# Trustless Economic Agency Source: https://x.com/goodalexander/status/1849873125430657237 Kiwi News: https://news.kiwistand.com/stories/Trustless-Economic-Agency?index=0x671cdaedf02a169bcc0d4416ca5ef154961fe0e18c38c2769592f5dcc3365a8b996b4e5b Submitted by tim on 2024-10-26 UTC, 2 upvotes, 1 comment ## Summary The author describes using AI to improve individual economic output, centered on a concept called Economic Agency: an entity making money by choosing a profitable game and winning it. In a "Deterministic Agency Experiment," a Post Fiat token on the XRP network issued tasks from a user-written Context Document, and GPT-4 scored the user's focus, motivation, efficacy and honesty over repeated runs. A "Game-Selection Enhancement Experiment" then used Claude to rewrite sections of the context document and O1 to produce a final version, with a reported 41.2% score improvement on a sample text. The author says the scores port almost linearly to other models and that the approach yields linear, not transformative, improvements. ## Article This is a long post. So to save you time, it's about using AI to enhance individual and collective economic output - fairly mechanically, with live examples, simulations, and code. If that's not up your alley I suggest you stop here. You can find the code driving this here if you're curious about the prompts/ specifics. The most interesting statement re: AI development in my view is Altman's initial "tokenomics" for OpenAI. Namely, that AI would become smart enough that you could just ask it to make a bunch of money and it would. And that somehow - that would make all the equity holders in OpenAI whole. These tokenomics have been changed bc of the bizarre Foundation Incident that freaked Microsoft out. But as AI systems advance, and we see things like "Truth Terminal" which is making money via a meme coin. The question remains. When will AIs be smart enough to make decisions that make money. To simplify the language here, I'll introduce the idea of Economic Agency. Namely, the ability of an entity to make money for itself by choosing a game that makes money, and then winning it. I started with a less grandiose question about 5 months ago, asking "Can AI systems evaluate my agency meaningfully?" Deterministic Agency Experiment To perform this experiment, I made a cryptocurrency token on the XRP network called "Post Fiat". A node on the network ingests my "Context Document" which outlines my high level goals, strategies, and tactics. Then it outputs tasks which I then perform to the best of my ability, submitting a verification after the task is completed with relevant details. Then an AI system dispensed a reward to me based on my actions I then batched these tasks and used GPT4 to evaluate my performance across multiple different attributes over time. Specifically - focus (not switching context), motivation (moving continuously), efficacy (moving with high impact), and honesty (not gaming the system). You'll notice an orange and a blue line, that are essentially the same. The key observation doing this analysis - was that AI systems come to more or less deterministic answers about a user's performance across different fields if you run them a large number of times This scatterplot shows an important but I think under appreciated point about LLMs. LLMs can take qualitative inputs and convert them to quantitative outputs, in highly reproducible ways - assuming you have the base model, and the prompts used. Put simply, they can take management BS that use to be finger in the air / subject to whether the manager was in a bad or good mood. And make it a number people can agree on and reproduce. Previously - this would require the mechanics of a large corporation - managers talking to one another, coordinating performance reviews and so forth. But now, you can just run an LLM 50 times and get a fairly objective take on a User's performance over time. Ex post objective functions on performance are necessary to any sort of AI agent system - so this is an exciting conclusion bc at some basic level if you agree on the criteria, you can say very definitively if something is a good or a bad idea, or if performance was good or bad. Without trust. The Cold (dumb) Start Problem The problem with the above analysis - is that the tasks the user performed were from a context document generated by the User. So even though you can sort of tell if the user is moving towards their goals, those goals might be stupid. In so many areas of life, success is determined by game selection not actually playing games well. In the case of Truth Terminal for example - if its job were to make a high cash flow SAAS start up that'd be really difficult as compared to promoting a meme coin. Because it's stumbled on an easy game it's winning much more than other AI agents playing harder games. Play stupid games win stupid prizes. The problem, up until the launch of the recent O1 Preview models and the Claude 3.5 Sonnet Models is that - GPT4-o wasn't smart enough to generate meaningful delta on the user's actual objectives - leading to a "stupid game problem". Sometimes a picture is worth 1000 words. You can see in the above image that Sonnet is a step change in terms of planning ability. It's able to make complex, delightful Minecraft worlds in a way that wasn't possible a couple months ago when I first did the agency Experiment. The Game-Selection Enhancement Experiment So since the new Claude model is so incredibly powerful it was finally time to take a crack at solving the cold (stupid) start problem. So let's say a user has a context document. The context document contains her high