From Karpathy's technical vision to boardroom reality—the first executive communication framework that bridges AI constraints with business strategy, complete with CEO-friendly ROI models and insights
@nate, this resonates. We are seeing the same split you describe: builders finding real value out of partial autonomy, while boardrooms chase fully agentic architectures that buckle in production. At Clevyr, our acceleration comes from Karpathy’s generation-verification loop: clear natural-language prompts, instant AI output, rigorous human review, repeat. (We just didn't call it that so elegantly...) Quiet, repeatable wins outshine slide-deck moonshots every time.
We treat LLMs like high-powered coprocessors, astonishing at pattern synthesis yet unpredictable at edge-case reasoning. Every workflow ships with guardrails, telemetry, and a human checkpoint. It lets us deploy AI safely today instead of betting the farm on a mesh that is still mostly marketing art.
Software 3.0 is not a future promise; it is already refactoring how we build and who gets to build. Thanks for grounding the conversation in working code, tight feedback loops, and honest constraints.
And on a personal note, your videos are perfect, and I love reading the details later in the day when it calms down for me. thx
glad the videos and articles combo is working for you! I love this example from Clevyr, makes a lot of sense and matches how I've seen actual AI work in practice in other engagements as well. It's all about letting LLMs do what they do best and sticking them in architectures where they can thrive.
I couldn't agree more with the code 3.0 standpoint. For my corporate implementation, I've found that adopting what Andrej calls 'LLM deep wikis' has been a very successful strategy. While I haven't tested it for agentic systems yet, it's incredibly successful from a human implementation standpoint, essentially creating an 'expert' in a given narrow focus. This has allowed my teams to onboard new technologies significantly faster, because the LLM's contextual understanding and grasp of the overall 'vision' provide them with a deep, foundational understanding of the topic.
This aligns nicely with the general idea presented in a Google DeepMind paper on causal models last year, where they suggest that efforts to improve LLM robustness may benefit from incorporating mechanisms that enhance causal reasoning.
For reference, the paper is "ROBUST AGENTS LEARN CAUSAL WORLD MODELS".
Great post. I’m encouraged by the idea that the best systems, at least for the near future, will best serve businesses that have the personnel to verify and sometimes correct the output of our LLM. Automation is intended to benefit both the customer and the human agents because it does 80%, and sometimes even 99%, of the lifting it is intended to do or help with.
We are currently “crawling” towards a solution for our team and customers. In the face of every other company trying to spend as little time as possible being an effective service model (cutting time and people), we see the technology as helping us deliver on the promise of dedicating MORE time to a customer and covering the things that may matter more, but we could never get to them.
I like that frame and it fits with a lot of the successful patterns I see in AI use—essentially expanding time-on-station for what a business does best by giving employees longer/wider spans.
How about treating Software 3.0 as the current reality while treating the "agentic mesh" as a vision for the future? There's nothing wrong with a vision as long as everyone is clear that a significant number of problems (context, identity, security, complexity, etc) still need solving. McKinsey can be blamed for selling a dream.
The problem to me is that it's not even a particularly compelling vision. McKinsey is definitely selling a dream here, but the dream is...suspiciously like the existing process wrapped up in an agentic trenchcoat. It's not particularly disruptive or novel. And so I take issue both with their positioning (selling what isn't really buildable) and also with the paucity of vision (there is much more interesting thinking around the future of AI for business out there).
Yeah, that's typical for McKinsey. Reminds me of their "Helix model" which was supposed to be completely different from a matrix structure but was actually the very same thing in all but the name. I destroyed that idea a few years ago.
If the goal is to apply the culture change you mentioned across all knowledge workers, it may be harder to achieve than the hypothetical automation of the entire customer success line. I say this because I’m not sure I see every knowledge worker thinking in terms of “software 3.0”, whereas people who have skills that product managers, software engineers, and system analysts traditionally have, possess a huge advantage in acquiring and applying these AI-augmentation skills.
So while all employees will need to become AI users, only some will go on to become AI “integrators” or “solution builders”. The bifurcated questions then become, what culture change will surface and foster integration/building skills in individuals who have them? And, what culture change will support future AI users in learning and embracing the solutions?
