I, along with several others, have decided to begin preparations to set up an "agent server" for our stock trading business (and other uses). This market will assist us in assessing the timing and most economically valuable tasks to focus on.

THE SERVER

It is our belief that there will be a narrow window of 6-8 months where there will be the opportunity to make an extraordinary amount of money. This time period will be limited between when the first AGI model that is powerful enough is released, and when a company like OpenAI creates a nice user interface and allows anyone to use agents at a low cost to pump out things that compete with us.

I am going to budget a minimum of $20,000 on the first server. It will consist of 5-8 4090 GPUs - 4-6 to run the AGI model and 2-4 additional GPUs for training other models. My opinion has always been that the task models are most capable at is model design, and model design is a self-contained task with a clear goal (optimize AUC) that doesn't require user input and doesn't require a GUI. Therefore, the primary goal of the agent server will be to automate machine learning research. The models it improves will be used to earn money trading stock options and other assets.

I have found that o1 already is successful at generating novel machine learning ideas for us, so the model knowledge of machine learning research already exists even now.

TASKS

When the agent server is waiting for the other GPUs to finish training a model iteration, we will also put it to work on other tasks. Here are the tasks we have envisioned for it at first:

• Improve the AUC of our stock options models (the models have earned $60,000 between January 2 and 23, 2025)

• Eliminate costly bugs from the code that executes the stock options models' predictions

• Write and debug Python code to recover the data from a RAID array that failed in 2023 and which we don't know the disk order and other parameters for (all disks and data are present; only the metadata is lost)

• Eliminate all PHP, for which packages are not being updated, from the PROHASHING backend and replace it with equivalent secure Python code (this code is ultra-critical, hasn't been changed since 2014 and has handled $150 million in transactions, and is thought to be impossible for a human to do without catastrophic losses)

• When given music, automate tasks like transposing keys and duplicating notes so that I can begin to sell very high quality music that isn't profitable to spend the time doing alone

• Do our taxes, which required 450 hours of labor in 2024

• Generate responses to the defendants' motions in our existing cryptocurrency scam litigation, defend myself against Wells Fargo (if necessary), and file two additional cases against new defendants responsible for our losses in the BlockFi and Celsius cryptocurrency scams

• Self-improve its own interface, so that the humans can more quickly and easily tell it what tasks to do

Once we earn money from these and other activities, we will use it to buy more hardware to run multiple servers in parallel, allowing us to train new models more quickly.

SETUP

To set up the server, we need to:

• Run additional electrical wires and install an additional box (this task has already begun)

• Research and purchase 4-8 GPUs, motherboard, 64-core processor, 1TB memory

• Assemble and test these components

• As this server will come online during the summer, determine how to deal with the heat

We plan to order the components when OpenAI's o3 is fully released to any Pro users in the United States willing to pay for it. Our reasoning is that we think that is a good balance between allowing parts to depreciate and being ready for an open source model of the required caliber to appear a month later. Right now, we are targeting the server to come online during April.

MARKET

The "useful task" questions will resolve to the date when the agent server is able to accomplish a meaningful part of a task that a human would previously have had to do. Examples would be when it trains a stock trading model that has a higher AUC or recall, when it can organize massive amounts of data from forms and create a database of our trades for taxes, when it can output a full correct complaint with a single-digit number of errors, etc. If it fails on the first attempt or needs a more powerful AGI model, the questions will remain open until it succeeds or we abandon it.

If it is decided not to bring the agent server online for whatever reason, all the questions in the entire market will resolve to NO. If the server comes online and it is determined that the setup will never work and the computing power will instead be used for manual training of models, all questions will resolve to NO. The market will not be left open for any server brought online in the future. It is for the first server only; a gap between abandoning the first server and trying again four months later doesn't count.

All remaining questions will resolve to NO when the market expires.

While the server will be extensively used for litigation as described above, any task or earnings related to litigation is specifically excluded from this market and counts towards none of the questions, and no comments will be offered in response to litigation activities.