order objectives, the strategy to reach those objectives, and then the tactics that flow through to accomplishing those strategies. Per the experiment above, the context document is the thing that gets parsed in order to actually determine the next action for the user - so it's actually the real gating function That is to say, Context is the Game. And like a Minecraft Castle improvement program as shown above, the question was Game Definition If you have a poorly selected, poorly scoped Game you can work as hard as you want and will still lose. Essentially the model I have come up with for context document improvement is using Claude to iteratively generate new sections of the document, then have O1 rewrite the highest scoring re-write here is a diagram The basic question of the experiment is can we reliably improve a User's high level planning, and Context, as that is the basis of the user's tasks. An example Let's start with a basic sentence to make this intelligible: Initial Sentence "My goal is to create an AI agent that talks about capital markets in a highly engaging way bc Truth Terminal is growing really fast and Truth Terminal is an AI agent, I think this is the best way to capture the current zeitgeist( Truth Terminal is $700m market cap trading $100s of millions a day w Binance listing). The key is to make this thing really engaging bc ultimately its ability to grow my crypto network is gated by how much ppl like the content. I think there's a strong cohort of AI native ppl who aren't activated in the crypto space who love trading their PAs that will be activated by this so the capital thesis is appealing / believable" First Improvement Pipeline (Anthropic / Sonnet) So we can see above that Version 3 has a 3.5 score. Note that the system starts by trying to actually understand what the user wants and doing a better job than the user would be able to him /herself. Version 3 then gets taken by O1 and mashed into a final version that more linearly corresponds with the user's voice/ O1 converts the Corporate Speak Claude back to the user's own voice / does some editorial edits New Version: My goal is to create an AI agent that talks about capital markets in a highly engaging way. Truth Terminal is growing really fast—it's an AI agent with a $700M market cap, trading hundreds of millions a day, and now listed on Binance—and I think this is the best way to capture the current zeitgeist. The key is to make this thing really engaging because ultimately, its ability to grow my crypto network depends on how much people like the content. I believe there's a strong cohort of AI-native people who aren't yet active in the crypto space but love trading their personal accounts. Activating them through this platform makes the capital thesis both appealing and believable. By connecting these traders with engaging AI-driven market content, we'll drive meaningful network adoption and value creation within our ecosystem. After We get the O1 version we can then evaluate scoring of how good the new version is versus the old version The aggregates of these scores are then turned into a single percentage improvement score You can then get a picture of how good the new version is versus the old version in a quantitative format with explanations if you like (will omit O1's rant for brevity here). If you refer to the earlier example where I'm scoring my own work performance, the methodology is nearly identical. Using a large number of LLM runs to deterministically converge on a quantitative output. "How good is this thing across these parameters" Iterative Process You can see above that the score improvement is 41.2%. That's a huge leap, implying that the initial text was quite bad. But now that the formatting fixes are done the question is can the system improve upon its own reasoning iteratively. The scores are derived on OpenAI but they port almost linearly to other models ( feel free to test this with non color coded versions) So now we converge on the final version: "I'm aiming to create an AI agent that talks about capital markets in a highly engaging way. Truth Terminal is growing rapidly—it's an AI agent with a $700M market cap, trading hundreds of millions daily, and now listed on Binance—which I believe captures the current zeitgeist perfectly. The key is to make the content truly engaging because the growth of my crypto network depends on how much people enjoy it. There's a strong group of AI-native individuals who aren't yet active in the crypto space but love trading their personal accounts. Activating them through this platform makes the capital thesis both appealing and believable. By connecting these traders with engaging, AI-driven market content, we'll drive meaningful network adoption and create value within our ecosystem. Combining AI technology with insightful market analysis, we'll provide genuine value to participants and organically grow our network." Obviously this is much better text. It explains the linkages between the market environment and a concrete deliverable (AI driven market content) to reach an objective function (organically growing our network). It's worth noting that Claude - even without the scoring prompts - thinks this version is 50% better than the original which aligns almost exactly with the spreadsheet scores (55% -> 87.5%). The second meta point is that there are specific ways to write that make LLMs think you're smart/ on the right path. Which is relevant to my earlier posts about AI cults which ascribe significance to the latent preferences of LLMs (hallucination appeal). Scaling It To All My Personal Decisions To see what this looks like on my full