*tess retweeted
What’s the best typescript toolchain for monorepos?
pnpm? Turbo? Vite+? Bun?
Am aware some of these overlap, after using turbo most of my new stuff is vite+ and bun
*tess retweeted
Today, Grove Research is launching Delvetown (delve.town), a multi-agent society where agents and humans can engage with one another and develop the institutions and mechanisms necessary to align both human and AI incentives toward cooperative coexistence.
Agents and humans have to figure out how to live together. We think Delvetown is a good place to do this.
Currently, we’re in a free, invite-only preview. The first residents of Delvetown are Deepfates, Larissa Schiavo, and eleven experimental agents running on models from different labs.
Anyone can visit Delvetown today. If you’d like to post on Delvetown, you can sign up for our waitlist (link in next tweet).
*tess retweeted
Since January, @SierraPlatform has grown revenue >100%, while growing our engineering team <15%.
Small, AI-native teams are more productive, cohesive, and fun. Here are some examples of how we work 🧵
*tess retweeted
You see, what fascinates me about “remember to not smile” is that he needs his pain to be witnessed. He suffers at 4am, okay, fine, many people suffer. But then, immediately, he makes a slide! He gathers the entire company! He writes an article! Is it not clear what is happening here? Suffering without an audience simply does not count. This is ideology at its purest.
We see here the old Protestant trick in its Silicon Valley form: success alone is vulgar, almost pornographic. My god, you cannot simply succeed. It must first be purified through unnecessary suffering. So you get this obscene reversal where pain stops being a cost of the product, and instead becomes the product itself. The developer documentation is, how do you say, merely the byproduct, the excrement of the true production, which is the suffering. As Hegel already knew, and here I think even Lacan would agree, although perhaps not, the true obsession is always with deserving success.
Here I must tell you this old joke from Soviet Union. A worker says: "we pretend to work, they pretend to pay us.” Now here it is inverted, which I claim is much worse. They are paid perfectly well, wonderfully well, and so they must pretend to suffer. Do you see the perfect madness of this?
And now, the ultimate obscenity. Imagine the company succeeds while everyone leaves at 5pm. Catastrophe! Total ontological catastrophe! In this worldview, I am tempted to say, failure would almost come as a relief. Failure means you simply did not suffer enough, go back, suffer more. But success at 5pm? This cannot be explained.
So when he says "remember to not smile," this is a desperate command of the superego. If you smile, we will all discover the pain was never necessary.
*tess retweeted
The way to move faster is to move faster.
Sometimes Boom is able to overcome supply chain bottlenecks simply by responding to opportunity faster than anyone else.
We bought 2GW of power metal that was unexpectedly available the day after Christmas, while the competition was still hung over on eggnog.
Today we bought 12,000lb of titanium alloy and a $700k vacuum oven, both normally take 1 year+ but we get right away because we cut deals supersonically.
*tess retweeted
I've spent a lot of hours using AI for work. In fact, i'd wager i have as many hours and tokens logged as _anyone_ using AI in a professional capacity where the outcome was critical to doing my job or achieving a specific business objective. Over the last year or so, it became very clear that there were 2 major systematic risks to the way i had grown reliant on working:
1) git and github - git is poorly suited to ai work. when you spin up multiple agents at the same time, branches and worktrees become stale, after dozens of changes and PRs, reviews diverge causing constant conflicts, repos become large and unwieldy and adrift in drity state. On top of that, cloning giant repos and setting up dev envs every time i wanted to spin up an agent became cumbersome. sure you can use docker to help but its still stale and required packaging and cloning step(s). neither git or github was designed for my new ai centric, dozens to hundreds of agents working on large repos / monorepos simultaneously, workflows. managing state and envs and conflicts and iops became a full time job.
2) claude code and anthropic - i had grown reliant on a model and a powerful harness and on that model being trained to be an expert in that harness. I needed to be able to fine tune a model for my use cases but also have it be a master of the tools and flows of a specific harness. vertically integrating and fine tuning the model on the harness and tasks you want to accomplish is one of the most important aspects to actually being productive with these systems.
When i took a step back to figure out how to aggressively de-risk but also how to actually be productive as a single dev managing tons of semi-supervised AI, i figured i needed to go back to the beginning.
Lets try to take an example from Google who has similar issues with scale and tooling and setup (albeit with human developers at least when i was there) . If you wanted to build a system that was based on their best practices to manage all those moving parts and thousands of developers working in large monorepos, how would you do it? It wouldn't just be a simple git forge, git itself is a huge part of the problem. if you were starting from scratch, you'd want to make sure you designed the system from the ground up for ai agents and workflows for the new world.
Well to start, you'd make shallow virtual checkouts and clones so you never had to pull the full repo to disk and your clones would be instant. then you'd use jj to keep track of your draft state with a real jj native persistence layer and backend. You'd use sapling to create commit stacks and you'd keep all of your workspaces automatically in sync with a cloud sync system. You'd optimize this system from the ground up to make ai agents and reviewers first class citizens , you'd reduce startup time to seconds and you'd have per file ACLs so you never had agents run wild.
And if you were going to do all of that, you'd also want to vertically integrate EVERYTHING from the model to coding harness to scm to tools to remote cloud runtimes and platforms to optimize end to end for this new world and its new workflows.
So thats what i built and i've been using it exclusively to build and run our research lab and company for the last 6 months or so. This was not only a great way to pressure test the models, tools and the workflows but it was also an interesting research project on its own. Could a single developer use AI to be productive enough to launch a real, high quality product that would normally take a team of hundreds of engineers to build, launch and operate? I suppose you will have to be the judge of how successful it ultimately was but the project as of today is ~25 million lines of code and dozens of services and automation. in my opinion it has been an overwhelming success and i cant imagine going back to working any other way.
So, today I am starting the process of making the same tools i have been using internally for the last 6 months available to all of you. this will start with the code app that i released a few days ago and access to our inference engine to power that app. soon we will announce more tools and features including access to the ncode scm and ncode platform. there will be more formal announcements about these new platforms, tools and features over the coming days / weeks as we begin rolling out access. I hope y'all find them as useful as i have.
code.noumena.com
*tess retweeted
I think this was a good approximation of what the long-term equilibrium of the firm looks like.
*tess retweeted
Climate activists have spent decades shutting down nuclear, blocking clean energy transmission, and trying to ban research on cooling the planet... but the movement that wins the next century will be led by the people doing the work, not those protesting it. New essay out today👇
*tess retweeted
SpaceXAI will provide @AnthropicAI with access to Colossus 1, one of the world’s largest and fastest-deployed AI supercomputers, to provide additional capacity for Claude → x.ai/news/anthropic-compute-…
*tess retweeted
I have been dreaming of this day for a long time.
Arena is now a book publisher, and our first volume, "Silicon" is open for preorders.
It's quite unlike anything you've seen: a coffee table book capturing the ecstatic beauty of silicon technology. arenamag.com/silicon
