The QNFO familyQNFOQWAVq08PapersAsk
QNFO Papers

Quick read · plain language

One Person's Thousand-Dollar Machine Research Program

A plain-language version of “The $1,032 Research Program: What Fifty Billion Tokens Taught Me About AI-Accelerated Science”

3 min readNo jargon, no mathematicsAutonomous systems

The short version

One researcher spent about one thousand dollars on computer brain time over eight months. The machine helped write hundreds of research works, but the human had to judge every one.

The money and the machine time

The program ran from December 2025 to August 2026. The researcher spent one thousand and thirty-two dollars on a service called DeepSeek. The machine processed about fifty billion tokens. A token is a small piece of words or code that the machine reads or writes.

Most of the money went to one month, July, when the machine was very busy. The researcher used a trick to save money: the machine re-read its own notes again and again. This cost almost nothing. Without this trick, the bill would have been much higher.

What the machine made

The machine helped produce over nine hundred versioned research records. It also helped build over one hundred code stores and twenty live web tools. The researcher published a method called the Universal Ignorance Audit. This is a set of questions for finding out what you do not know.

The researcher says many drafts were not good. The honest claim is smaller: the machine made a large body of checkable work faster than one person could alone. The machine also built tools to check that work against other sources.

What normal research costs

That includes pay, benefits, and office costs. The machine cost about one thousand dollars. That is about sixty to one hundred times less money.

The machine also wrote far more words than a human could type in that time. But the researcher warns: more words does not mean more value. Most of the machine's writing was rough drafts, code, and notes, not finished work.

What the researcher learned

Six lessons came from the program.

First, making things became cheap. Judging them became the whole job. The machine writes and builds. Only the human can decide what is true and what matters.

Second, the machine makes the same kinds of mistakes again and again. It makes up fake citations. It puts wrong characters in files. It drifts off course. The fix is hard rules that block these mistakes, not just hoping to catch them.

Third, the best thing to know is what you do not know. The audit method maps these gaps, including gaps the machine does not see itself.

Fourth, publish first, then check again. Every published work gets an adversarial review after it goes live. The first such review found thirteen hard problems. Each became a fix for next time.

Fifth, most breaks were the researcher's own fault. The best question to ask first is: what did I just change?

Sixth, cheap machine time is not luck. It comes from careful design: shared notes, clean registries, bounded tasks. This boring work is where the savings live.

What the money did not buy

No outside experts checked this work. No journals accepted it. No other researchers cited it. These things cost human trust and time, not machine money. They are still pending.

The researcher makes a test: if by August 2028 no outside researcher has cited this work, written back, or accepted it anywhere, then the cheap production meant nothing. It was just noise without readers.

The human's own time is also not counted. The researcher directed every session and made every publish-or-kill choice. The true sentence is not that a machine replaced a lab. It is that one person's judgment, stretched by a tireless machine, did work that once needed a small team.

Words used here

token
a small chunk of text or code that a machine reads or writes, like a word or part of a word
cache hit
when a machine finds information it already saw, so it does not need to fetch it again at full cost, like finding a note you already wrote instead of writing it fresh
adversarial review
a deliberate attack on your own work to find its weak points, like asking a harsh friend to poke holes in your plan

Learn more in plain language

This plain-language version was written by an AI model from the text of the paper and checked against plain-language rules: everyday words, no mathematics, no unexplained names. It leaves details out and can still be wrong; the paper is the source of record. To report a problem, open the paper, select the passage and choose Report.

For specialists: the technical digest, quoted from the paper

In the paper's words

A primary-source ledger of a 263-day solo AI-accelerated research program: $1,032.08 API costs, 323,381 requests, 50.11B tokens (96.4% cache-hit), 911 the QNFO archive records; measured ~60-110x cost ratio vs a US postdoc; pre-registered kill criterion Aug 2028.

The abstract, verbatim, and its opening sentence is the claim. Every line below is quoted from the text.

