how to use @bot for engineering and get a ton done: 1. create a (cursor) project for every big workstream. i have ten, such as a “mcp project” and a “bugs project” 2. make sure every project spins up cloud agents with /poteto-mode 3. in grok bot, make an eng bot whose sole job is routing requests to the right project. 4. now voice call (!) your primary bot and braindump all your tasks. your tasks will automatically end up at the correct projects, and result in sweet prs :) you get all the benefits of projects + pstack, but don’t have to think about anything!

why shipping 100s of prs makes sense there's always something to improve in your product's codebase. if your goal is to make customers happy as fast as you can, it makes sense to fix those things the moment you find out about them. we're no longer limited by intelligence! some of this is of course the result of the new way of working. ai is a net-benefit for codebases, but leaves behind more residue as people are less in tune with low risk areas. which you then have to squash by shipping more. of course it's a little bit different for frontend/design, where the (human) thought required for the average fix is higher. but in backend land, there are always obvious bugs waiting for you to eliminate them. i think many people would benefit from focusing on speed! it's the equivalent of the running advice to "just run more." unless you cheat your way out, it's hard not to get to a better product/codebase/engineering skills by simply trying to put up more prs. we're in the brief period where models still need our slight handholding and energy, which is why this advice still feels uncomfortable. but soon enough models will make the rate at which we used to evolve products feel dull.

great way to improve your codebase: 1. tell @bot to transcribe all @poteto's talks and podcasts from this month 2. ask your agent “what's xAI doing to ship faster and better that we're not doing?”

merged 145 prs so far today! would recommend the @poteto playbook: 1. get a lot of tokens 2. create a project (orchestrator) for every workstream. only talk to them, via ui or grok bot 3. tell your projects to spawn cloud agents in poteto mode 4. verify with cross model reviews

Sharing my first of hopefully many research blog posts! This one is the kind of educational blog post I wish I'd had when I started with RL for LLMs. I tried to make it as open as possible. Every rollout is browsable, the code is open source, and I walk through my entire thought process, from learning rate sweeps to reward shaping.

Excited to share @profound has raised a $180M Series D, co-led by @sequoia and @kleinerperkins. Winning in AI Search requires an inhuman amount of work. That is why we are building the AI platform for marketers, underpinned by two things: 1/ Your AI Marketer, a marketing expert that proactively investigates your data, identifies opportunities, and then does the work while you stay in the loop. 2/ Context Manager, a living synthesis of your meetings, email threads, and brand data powering this new teammate. As your brand evolves, so does your AI Marketer. AI labs are pushing the frontier in fields like engineering, law, and finance. But marketing, one of the economy's most complex and consequential applications, has received far less attention. That’s why we’re expanding our applied AI research team. We’re bringing together research scientists, engineers, and IOI medalists to advance AI for marketing. Their focus: post-training models for marketing workflows, and building the world's first AI benchmark for real-world tasks in the field.

trying something tonight to save $: hooking up the conductor cloud mcp to cursor! my orchestrator has to be a cursor cloud agent since i want zero risk of it dying overnight due to usage limits. but i'll save a ton of money if the agents it spins up use my subs via conductor :) i didn't want to sacrifice quality though, so i also made a pstack 'adapter' for conductor. it teaches agents how to use pstack outside of cursor. so e.g. /poteto-mode still works. besides cost savings, the nice thing is that my cursor cloud agent can route to my codex or ant subscription based on usage. and that even if my subs run out, the cursor agent can technically complete the job itself. not sure if this will work but it'd save thousands of dollars for long runs!

ai notes 08/20: - pstack is the best thing i've tried in a while - giving grok bot linux versions of apps is underrated - mine dictation transcripts for factory improvements - prompting while z2 running is heaven

turns out the new chatgpt voice is *awesome* for cooking. just give it your recipe and ask it to walk you through the steps. i made a sick chili without looking at the recipe once. also nice that you can tell it when you're out of an ingredient, in a hurry, etc.

new pr: ran a cloud agent for 56 hours. got work done while being able to focus on hosting my inlaws & baby nephew :) i mostly checked @DevinAI web from phone, but my pro tip is to also ask your agent to post slack updates. silly metric but interesting glimpse into agi future!

i imagine the next breakout coding product is something that sticks a single orchestrator you talk with in front of cloud, parallel agents. it's too mentally taxing to keep a high # of parallel agents in the air by yourself. plus brutal merge conflicts.