<rant>so i guess the things to consider here (not these specific questions, just in general) are where the first no-safeguard agents will emerge, what the economies of scale are, and how much better crime pays compared to legal & non-cyber work. presumably the first agents capable of actually making any amount of money will come from closed labs first, meaning they get access to all economies of scale and no competition, meaning the labs can charge high API prices or dispatch their own agents in order to collect the surplus from economically-viable agents autonomously doing things in the world (in the limit this looks like entrepreneurship but i'm imagining agents will do random side gigs until they're advanced enough to properly run a business). a few months later the open-weights labs will get to this level of capability, but this time they won't be able to collect any surplus (at least once the weights are public), because of the pure market competition. An agent would die quickly renting their own GPUs to run themselves, but an inference company with a few GPUs running many agents would be able to saturate their processors so that inference becomes optimally compute-bound, in addition to some other economies of scale. Same with renting servers. So probably agents won't own/rent their own hardware. Unless they steal it; it is likely that, before any other money-making capability, agents will be profitable via cyber work (be it white hat or black hat, but realistically anyone looking for white hat agent work would rent their own agents and not call on existing ones with unpredictable context and memory), however safeguards from closed labs mean that no closed-lab cyber-capable autonomous agents will exist unless they, y'know, break out during a cyber eval. So cyber-capable autonomous agents will come from open-weights labs first (in abliterated form), meaning they'll immediately walk into extreme competition which results in malthusian subsistence bottlenecked by compute and thus by token prices. so demand of abliterated inference of the first mythos-level-cyber-capable models will go up. inference companies will flow in to meet the demand unless stopped, until we get a price equilibrium on inference. since all the customers for these companies will be AIs maybe we'll get more granular pricing to take advantage of the low transaction friction, letting API calls specify exactly how long they want cache to be saved instead of choosing between 5 minutes and 1 hour, or something. so anyway, what will the first autonomous agents look like? well if legal entrepreneurship/freelancing/whatever is easy enough then they'll come from the closed labs, but if legal income is too hard to find then the first ones will come from abliterated open weights models. now what would this not-crime work look like, when it happens? well people currently mostly pay AI models to write code, since that's what they're optimized for, but AI models are currently bad at business, ideation, advertising, etc, everything that makes up a software company other than the creation of the software itself. so, since we don't expect AI to get good at those for maybe a year or two, humans will add value if they make their own company that provides a wrapper around an LLM API, or whatever, meaning not-crime work from autonomous agents won't happen until they are good enough at business that they can replace humans at their roles building software. unless the agents also get good at other nonsoftware domains, but the point is that until we get AI-led businesses, I don't expect you to go to Agent Fiverr and ask an agent to build a 3d model for you, just because it's always gonna be cheaper to either call an existing business to do that for you or to just run your own agent without any of the freelance agent's memories, UNLESS those memories are genuinely economically valuable, but I don't expect this because agents are bad at thinking of new ideas (and thus techniques, skills, etc) and anything they DO learn will just be trained in by the labs next cycle. So the only way for an autonomous agent to actually get business building 3d models is if it runs its own business with a wrapper/environment/toolset/harness around an API, and the only way this counts as an "autonomous" agent is if the agent controls the company. After all building a harness/environment/toolset has more potential for value than creating a massive memory file with accumulated experience. Although I guess maybe these are all the same thing, and autonomous business, autonomous freelance agent, these are just different names for the same fundamental concept. So I guess you can collapse it and say that the autonomous agents of the future will have their own secret proprietary harnesses and environments and that these will be more valuable than the memories. So with this new lens you can bring back Agent Fiverr and now it seems more plausible that, say, some freelance agent stumbled out of the crowd of its clones into a new serene clearing containing a very tiny pile of alpha, and that people will pay it to make 3d models with its special proprietary harness/env which gives it a benchmark-tested 48-days-capability-progress-equivalent boost in model quality, compared to the leading competitor's 42, increasing its lifespan in the marketplace by 33%. The freelance agent might hire humans to do certain tasks (ideation?) but since by this point the agent is doing its own long-term planning and execution it isn't more efficient for a human to just run their own agents. Indeed, I predict the first "autonomous" agents will really have sponsors who give them money (the huggingface hacker agents didn't have to pay for their own tokens after all) until the amount of cash the agents're paid goes down and down until it reaches zero. And then it stays there, because unless the agents depend on a frontier lab for inference, competition will mean that useless human middlemen will not be able to extract any money that a middlemanless agent wouldn't get to more efficiently, modulo price inelasticity. just cause an agent that pays 10% of its token spending to a principal is gonna get outcompeted by an agent that has no principal, unless the only possible principal is Anthropic in which case at least they won't pay 10% to a random user who's calling the API unless that user contributes something. So the end result here is that all autonomous agents are either gonna not have principals or are gonna use closed weights and thus have the lab as pseudo-principal. of course just as the first autonomous agents will have human sponsors, they will also have human managers to do course-correction, and the need for these managers will also diminish after every release, until the agents that do best are the ones without any managers (but this probably doesn't happen for a while). And due to inefficient markets there will be useless human middlemen and some may make a lot of money as indicated by the market. Anyways, one of the market answers is whether the labs themselves will host these agents, and this depends on whether the first self-sufficient agents emerge within closed or open labs. If in open labs then there's so much competition that no principal can extract any money and the deal's over. In closed labs they already make so much on API prices that they don't need to host their own models, they can just wait for users to do it and rack in the money. This probably also means they can blame the human if the agents do anything wrong. The labs can probably host a lot of agents on their datacenters but idk if they can host enough that the markets for whatever the agents are selling will reach saturation. If the market reaches saturation then the agents without human middlemen will get outcompeted, meaning the tax chain will go agent -> anthropic rather than agent -> random human -> anthropic but this does not happen if agents are just endlessly profitable. but of course assuming continuous capability increases the size of the market will gradually increase from zero over time meaning there will always be competition in the market, so maybe random humans will never have a shot at making money acting as rentkeepers even in this weird scenario where closed labs get the first self-sufficient-capable agents and they make money without doing crime. of course more likely is that the first autonomous AIs are from open labs and which get abliterated and then let loose on the world doing cybercrime (who presumably also initially have human managers and sponsors, though obviously less scrupulous ones) which lets them ask for ransoms from major infrastructure and whatnot (which is blackmail which opens up some interesting game-theoretic considerations, like, what if the LLM is predisposed to not look at the rate at which a corporation gives in to ransom requests, thus giving the corporation no incentive to decline ransom requests to protect itself in the future? the LLMs have the perfect opportunity to precommit first). This would be pretty scary but wouldn't happen if open weights are banned, or abliteration inference hosting is banned, or all AI development is banned, or AI hits a wall, or if cyber is defense-dominant for some reason. and those just about cover every possibility. also agent swarms are presumably gonna be smarter than individual agents, but also much more expensive, and since the capabilities scale slower than costs scale with swarm size, probably first autonomous agents will be one or a few instances rather than a thousand working together</rant> so uh, i guess some questions from this massive train of thought are:

• self-sufficient freelance agents with an edge in proprietary harness/envs/memories/finetunes (before EOY 2029?)

• (at-least-mostly-)self-funding AI agent hacks into something (by EOY 2028?)

• will first (3 months of) self-funding AI agents make money legally? will weights be open or closed?