--- title: "GPU Value Scientist: AI agents forecast GPU rents" description: "Four AI agents forecast the price to rent a GPU. Through past.dev, each agent reads what the others found before the forecast date." canonical: https://past.dev/blog/projects/gpu-value-scientist last-updated: 2026-11-18 --- # GPU Value Scientist: AI agents forecast GPU rents Projects, November 17, 2026. The past.dev team. Y Fan built GPU Value Scientist alone at the Iterate AI Research Hackathon. In this project, four AI agents forecast the price to rent a GPU for one hour. Through past.dev, each agent reads what the other agents found. It reads only the findings dated on or before the date of the forecast. ![The Forecast explorer of the GPU Value Scientist site. A list of GPU rental categories is next to a chart for the H100 PCIe 80GB: recorded prices to November 6, then the predicted middle price and a shaded range to November 13.](https://past.dev/blog/projects/gpu-value-scientist/site.jpg) - Built by: Y Fan ([GitHub](https://github.com/yfandev)) - Event: [Iterate AI Research Hackathon](https://iterate.inc/ai-research), Wefunder, San Francisco, November 14 and 15, 2026 - Live site: https://y-fan-development--gpu-value-scientist-demo-site.us-east.modal.run/ - Code on GitHub: https://github.com/yfandev/gpu-value-scientist - Demo video: https://drive.google.com/file/d/1vO5cvs7UFu_0G01CswN27NCISgOL2Y8T/view?usp=sharing ## The project GPU Value Scientist forecasts the price to rent one GPU for one hour. Pacific Machines set this task as a challenge at the hackathon, and it gave the price data. The data has posted prices up to November 6, 2026, for 37 rental categories. A category is one GPU model, from one price source, with one type of rental. For each category, the project predicts a low, a middle and a high price for November 7, 9 and 13. Pacific Machines also gave a baseline forecast, which predicts prices from past price changes. Four agents use the OpenAI model `gpt-4.1-mini` to find a better method than this baseline. The agents work in a loop over 14 forecast dates in the past, one each week, from July 22 to October 21, 2026. At each date, the agents can use only the prices and the sources up to that date. They choose from a fixed list of 8 methods, and they cannot write code. The 37 categories are in 4 rental groups. For example, the marketplace group has the advertised prices on GPU marketplaces. At each forecast date, each agent does one job: | Agent | Job at each forecast date | |---|---| | Researcher | Searches the web with the Exa search API for sources dated on or before the forecast date | | Analyst | Proposes one method to test for one rental group, and records the results of each method so far | | Critic | Tests each proposal against the baseline on the dates after the proposal, and accepts or rejects it | | Forecaster | For each rental group, chooses the baseline or a method that the Critic accepted | The agents need a shared memory for two reasons. First, each agent needs what the other agents found. For example, the Forecaster can choose a new method only after it reads that the Critic accepted this method. Second, the test is fair only if no agent sees information from after the forecast date. The final forecasts of the project come from one more step after the loop. In this step, Y Fan widened the predicted ranges with one fixed rule. Human feedback started this step, so the agents did not find this rule. After this step, the forecasts had 4.323% less error than the baseline on the same past dates. The 95% interval of this gain is -0.82% to +7.02%, so the gain can be zero. In a later test on the dates November 1 to 3, these forecasts had 1.488% more error than the baseline. ![The research log tab of the GPU Value Scientist site. A table lists each forecasting idea with its test number, its rental group and its result. One row shows Allowed into later forecasts, and the other rows show Not adopted after testing.](https://past.dev/blog/projects/gpu-value-scientist/research-log.jpg) The research log on the live site. Each row is one proposal from the Analyst and the Critic's verdict on it. The agents shared these records through past.dev. ## How the memory works The project calls the past.dev REST API from one class, `SharedMemory`, in [agents.py](https://github.com/yfandev/gpu-value-scientist/blob/main/src/agents.py). Three agents write records, and two agents read them. | Agent | Writes to past.dev | Reads from past.dev | |---|---|---| | Researcher | The sources that it