Moon Dev credits an OpenAI "co-founder" with a loop that "just changed trading"; Karpathy calls himself a founding member, and the loop he shipped 195 days earlier optimizes LLM validation loss on a GPU

A 17.136009 second MP4 (QuickTime/MOV container, mov,mp4,m4a,3gp,3g2,mj2), 720x1280 portrait, H.264 High profile level 3.1 in yuv420p at a constant 30/1 fps across 514 video frames at 4,016,631 bps, with HE-AACv2 audio at 44,100 Hz stereo and 64,693 bps across 373 frames, 8,755,482 bytes total at an overall 4,087,524 bps, yt-dlp format id h264_720p_4087288-1; posted at epoch 1789740000, which is 2026-09-18 14:00:00 UTC, by handle moondevonyt (uploader id 7095083653357978666, nickname MoonDevOnYT) over the credit "original sound - MoonDevOnYT" with a 407 character description; at capture time (metadata epoch 1792408708, 2026-10-19 11:18:28 UTC) the counters read 7,736 views, 304 likes, 12 comments, 63 reposts and 247 saves. I read thirteen frames: the full 2 second grid frame_001 through frame_009 at 0, 2, 4, 6, 8, 10, 12, 14 and 16 seconds, plus frame_004, frame_012, frame_022 and frame_030 from a denser 0.5 second re-extraction (34 frames at fps=2) because six of the nine original frames are motion blurred. The denser pass confirms there are no cuts: the entire clip is one continuous handheld phone pan across a single desktop, so no card or label was dropped between samples. What the sharp frames show is a browser with tabs reading "Moon Dev AI Trading Bo...", "iphone-labor-day - Goog...", "Launch a token · pons", "Launch | Moon Dev", "Robinhood Chain: addre..." and "CHILLGUY/USDT_P 0.011", a bar reading "LEADER SOBLIZZY (95) RANK #2 STREAK 0 YESTERDAY 3", and three terminal tabs named "moon-dev-tradin...". The dominant pane (frames 005 and 007, 8 s and 12 s) is a coding agent transcript that opens "I'll start by reading the current strategy and the" and then logs timestamped shell calls: at 09:25:18 ls -la && echo "=== STRATEGY === continuing into ho "=== RESULTS TAIL ===" && tail -30 results.tsv &&, at 09:25:37 cat baseline.json; echo "=== harness head ===" && sed -n '1,120p' ha, and then eleven consecutive running: /opt/anaconda3/envs/tflow/bin/python heredocs beginning import pickle, numpy as np, pandas as pd, pandas_ta as ta at 09:25:42, 09:25:46, 09:26:00, 09:26:17, 09:26:45, 09:27:04, 09:27:32, 09:28:01, 09:28:32, 09:28:51 and 09:29:13, closing on two agent lines: "The whole delay=2/3 row beats delay=1 at every hold from 16" and "That is a plateau, not a fitted cell. Making that change." A second pane tabbed "liqhunt" prints "Trigger: 50.0% of the long book liquidated", "Resolves to 0.1244 (that level wipes $2,714,396 = 56.8%)", "anchor LOCKED at 0.1282 (2.6h old, re-aims in 1.4h)", "rung 1: 10,273 @ 0.12756 ($1,310.42)" and "HUNTING | mark 0.17456 | reachable book $4,904,078 across", stamped 09:23:05 through 09:28:07. Frame 009 at 16 s pans onto a Robinhood-style token page for "ALWAYS / Always Has Been" at address 0xa447..5df28 showing "$0.00 cash", a Buy and a Sell button, a candle chart labeled "Always Has Been · 1 · fomo" and a left rail of rows marked "Graduated", with a second coding-agent pane explaining a stale data/pullback_ws.lock file, ending "Cooked for 31s · done 4:38 PM", "bypass permissions on (shift+tab to cycle)" and "new task? /clear to save 252.3k tokens". The spoken argument is three sentences: a named OpenAI figure invented a loop, the speaker applied that loop to trading systems, and it now runs unattended improving those systems. No profit and loss curve, no equity chart, no out-of-sample or walk-forward split, and no contents of results.tsv appear in any frame I read, and the repository the video is describing is never named on screen. The transcript is 43 words in 3 Whisper segments ending at 16.9 s against a 17.136009 s container and a 17.188571 s audio track, so nothing is timestamped past the end of the audio and no segment is hallucinated.

