owaowo 👾

People keep asking how companies like Ramp, Linear, and Anthropic ship so much software with agents. There is no new SDLC or secret set of tools. The tools are increasingly available to everyone. You can build your own, use Linear, Claude, or whatever works. The bigger difference is ambition. These companies encourage their teams to build more, experiment more, and let agents do more of the work. That's the main different between "AI native" or not. You try to use the technology to the max. But don't get there by adopting a tool. You get there by changing what you believe your teams are capable of. So in the very simplest way: - let agents build - make it easier and faster to contribute - increase the ambition of what your teams attempt. In practice, that means things like: - Standardize around cloud agents. Most of our code (5M lines last quater) is now written in Linear with our Linear cloud coding agent. Cloud sandboxes make standardizing the system and environment easier for everyone and enables faster contributions - Linear agent (with code access and cloud agent) operates in Slack too. If there is a discussion or someone notices a problem just say "@linear fix this" (decrease the time from something noticed to fixed) - As bugs, errors etc arrive in Linear, send them to Triage, and have Loop to auto-investigate (connect to tools like Sentry, Datadog), and write a fix - Let designers, PMs, non-technical people contribute code where useful. If they notice a problem, tell them they can submit a fix. If they want to try something, ask them to build a prototype. - Set up code review agent and system to make sure PRs are safe to merge. - Run security as an automation, have agents constantly try to find security issues in your codebase, as they find, send it to a coding agent - Instead of working on issue level, increase ambition, make a whole project, ask the agent to ask questions, investigate code base, create issues and have the agent work on the whole project at once. Set up goals like "give me proof that what you build works". All our PRs now come with screenshots and videos and preview build you can - Bring customer feedback to the system like Linear, and make it available for your team. At Linear every project now starts with 10-500 customer notes that influence the direction of the work, and which customers we should ask to test the solution first. - Set up Loops/automations to constantly update reference documents for agents, like product area/code owners so work can be routed to right humans Then what is left is how do you set direction for your teams (initiatives, projects they should work on) and how do you review results (better code review, Linear does it with Guided Reviews), how you maintain quality/cohesion which is not a new problem either. Basically most of the problems you always had remain, but the execution bit is now more and more solved. At @linear we can help with automation, cloud agents, and code review, customer feedback and planning bits. But we can't change your mindset or how your company operates. If your company is still spending energy debating jira workflow steps, you're probably solving the wrong problems.

so far this is working great. i tried this in the past and it wasn't as great. some details - we have long running sessions per project - opus is told to do all work in deepseek subagents - be strict about scope when creating a task - ask for proof of work instead of a summary - do a one step review of the diff to make sure it's not totally off opus is amazing at this role

The biggest advantage of Codex over Claude Code right now is computer use imo, people asking me why. 1. APIs/MCPs are painfully slow. Reading a Slack channel can take forever. 2. Legacy enterprise software is often so confusing that humans struggle to use it. I couldn't figure out an expense reimbursement app. Codex figured it out and completed the task in minutes. Computer use turns every UI into an API basically.

Since August, a team within METR has worked with internal Anthropic information to assess whether AI itself is accelerating AI development. We are now publishing a report summarizing the results of this risk assessment.

Background subagents in Claude Code now use less of your usage limit. A running subagent updates its status line every 30 seconds, and each update used to re-read its whole conversation. Now Haiku 5.5 writes it from short notes instead, so each update costs about 99% less.

Releasing the model weights of Kimi K3.1 Pro: huggingface.co/moonshotai/Ki… We'll keep advancing open intelligence.

Humans are ~1,000,000x more sample efficient than LLMs, and there are no good explanations. We’ve found that a big part of the answer is that LLMs are *massively* underparameterized by many orders of magnitude. The brain has 100T synapses learning from ~200M tokens of language. Chinchilla prescribes a 1T-parameter model on 20T tokens. That’s 7 OOMs apart in parameters per token. After a lot of painstaking work, we present new data-constrained scaling laws and learning algorithms to scale to (extrapolated) quadrillion-parameter transformers. w/ @akshayvegesna, @EthanBaron78553, @asatheesh6561 Full paper led by @EthanBaron78553, with @ShmuelBerman, @andrewgwils

We’re introducing Hydrogen-1.0, our first open-weight model for scientific research. We post-trained GLM-5.3 (@Zai_org) on our internal scientific data, lab records, and discovery trajectories. It scores 38.3% on FrontierScience-Research vs GPT-6-Astra’s 33.3% and scores 93.0% on GPQA-Diamonds. We believe science should be open, collaborative, and democratized, rather than proprietary. Only open weights let researchers inspect, adapt and build on the unprecedented speed of AI growth. There are many recent models named by chemical elements, but we believe Hydrogen-1.0 is the first aiming towards and open scientific superintelligence for discovery. 🧵

There has been a running debate about what RL post-training actually does: does it teach new capabilities, or does it mostly "sharpen" the base model, i.e., pile probability onto a few behaviors the base model already had and suppress the rest? The evidence so far came almost entirely from math and coding, while this paper study agentic tasks. They find that: 1. Post-trained models win at small K, but the base models' pass@K curves climb more steeply and cross over in most settings before K = 128. 2. The bigger the model, the earlier the crossover. 3. Post-training amplifies the base model's wrong modes as well as its right ones. The paractical take away is that If you have a large compute budget and a reliable verifier, a big base model with a light harness may reach more solutions than post training it. If you get one shot, you should use the post-trained model instead.

Introducing agent-desktop agent-browser, for your whole computer → Computer Use for any coding agent → macOS, Windows and Linux → Runs in bg while you keep working → Agents use your computer in parallel, each with its own cursor

Stageshow: a skill that helps your agent make better faux UI stills, animations, bento grids, and more. npx skills add benjitaylor/skills --skill stageshow