
Inside FarmTable AI Lab
We run an AI research lab in North Texas. We’re building our own coding swarm and studying which open models work well together on hardware we own. Come take a look at what’s on the table.
How much can one GPU do?
We start with a question someone with a consumer RTX card can ask: will the model fit, and how long will I wait for an answer? Our DGX Sparks, RTX PRO 6000 Blackwell GPUs and L40S GPUs let us ask the next question. What changes when several agents need answers at once?
How we compare local setups
For a given workload we vary the model, runtime and hardware, then look at the wait for the first token, throughput under load and the quality of the answer. Our setups range from Ollama to TensorFold and Strata. We want useful open models to stay runnable on hardware people control, even as hosted services change.
Qwen3.8 Flash Next is one of the models on the farm. Our reported peak in local testing is about 4,200 tokens per second. The rate depends on the workload and how many requests run together.
Which model should review the code?
FarmTable’s development swarm is software we build here. One agent can write a change and another can examine its work. Running the models ourselves lets us experiment with who builds, who reviews, and how much GPU time each gets.
On the laptop is one change we made so we could study that question. The reviewer used to be fixed in the code. Now we can choose its model independently and keep the builder the same. That gives us a way to test when a stronger reviewer earns the extra compute.
Read the change on the laptop
A task can name its reviewer, inherit the project’s choice, or use the default. Sixteen routing tests passed for this feature. Whether a particular reviewer catches more mistakes is a separate experiment.
for source in (task, plan):
if source and source.get('auditor_model') is not None:
return source['auditor_model']
return DEFAULT_AUDITOR_MODELWhat carries over to a new problem?
Ryan brings that question to his AI PhD research and his Kaggle projects. This entry comes from an ARC Prize experiment. After an agent solved a game, could a short note about what worked help it get unstuck in a different game?
We looked at 61 stalled positions. In none of them did both readers find a suggested lesson that fit the new game, had evidence behind it, and gave the agent something it could do. We shelved that version of the lesson library.
Inside the notebook · September 19, 2026
The study used records from 12 Kaggle runs. The readers assessed 146 candidate lessons from other games. Both found four relevant, but none supported by what the agent could see in the new game. That finding applies to this way of selecting lessons.