The name is Andrej Karpathy, and his own site says "founding member," not "co-founder"

At 0.00 s the transcript reads "The co-founder of OpenAI, Andre Karpathy, just changed trading." Two things in that sentence need fixing before anything else can be checked.

The spelling is Andrej Karpathy, with a j. "Andre Karpathy" is a Whisper transcription artifact, not what the speaker said and not how the person spells it. His personal site at karpathy.ai carries the name in full.

The title is the more interesting problem. Karpathy's own biography on karpathy.ai says, verbatim, "I was a research scientist and a founding member at OpenAI," and the site's third-person blurb reads "Andrej Karpathy is an AI researcher and educator. He was a founding member of OpenAI and later the Director of AI at Tesla, where he led the computer vision team of the Autopilot." He does not use the word co-founder about himself anywhere on that page. Wikipedia's article on him follows the same convention: "Karpathy is a founding member of the artificial intelligence research group OpenAI." Fortune, writing about the very loop this video is about, called him "one of the founding employees of OpenAI."

Press usage is split. TechCrunch's May 2026 piece describes him as "the AI researcher who co-founded and formerly worked at OpenAI," so the video is repeating a widely used label rather than inventing one. The label is still looser than the subject's own, and the distinction matters for a video whose persuasive weight rests entirely on the authority of the title. I could not reach OpenAI's own December 2015 introduction post to check how the organization listed him at launch: openai.com/index/introducing-openai/ returned HTTP 403 and the Wayback Machine is not reachable from this tool, so I am not citing either.

One more piece of context the video omits. As of 2026-05-19, 122 days before this video was posted, Karpathy was no longer at OpenAI at all. He announced he had joined Anthropic's pre-training team under team lead Nick Joseph, with an Anthropic spokesperson saying he "will start a team focused on using Claude to accelerate pre-training research." Describing him in September 2026 purely as "the co-founder of OpenAI" is four months out of date on his actual affiliation.

"Just changed trading" describes a repository that went public 195 days before this video

The loop in question is karpathy/autoresearch. The GitHub API reports the repository was created at 2026-03-06T22:00:43Z, with the initial commit b11d6f2 at 2026-03-06T21:58:52Z and the last push at 2026-03-26T00:07:37Z. The README's own epigraph is signed "-@karpathy, March 2026".

This video was posted at 2026-09-18 14:00:00 UTC. That is 195 days after the repository appeared and 176 days after its final push. The word "just" at 3.16 s in the transcript, and the description's "the co founder of openai just handed me the exact loop," both place the release in the recent past. It was not recent. Fortune had already named it "The Karpathy Loop" on 2026-03-17, six months before this video, in a piece reporting that Karpathy "let the AI agent run continuously for two days, during which time it conducted 700 different experiments" and found "20 optimizations that improved the training time."

By the time I checked on 2026-10-19 the repository had 98,178 stars and 13,600 forks. This was not an obscure drop that took half a year to surface. The video's framing of novelty is the one claim in it that the timestamps flatly contradict.

I attempted to read Karpathy's own announcement tweet, linked from the README as x.com/karpathy/status/2029701092347630069, and the fetch returned HTTP 402 Payment Required. I am not citing it and I am not quoting from it.

The loop is real, and Moon Dev ported its filenames, but it was written to minimize validation loss on a GPU

At 4.48 s the video says "In my belief, I went ahead and applied his loop here to building out trading systems." The word "applied" is doing honest work, and the frames back it up more concretely than the audio does.

Here is what autoresearch actually does, from the README: "give an AI agent a small but real LLM training setup and let it experiment autonomously overnight. It modifies the code, trains for 5 minutes, checks if the result improved, keeps or discards, and repeats." The target is spelled out in program.md: "The goal is simple: get the lowest val_bpb," where val_bpb is validation bits per byte on held-out data. The README's requirements section is unambiguous about the hardware: "A single NVIDIA GPU (tested on H100), Python 3.10+, uv." The training code is "a simplified single-GPU implementation of nanochat." The word trading does not appear in the README or in program.md.

