Read, Write, Own - LLMs Edition

In the grand New Year of 2025, I will be writing again. This is a framework I have been sharing with founders and hope that you will find it useful as well.

Chris Dixon created this 3 pronged framework called Read, Write, Own to describe how the data paradigm and user behavior has changed over time.

To roughly summarize: 

• Web 1 consisted of static web experiences that were read only allowing users to passively consume

• Web 2 consists of a more dynamic web experience in which users can also write their experiences enabling content creation ie video, blogging. However, this is done on centralized platforms like YouTube and Twitter which can arbitrarily change underlying frameworks and algorithms. Examples of adverse changes include Meta’s infamous video algorithm changes and Twitter/Reddit turning off their APIs and killing a thriving 3rd party ecosystem.  

• Web 3, consists of true ownership in the decentralized protocols that one publishes on and turns users into active stakeholders with a financialized incentive. This unlocks new models of collaboration, monetization, and automation, but carries immense risk as we can clearly see from the selfishness / greed of those within the space. 

As a sucker for 3 pronged analogies myself, I have adapted his R,W,O for myself regarding Large Language Models (LLMs) which create a new paradigm of nondeterministic programming. The distinction between deterministic and nondeterministic programming lies in the predictability of the outcomes based on inputs and condition; deterministic programs produce the same outputs every time and are inherently reproducible while non deterministic programs produce different outputs based on the system and session state. Once LLMs mature, they will help be integrated into nondeterministic systems that are adaptable and capable of generating diverse outputs tailored to context, opening new possibilities in user interactions as described below

• The first set of capabilities with LLMs consists of only reading / summarizing / accessing data. LLMs can be used to interact with documents, underlying databases, and with the right product abstractions, can provide detailed explanations of rendered content like visualizations, reports, and dashboards or break down complex workflows into step by step processes tailored to a user’s needs. This new paradigm of nondeterministic programming is powered by API providers fighting on price and increasing performance on every metric, providing developers new cheap (albeit uncontrollable) ways to add sentiment analysis, data extraction, and other novel forms of data manipulation.

  • I am especially interested in the use of LLMs to accomplish sorts and ranking for semantics. Imagine creating an opening at your company and having 1000 applications. With the right prompts and context, you could have an LLM read every application and rank each by its chance of success at the job and highlight any standout green and red flags for you. This ranking feature is tied to reasoning and data interpolation and will improve as the models improve ¹

• The next tier of LLM capabilities is to write. I believe that this is best shown through the advanced set of LLM capabilities like tool calling and copilots / assistants that work in tandem with humans. This will start in deterministic fields in which an answer can be discovered via product documentation or code execution, but will eventually subsume a majority of knowledge work. I believe that LLMs are in the nascent stages of a reasoning explosion which will make copilots even more effective.

• The final tier in the framework is own; LLMs will consist of owning their work without any human input in an agentic self orchestrated loop. This fits into the paradigm of ‘selling the work’, and ‘selling the outcome’ which can create value independently of a human. I believe that this type of LLM workflow will have a human setting the parameters for the agent and letting the agent do the work almost as a conductor. I cannot fathom the economic implications - divorcing labor from human work in a nonlinear way hurts my brain but it is a concept we should become comfortable with,

Now that I’ve established this framework, what does it mean and how can we tactically apply it? I believe being an AI native firm will lose its meaning as LLMs are just another tool for achieving business outcomes. An AI native accounting firm and a regular accounting firm look vastly different today but that rift will decrease as the knowledge / alpha of using LLMs permeates through the work force and as implementation costs decrease. But what does that tell us about building businesses today?

My own philosophy for building and advising builders to to build the best product experience for the user. When incorporating LLMs, you want to ensure reliability, usability, and value in responses and if you cannot ensure that, you need to lower down a tier. That will eventually mean making a decision of incorporating LLMs - or not!³

If your product does not need ‘AI’ or you can create the optimal experience without it - do not incorporate it. You will distract your customers and employees from the real goal of pushing forward business outcomes. Only incorporate LLMs if you see a real way to increase user delight and win on either price/experience/usability and start at the Read Tier. There are a variety of LLM experiences that are too expensive to implement today to achieve viable unit economics.  I believe that LLM calls will reduce in price whether through competition among API providers, increasing viability for local models, or the availability of open source best practice frameworks which will significantly reduce the implementation of costs over time. 

Building out LLM capabilities past Read will require refactoring infrastructure and implementing evaluation (evals). I would recommend watching from the sidelines; now is not the time to struggle with implementation when there are customers to serve. However, organizations that adopt a strategic, incremental approach beginning with foundational use cases that prioritize measurable outcomes will not only avoid the trap of unnecessary complexity but will also position themselves to scale effectively as costs decline and capabilities expand.

I do believe that the reasoning explosion I mentioned above will happen sooner than later and the largest question going forward is: What level of investment should I make into future LLM capabilities before they have been built. I spent much of the second half of this essay convincing you not to incorporate LLMs unless there is a good reason. However, if you see that your business can take advantage of LLMs this is a real conversation to have. This prebuild can range from refactoring infrastructure, creating product abstractions that pass context to LLM calls, and beginning to actually build agents with current frameworks even as their functionality is not quite there yet.

Time will tell and if any founder wants to talk through this quandary, please reach out to [email protected] and I will expeditiously return the message.

P.S.

Another open question to me is: will we have 3 separate products, one that has different levels of human interaction according to the framework. Firms like Cursor and Codeium are creating IDEs that provide real time suggestions for completing code snippets as developers type and have various ‘talk to the codebase’ features that satisfy the Read and the Write. Once they create their agentic coding workflows, will their UI still be IDE based or something entirely different? Will they break their product up into different components? Will the profile of the coding copilot differ from the profile of the agent conductor? Answering these questions will be done by designers with far greater taste and I am excited to see what we as an industry create. 

1. This is not an original thought - I came across this idea in a newsletter or podcast but I have forgotten it. If you know who mentioned this, please let me know and I will promptly give credit.

2. I very much hope that we as a tech community find a new word to replace agentic.

3. As a web3 degen, I truly dragged myself everywhere looking for real use cases outside of speculation and near instant transactions because I wanted to be clever. It turns out that those two use cases that I thought I was bigger than have been where most of the mindspace as accrued to, culminating with a unicorn acquisition of Bridge to Stripe, both the crypto industry and Stripe’s largest acquisition.