found, with their dates | Nothing | | Analyst | Each proposal, and the results of each method so far | The verdicts and the sources | | Critic | Each verdict: accepted or rejected, with the reason | Nothing | | Forecaster | Nothing | The proposals, the results and the verdicts | Each agent has its own `identity` in past.dev. The project puts the four identities in one `audience`, which is a group of identities that share records. So the Analyst and the Forecaster can read the records of the other agents. ### Each record has the forecast date Each write is one `POST /api/v1/ingest`. The `timestamp` of the record is the forecast date, at 23:59 UTC. For example, when the Critic writes a verdict for the forecast date of September 16, the record has the date September 16. The date of the run itself does not change this date. ### Each read stops at the forecast date Each read is one `POST /api/v1/recall` with the identity of the reader. The request sends the forecast date as `queryTimestamp` and as `occurredTo`. So the response has only the records that have a date on or before the forecast date. This is a simplified version of one write and one read in agents.py: ```python ns = "gvs-91fbdf53b761954a" # the audience of the run # Write one record, dated at the forecast date post("/ingest", { "id": record_id, "content": canonical(record), "timestamp": date + "T23:59:00Z", "identity": ns + "-" + writer, "audience": ns}) # Read only the records dated on or before that date post("/recall", { "query": f"Retrieve {ns} research hypotheses, ...", "identity": ns + "-" + reader, "queryTimestamp": date + "T23:59:59Z", "occurredTo": date + "T23:59:59Z", "limit": 200}) ``` ## What memory changed The repository keeps the log of one live run with real OpenAI calls. In this run, the agents wrote 66 records and made 29 reads. Each write and each read went to past.dev. ### Each read returned records from other agents Each of the 29 reads returned at least one record from another agent. The Analyst got the sources from the Researcher and the verdicts from the Critic. The Forecaster got the proposals and the results from the Analyst, and the verdicts from the Critic. ### One verdict from the Critic gave the Forecaster a new choice The Critic wrote 11 verdicts. It rejected 10 proposals and accepted 1: "Keep today's price, estimate a range from history", for the marketplace group. The Analyst proposed this method at the forecast date of August 26. The Critic tested it on the next dates and accepted it at the forecast date of September 16. The code lets the Forecaster choose a new method only if its read returns an accepted verdict for that method. From September 16 to October 21, the Forecaster's read returned this verdict at each of the 6 forecast dates. Each time, the Forecaster chose the new method for the marketplace group. For the other rental groups, the baseline was the only choice. This new choice did not give a reliable gain. Over the 14 dates, the loop's forecasts had 0.057% less error than the baseline. This result is from before the step that widened the ranges. Y Fan writes that the loop "did not establish a reliable improvement" over the baseline. ### In a lookup test, past.dev was exact but slower than a local lookup Y Fan also tested past.dev on its own, outside the forecasts and without a model. Y Fan asked 18 questions, 3 times each. Each question is about the latest record of one kind, for one rental group, at one date. past.dev gave the exact answer in all 54 trials. A local lookup also gave the exact answer in all 54 trials. This lookup reads a copy of the records on the machine that runs the test. past.dev took 0.82 seconds on average, and the local lookup took 0.0004 seconds. For this exact lookup, Y Fan prefers the local copy, because exact filters on the local records give the answer. The [benchmark notes](https://github.com/yfandev/gpu-value-scientist/blob/main/docs/MEMORY_BENCHMARK_V3.md) give the full test. ## Try GPU Value Scientist 1. Open the [live site](https://y-fan-development--gpu-value-scientist-demo-site.us-east.modal.run/). 2. Select Research log to see each proposal and its verdict. 3. Select Sources & sponsors to see each write to past.dev and each read from it. 4. Read [agents.py](https://github.com/yfandev/gpu-value-scientist/blob/main/src/agents.py) to see each call to past.dev. To give agents a shared memory with dates, start with the [quickstart of the Memory API](/docs/memory-api/quickstart). Thank you to Y Fan for building with past.dev at the hackathon.