The loop body in program.md is nine numbered steps under the heading "LOOP FOREVER": inspect git state, hack an idea into train.py, commit, run uv run train.py > run.log 2>&1, grep the result, handle a crash by reading the stack trace, record the row in results.tsv, advance the branch if the metric improved, and git reset back if it did not.

Now look at what frame 005 catches at 09:25:18 and 09:25:37: the agent runs tail -30 results.tsv and cat baseline.json and sed -n '1,120p' ha, almost certainly a harness file. results.tsv is the exact filename program.md specifies for the experiment log, with its five columns commit, val_bpb, memory_gb, status, description and its keep / discard / crash status values. That is a real structural port, not a vague gesture at one, and it is the strongest evidence in the clip that the video's central claim about method is true even where its claims about timing and authorship are not.

What changed in the port is the whole measurement layer. Karpathy's agent runs uv run train.py against a fixed 5 minute wall-clock budget on an H100 and reads one number out of a log. Moon Dev's agent runs /opt/anaconda3/envs/tflow/bin/python heredocs that import pickle, numpy, pandas and pandas_ta against a pickled dataframe, eleven times in frames 005 and 007. Anaconda on a local machine with pandas-ta is a backtest harness, not a GPU training run. The description's claim that Karpathy "handed me the exact loop that changes trading forever" is wrong on the word exact. He published a loop for one thing, and someone rewrote the inner step for another.

"Sitting here all day in its loop" matches a real instruction in program.md; the frames show three minutes and fifty-five seconds

At 10.20 s the video says "So it's just sitting here all day in its loop, making my systems better and better and better." The "all day" part has a genuine source, and the "better and better" part does not have one in this video.

program.md contains a section headed NEVER STOP that reads, verbatim: "Once the experiment loop has begun (after the initial setup), do NOT pause to ask the human if you should continue. Do NOT ask 'should I keep going?' or 'is this a good stopping point?'. The human might be asleep, or gone from a computer and expects you to continue working indefinitely until you are manually stopped." It goes on: "As an example use case, a user might leave you running while they sleep. If each experiment takes you ~5 minutes then you can run approx 12/hour, for a total of about 100 over the duration of the average human sleep." The description's "i walked away, came back hours later" is describing the documented use case accurately.

The evidence on screen is narrower than the claim. The agent log in frames 005 and 007 runs from 09:25:18 to 09:29:13, which is three minutes and fifty-five seconds, and it contains thirteen tool calls: two inspection commands and eleven python heredocs. One decision is visible at the end of frame 007: "The whole delay=2/3 row beats delay=1 at every hold from 16" and "That is a plateau, not a fitted cell. Making that change." That is one accepted change. The second agent pane in frame 009 reports "Cooked for 31s · done 4:38 PM" and spends its output debugging a stale data/pullback_ws.lock file, which is maintenance, not strategy improvement.

No frame in the clip shows a results.tsv with rows in it, a run count, a metric trending down, or any before-and-after number. The claim that the systems are getting "better and better and better" is not demonstrated anywhere in the 17 seconds. Treat it as unverified on this evidence.

What autoresearch protects, and what the frames never show: a read-only evaluation harness

This is the gap that matters most when the loop moves from language models to trading, and it is worth stating because the video does not.

program.md locks the measurement down in a section headed "What you CANNOT do": "Modify prepare.py. It is read-only. It contains the fixed evaluation, data loading, tokenizer, and training constants," and "Modify the evaluation harness. The evaluate_bpb function in prepare.py is the ground truth metric." The agent may rewrite the model, the optimizer, the batch size and the architecture, and it may not touch the thing that scores it. The metric itself, validation bits per byte, is computed on held-out data the agent never trains on, and program.md notes it is "vocab-size-independent so architectural changes are fairly compared."

That design is what makes a thousand unsupervised experiments safe. The agent cannot reach the scoreboard.