@nate, this resonates. We are seeing the same split you describe: builders finding real value out of partial autonomy, while boardrooms chase fully agentic architectures that buckle in production. At Clevyr, our acceleration comes from Karpathy’s generation-verification loop: clear natural-language prompts, instant AI output, rigorous human review, repeat. (We just didn't call it that so elegantly...) Quiet, repeatable wins outshine slide-deck moonshots every time.
We treat LLMs like high-powered coprocessors, astonishing at pattern synthesis yet unpredictable at edge-case reasoning. Every workflow ships with guardrails, telemetry, and a human checkpoint. It lets us deploy AI safely today instead of betting the farm on a mesh that is still mostly marketing art.
Software 3.0 is not a future promise; it is already refactoring how we build and who gets to build. Thanks for grounding the conversation in working code, tight feedback loops, and honest constraints.
And on a personal note, your videos are perfect, and I love reading the details later in the day when it calms down for me. thx
glad the videos and articles combo is working for you! I love this example from Clevyr, makes a lot of sense and matches how I've seen actual AI work in practice in other engagements as well. It's all about letting LLMs do what they do best and sticking them in architectures where they can thrive.
I couldn't agree more with the code 3.0 standpoint. For my corporate implementation, I've found that adopting what Andrej calls 'LLM deep wikis' has been a very successful strategy. While I haven't tested it for agentic systems yet, it's incredibly successful from a human implementation standpoint, essentially creating an 'expert' in a given narrow focus. This has allowed my teams to onboard new technologies significantly faster, because the LLM's contextual understanding and grasp of the overall 'vision' provide them with a deep, foundational understanding of the topic.
This aligns nicely with the general idea presented in a Google DeepMind paper on causal models last year, where they suggest that efforts to improve LLM robustness may benefit from incorporating mechanisms that enhance causal reasoning.
For reference, the paper is "ROBUST AGENTS LEARN CAUSAL WORLD MODELS".
Thanks Alex, makes a lot of sense! Love getting the specifics here. Linking the paper so I remember it: https://arxiv.org/abs/2402.10877
From my experience, human help at the design end at the beginning, as well as validating at the end
Great post. I’m encouraged by the idea that the best systems, at least for the near future, will best serve businesses that have the personnel to verify and sometimes correct the output of our LLM. Automation is intended to benefit both the customer and the human agents because it does 80%, and sometimes even 99%, of the lifting it is intended to do or help with.
We are currently “crawling” towards a solution for our team and customers. In the face of every other company trying to spend as little time as possible being an effective service model (cutting time and people), we see the technology as helping us deliver on the promise of dedicating MORE time to a customer and covering the things that may matter more, but we could never get to them.
I like that frame and it fits with a lot of the successful patterns I see in AI use—essentially expanding time-on-station for what a business does best by giving employees longer/wider spans.
I loved karpathy’s presentation
Now I need to go check out the other one 😳
LOL it’s quite a read.
How about treating Software 3.0 as the current reality while treating the "agentic mesh" as a vision for the future? There's nothing wrong with a vision as long as everyone is clear that a significant number of problems (context, identity, security, complexity, etc) still need solving. McKinsey can be blamed for selling a dream.
The problem to me is that it's not even a particularly compelling vision. McKinsey is definitely selling a dream here, but the dream is...suspiciously like the existing process wrapped up in an agentic trenchcoat. It's not particularly disruptive or novel. And so I take issue both with their positioning (selling what isn't really buildable) and also with the paucity of vision (there is much more interesting thinking around the future of AI for business out there).
Yeah, that's typical for McKinsey. Reminds me of their "Helix model" which was supposed to be completely different from a matrix structure but was actually the very same thing in all but the name. I destroyed that idea a few years ago.
https://unfix.com/blog/lets-unfix-helix
Agentic Mesh sounds like a marketing buzzword.
If the goal is to apply the culture change you mentioned across all knowledge workers, it may be harder to achieve than the hypothetical automation of the entire customer success line. I say this because I’m not sure I see every knowledge worker thinking in terms of “software 3.0”, whereas people who have skills that product managers, software engineers, and system analysts traditionally have, possess a huge advantage in acquiring and applying these AI-augmentation skills.
So while all employees will need to become AI users, only some will go on to become AI “integrators” or “solution builders”. The bifurcated questions then become, what culture change will surface and foster integration/building skills in individuals who have them? And, what culture change will support future AI users in learning and embracing the solutions?