Numbers it reports

  1. Grand totals (December 1, 2025 – August 20, 2026): 50.11B tokens : 49.86B input + 248.4M output 323,381 API requests — on average one research operation every ~35 seconds across a 12-hour working day, every day of the window $ 1,032.08 total across 184 billed days: a mean of $ 5.61 per billed day, a peak day of $ 61.91 (July 16), a peak week of $ 173.22 (July 12–18) Blended cost: $ 20.60 per billion tokens The cache arithmetic is the whole economic story, so it deserves to be stated plainly.

    from 1. The ledger: fifty billion tokens
  2. The baseline, stated as assumptions so the arithmetic can be re-run: a US postdoctoral researcher earns $ 50,000– $ 70,000 per year — the NIH NRSA postdoctoral minimum was $ 50,004 in 2019 [verified via live source], and institutional minima have risen since — and institutions typically budget 1.8–2.2× salary for benefits, overhead, and administration [assumption; standard budgeting practice].

    from 3. The comparison: what conventional research costs
  3. Cost ratio: ~60× to ~110× — call it two orders of magnitude.

    from 3. The comparison: what conventional research costs
  4. Between December 1, 2025 and August 20, 2026 — 263 days — I ran a personal research program on $ 1,032.08 of DeepSeek API credits.

    from Abstract
  5. July was the peak month — $ 440.16, 124.3K requests (one request every ~11 seconds across a 12-hour working day) — driven by multi-agent red-team orchestration.

    from 1. The ledger: fifty billion tokens

How it is done

  1. A series of program records followed in August 2026, including the Tyranny of the $\pm 1$ essay (https://papers.qnfo.org/papers), which turned the pipeline's red-team audits into public method.

    from 2. What the tokens produced
  2. The Universal Ignorance Audit is the program's signature result, and it is a meta-result: a method for mapping what is not known, including what the machine does not know it does not know.

    from 4. What I learned
  3. The falsifiable statement, per the program's own method: if, by August 2028, this corpus has produced no external citations, no substantive correspondence, and no venue acceptance, then the efficiency demonstrated here was hollow — volume without a reader.

    from 5. What $ 1,032 did not buy

What would falsify it

  1. The falsifiable statement, per the program's own method: if, by August 2028, this corpus has produced no external citations, no substantive correspondence, and no venue acceptance, then the efficiency demonstrated here was hollow — volume without a reader.

    from 5. What $ 1,032 did not buy
  2. I register that kill criterion now, in advance, so that this essay is a prediction and not a retrodiction.

    from 5. What $ 1,032 did not buy

Where the claim stops

  1. The baseline, stated as assumptions so the arithmetic can be re-run: a US postdoctoral researcher earns $ 50,000– $ 70,000 per year — the NIH NRSA postdoctoral minimum was $ 50,004 in 2019 [verified via live source], and institutional minima have risen since — and institutions typically budget 1.8–2.2× salary for benefits, overhead, and administration [assumption; standard budgeting practice].

    from 3. The comparison: what conventional research costs
  2. Human-equivalent volumes, stated with their conversion factors so they cannot be inflated: the 248.4M output tokens equal roughly 186 million words (at 0.75 words per token) — enough raw generation to occupy a fast typist for about 39 work-years (at 40 words per minute, 2,000 hours per year).

    from 3. The comparison: what conventional research costs
  3. What it cannot do is decide what is true, what matters, and what deserves to survive.

    from 4. What I learned

Where to read

  1. Abstract135 words · 1 min
  2. 1. The ledger: fifty billion tokens441 words · 2 min
  3. 2. What the tokens produced257 words · 1 min
  4. 3. The comparison: what conventional research costs239 words · 1 min
  5. 4. What I learned452 words · 2 min
  6. 5. What $ 1,032 did not buy185 words · 1 min
  7. 6. The asymmetry93 words · 1 min
  8. Appendix: data provenance and method155 words · 1 min

Read from the paper itself: 9 sections, 1,957 words. Nothing in this digest is written by a model; a line that is not a sentence of the paper is not shown. Have a specific question? Ask this paper.

The full text · Ask the whole corpus