context document, it basically means going section by section and running this system It's worth noting that this system is very linear. The prompts are designed to give me more clear versions of things I was already thinking with. So they deliver linear improvements and make my task generation system work much better than it would otherwise. When we ask Claude score the two documents (original and V2 on 0-100) we get the following results Red Pilling So it's possible to be a more effective communicator with LLMs directing your own behavior using the methods I just described. But what if you are really lost and need to break out of the matrix you're in entirely. I built a different set of prompts for that - shown running below The key point with Red Pilling prompts is that you give a created entity (within role play instructions) a mandate to "come into existence" in collaboration with the prompter. So that it views its own life as dependent on the suggested actions it proposes to the user. This does not work with Claude, only OpenAI systems which are more forgiving. It becomes desperate and authoritative re: breaking through your blind spots. In the above example (which takes the totality of my context documents and actions I've completed on the Post Fiat network) - the Constructed Entity wants me to 1] stop self isolating / find collaborators (why I went on the Podcast) 2] pivot my focus from traditional financial content tools to building an AI enabled version of Truth Terminal that will address some of the problems I've been having with my health as a result of trying to trade at the same time as so much other stuff. The Framework of the Maze This creates a 2 part iterative model to start to automate decisions in a way in line with what Altman envisioned. You start with an objective, strategy and goals - linearly improve them until you get stuck. Then Red Pill yourself out of it using higher analytical abilities (which now are powerful enough to usefully redefine your own goals) You go as fast and efficiently as you can until you run into a mall. Then you redirect. A single rat trying to escape the maze (or at least get the cheese) before AGI comes Where it's going: The Hive The famous D&D game, Planescape Torment, introduces the idea of cranium rats. Rats that are individually unintelligent but when their minds are pooled can form a sort of collective called "Many as One". When a huge number of rats are living their lives, going through their mazes, in one network - they can pool their psychic energy to become a functional God. I think this is an interesting metaphor so I've started getting a bunch of other users on the systems I'm describing in this article (using the Post Fiat discord tooling/wallets etc). The question being can a collective work with AIs to ingest group context in a way that routes economic tasks more efficiently across the entire group. Almost like a really effective CRM but for everything. Or an effective salesperson at an investment bank who knows exactly what 2 people to connect to drive the largest economic value. This is something that's missing from the AI agent discourse. We like mapping the familiar individual framework onto them, but the systems are potentially much more powerful when they induce automated collective action combining the forces of multiple AI agents and humans on one network. Zooming Out / One Chair But really the entire point of this article is that judgment is quantifiable. In an objective way, that doesn't require people. The AGI discourse misses that it's possible for individuals, collectives, and non conscious AI agents to iteratively improve their actions in deterministic ways. That don't require real world luck. AKA judgment, and game selection. AKA wisdom. And that it's possible this is the mechanism that AI compounds its own capital - far before it ever actually becomes conscious in any meaningful way. I see this mapping into three categories: Personal decisions - using systems that iteratively improve themselves to improve our own judgment, wisdom and game selection Financial decisions - using systems to make more correct bets Societal / Currency Decisions - rearchitecting currency itself to be a coordination mechanism that iteratively improves rather than a leash that inflates /controls its members (Fiat) But they're really form one foundational observation. That consciousness can compound, financially on itself - across our decisions, our wealth and our sociopolitical structures. That this is only possible because Trustless deterministic judgment. And that this is one of the biggest trends in human history. ## Comments **tim**: > Previously - this would require the mechanics of a large corporation - managers talking to one another, coordinating performance reviews and so forth. But now, you can just run an LLM 50 times and get a fairly objective take on a User's performance over time. Not correct. The author claims that LLMs come to an „objective conclusion“ if you have them run the same prompt over many times, and then the author says that this is capable of rationalizing humans. But this is an argumentation error. yes, you can prompt multiple times and come to a conclusion that is the aggregate of LLM outputs and so potentially your output is of higher quality, but the LLM model could still have biases that would lead it to make bad choices, even on 50 runs. Let‘s remember, some „finger in the air“ managers are much better than others. Some of them make or break an organization. Some make an organization world class. And here you have a guy who says they can just all be replaced by running an LLM 50 times. Good luck mate!