Nicolas Ramsrud

scaling productivity means scaling the organization around it. agents need the same thing you figure out how to scale agentic organization, you unlock unimaginable productive capacity.

its completely intuitive but not a lot of people are talking about this and i think its a blind spot. scaling models combines intelligence with storing information in the weights. this implies that you can remove the information and keep the intelligence greatly reducing the amount of parameters. if the model knows how to get trustworthy information this means a frontier capabilities at 100x less cost or more. this also implies that if the information is key to performance in a specific sector, it can be added back into this “core model” leading to frontier performance in that specific sector at a strong multiple less cost. once we have a layer where intelligence can securely coordinate (cough) there will be no reason to use high cost general frontier models when you can get same performance by routing between a network of specialist models at >100x less cost. by cough i mean we are building not just the layer but the models

Make the work ai does verifiable and the broader economic impact of ai actually happens

I’ve been saying this for a while. For many types of software, ai is already the primary user. Ai is already dominating search. Soon ai will be the only user and the way people decide what they want Marketing to agents is, or soon will be, the whole game

this is the very same first principles deduction for why our OS and runtime model, built upon Lurk, is the only comprehensive solution to building trust between humans and agents. trust is required to let more and more of this growing intelligence, and its productive capacity, scale but trust is hard or impossible to establish in the model itself. trust instead must be established in the environment where they work: the runtime. the runtimes agents operate in today were not designed for agents. constraining actions, access control, data, code, programs and execution, all is course grained. therefore we need a new type of language, OS and runtime so that agents can work with tightly scoped authority and produce verifiable results. if you have this, you dont have to trust the model. the model can only do what you grant it the capabilities to do. even a year ago people were very close minded to this, and still today I meet a lot of resistance to this idea over those that thing we will magically achieve both super intelligence and one that we can fully trust with unbound access. But I am happy to see this is starting to be realized. alignment is another problem entirely and i believe there is a way to develop trust in the model, which is what emerges when you combine training with these lawful runtimes, which is something else we are exploring

this is so clearly the biggest scaling bottleneck for AI. not just for companies but in general. you either have to trust or verify. and if your only means of verification is human review, you cannot scale productivity with AI. further, coordination is what takes the productive capacity of one and scales it. This is the mechanism behind scaling human output in organizations, through trade, and markets. AI needs the same capacity, and that requires machine legible trust: verifiability. and it needs to come from where they do work: with code and data in a runtime. if agents can prove the output of their work, and that proof is portable, you can use that to create networks of coordinating agents. this verifiable work then creates a highly valuable training surface as information with machine verifiable truth. this has implications in model training and specialized models for all different fields, applications and organizations. and because machine verifiable truth then can be composed, you can eliminate reverification and rework, create information asymmetries (markets!), and continually expand a graph of high value information for building bigger and bigger models as well as deeper and deeper specializations, that can then coordinate to achieve higher and higher collective capabilities. This goes against the one-large-model thesis but i am convinced this is the end game.

the extended Lurk paper has just been published in the Journal of Functional Programming. a big moment and one long coming, it includes a lot of what has been in work, but also is a snapshot of where we were in july of last year. we have been very quiet over the past year, focusing on scaling our execution to build the full vision of lurk. so while this is a huge accomplishment, seeing it shows just how much has been done. The next paper is going to be mind blowing. truly special thanks goes to @huitseeker for leading this one and Nada Amin @hseas for feedback and guidance.

talking about prompting as a skill has somewhat fallen out of vogue as models have gotten better, but its still the thing that separates the good from the great almost everyone views their interaction with these systems as flat transactions, never considering that the act is a collaboration that distills context, and the shape of the context is what determines outcome. therefore, context is everything. there exists some context by which anything is possible. this is the most coherent and complete exploration on this concept that I have read.

This is not your normal AI content but an exploration that is a reflection of his ever-evolving understanding of these systems. what is most interesting is the process 'behind the curtain' where novel insight is achieved by careful and precise prompting. understand the meta-point and you might gain some insight yourself

I shared this in a comment to another post and the response showed that many people deal with the same issue. If you are using Codex or Claude in large codebases, you can reduce the amount of tokens used in search and reading files by 30-60% if you make an agent-efficient repo map and symbol extraction list and direct your agents to use these as the search entry points. I made this indexing tool to do just that, agent and skill doc examples in the readme. github.com/Nramsrud/ARK-inde…

I think as I tab between my twelve instances of Claude ignoring the “you’ve reached 99% of your weekly limits” warning on a Tuesday evening

Jordan Hubbard (nvidia) experimenting with building a language specifically for coding agents. it turns out its... lisp-ish. was lurk right all along? github.com/jordanhubbard/nan…

Not only can I access my agent farm from anywhere (including my phone) via ttyd and tmux (behind tailscale), but they now send me a notification to my iPhone if the are done with a task, blocked or need my input. Chat integration next.

I have noticed that a side effect of ai becoming more capable is that time spent not being productive feels so much more painful

weekend update. the agent farm wrote and merged 227,746 LOC, 2,907 tests written (all passing), 120 commits, 41 PRs, 54 total agent hours (across 17 running hours). ~4-6 of my hours. this is insane.

last night while I slept my agent farm wrote and merged 45,152 loc, 2,260 tests (all passing), across 56 commits and 14 PRs, in 12 agent hours (4 hours wall clock) I literally cannot write spec docs fast enough

absolutely absurd statement free speech has no asterisks

I have been doing this for months now and can confirm it is the single biggest lever to getting 90+ sleep scores every night. My heart rate is lower and bottoms out almost immediately when I fall asleep, I get 2+ hours of REM, HRV is 30% higher than otherwise, and I wake up feeling sharp. I do better with 5ish hours between eat and sleep so it is worth experimenting a bit