (though some gaps with things like git) and is more cross platform and then could run that in this

@danveloper This is really cool 😎

@thdxr Apple has been playing this game for years

i wish there was more information about this like what model they started with feels weird that all of a sudden the tools we're relying on are super opaque hasn't been like this for decades

This track rules. Open the door! youtu.be/LyhINTNK3VE?si…

@AprilNEA Internal go SDK!? C'mon!

🧵 I just reverse-engineered the binaries inside Claude Code's Firecracker MicroVM and found something wild: Anthropic is building their own PaaS platform called "Antspace" (Ants + Space). It's a full deployment pipeline — hidden in plain sight inside the environment-runner binary. Here's what I found 👇

@doodlestein @clanker_ Models love to build an autoscaling Kubernetes cluster if that's what you need. Even if you don't.

@clanker_ It’s because the average user is having the models do incredibly basic, formulaic stuff. But I agree with you that, for anything nontrivial, you should only ever use the best frontier models with reasoning effort cranked to the maximum.

there's this bizarre influenza psyop going around, endorsed by the frontier labs, that you should be using the mini version of a model, medium reasoning, & whatever other bs, when literally the only usable combo has only ever been the best model with the highest reasoning effort

Knowledge work actually. I know an epidemiologist balancing multiple max subs.

@mattpocockuk I mostly agree here, but I have also find it opens up some new use cases that were previously too chatty. E.g. building ,deploying and tailing logs. I know I could do things to make both ends of this more token efficient but nice not to have to.

Doing some experiments today with Opus 4.6's 1M context window. Trying to push coding sessions deep into what I would consider the 'dumb zone' of SOTA models: >100K tokens. The drop-off in quality is really noticeable. Dumber decisions, worse code, worse instruction-following. Don't treat 1M context window any differently. It's still 100K of smart, and 900K of dumb.

@signulll The fact they can't get traction in an app store is all you need to know

let's simulate what sama & mark chen are going through for the next few weeks on what to prioritize & what to chop shall we? this will be the walmart version of it lmao. openai has 29 "products" in the portfolio. - chatgpt - chatgpt search - deep research - chatgpt agent - chatgpt atlas - chatgpt pulse - sora - codex - image generation / images - voice in chatgpt - projects - canvas - tasks - shopping research / commerce in chatgpt - gpt store / gpts / apps - chatgpt business - chatgpt enterprise / team / edu packaging - atlas for enterprise - oit's vepenai api platform - responses api - realtime api - agents / agentkit - agents sdk - web search api/tooling - image generation api - codex cloud / developer codex surfaces - computer-using agent (platform/model layer) - operator (legacy/transitioning standalone identity) - let's say 4 hardware devices jony ive's team is working on (this probably does not get killed) now obviously most of these fit into the enterprise & coding territory. so what do you keep vs kill? most likely to be killed / deprioritized: atlas pulse gpt store (at least put in sunset) shopping & commerce sora (as a “hero product”) merge & consolidate everything else into chatgpt app. intense focus: codex agents enterprise workflows

The cost of not tinkering and exploring has never been higher.

@GergelyOrosz @turbopuffer - 10x cheaper (at scale) search. game changer.

What is an innovative startup you know of - and what makes them interesting?

Extract insights from Reddit using hiveminer. My agentic search harness hevmind.com/2026/03/01/har…

Production logs is the bible. Traces are the stories of what happened. How we interpret is what the Prophet told us.

@bcherny Ya'll should call this "happy hours"

We doubled Claude usage on weekends, and outside 5–11am PT on weekdays for the next 2 weeks.

@michalbrojak @trq212 @bcherny This is bonkers

What??!! 1M tokens context window?? Insane. Thank you Claude ❤️ But, when to clear it? @trq212 or @bcherny any tips?

@petekp Welcome

alias c="claude --allow-dangerously-skip-permissions" they see me rollin'... 🎶

@dexhorthy @trq212 I have already been confused by is default but requires opt-in fwiw

many folks are asking me what I think of Opus 1m context window - i am working hard with @trq212 and team to get this working so we can eval it but hit a couple speed bumps stay tuned!