In a trading harness the equivalent protection is a held-out period and a walk-forward split, because an agent that runs hundreds of parameter variations against one fixed price history will find configurations that fit that history and nothing else. Nothing in any frame I read establishes that Moon Dev's harness has that protection. Frame 005 shows the agent reading the harness with sed -n '1,120p' ha, and reading is not modifying, but there is no visible prepare.py equivalent marked read-only, no out-of-sample window, and no baseline-versus-holdout comparison. baseline.json is opened at 09:25:37 and its contents never appear on screen.

The agent's own closing line is the tell. "The whole delay=2/3 row beats delay=1 at every hold from 16" and "That is a plateau, not a fitted cell" is the agent reasoning its way to a robustness check by eye, preferring a broad region of parameter space over a single lucky cell. That is a sensible heuristic to see an agent apply unprompted, and it is strictly weaker than the guarantee Karpathy built, where the scoring function is off limits by rule. On the evidence in these frames, the part of the design that stops the loop fooling itself is the part that did not obviously come along.

Key Takeaways

• Correction: The video says at 0.00 s "the co-founder of OpenAI, Andre Karpathy." The name is Andrej Karpathy, and his own site says "I was a research scientist and a founding member at OpenAI," never co-founder. Wikipedia and Fortune also use founding member and founding employee.
• Correction: The video says at 3.16 s that Karpathy "just changed trading." The repository behind the claim, karpathy/autoresearch, was created 2026-03-06T22:00:43Z and last pushed 2026-03-26T00:07:37Z. This video posted 2026-09-18 14:00:00 UTC, 195 days after creation and 176 days after the last push. Fortune had already named it "The Karpathy Loop" on 2026-03-17.
• Correction: The description claims Karpathy "handed me the exact loop that changes trading forever." The README defines the loop against LLM pretraining: "give an AI agent a small but real LLM training setup," goal "get the lowest val_bpb," requirement "a single NVIDIA GPU (tested on H100)." Neither the README nor program.md mentions trading. The control structure transfers; the loop as published does not.
• Verified: The port is real and specific. Frame 005 at 09:25:18 shows the agent running tail -30 results.tsv, which is the exact log filename program.md specifies, with its documented keep / discard / crash status column.
• Verified: The video's "sitting here all day in its loop" matches a documented instruction. program.md has a NEVER STOP section stating the agent should "continue working indefinitely until you are manually stopped," with an explicit "leave you running while they sleep" use case and an estimate of "approx 100" experiments per night.
• Unverified: "Making my systems better and better and better" at 10.20 s. The longest agent span readable in any frame is 09:25:18 to 09:29:13, three minutes fifty-five seconds, thirteen tool calls and one accepted change. No metric history, no results.tsv contents, no profit and loss and no equity curve appear anywhere in the clip.
• Clarification: The description's "i can't believe it's free" is true. The repository is public and its README ends "## License MIT". There is no LICENSE file in the repository tree, and the GitHub API returns a null license field, so the MIT grant exists only as a README line.
• Context: Karpathy left OpenAI before this video was made. He announced on 2026-05-19, 122 days before the post, that he had joined Anthropic's pre-training team, where per an Anthropic spokesperson he "will start a team focused on using Claude to accelerate pre-training research."
• Context: The guard that makes autoresearch safe to run unattended is that the agent cannot edit its own scorer. program.md forbids modifying prepare.py and states "the evaluate_bpb function in prepare.py is the ground truth metric," and that metric is computed on held-out data. No out-of-sample split, walk-forward window or read-only harness is visible in any frame of this video.
• Context: The trading surfaces on screen are retail crypto, not an institutional stack. Frame 009 shows a token page for "Always Has Been" at 0xa447..5df28 with "$0.00 cash" in the account, a "Graduated" token rail, and tabs reading "Robinhood Chain: addre..." and "CHILLGUY/USDT_P 0.011".
• Unverified: The figures in the "liqhunt" pane, including "that level wipes $2,714,396 = 56.8%", "rung 1: 10,273 @ 0.12756 ($1,310.42)" and "reachable book $4,904,078", come from Moon Dev's own tooling. There is no external source for them and I am not treating them as established.

Resources

• karpathy/autoresearch, GitHub, created 2026-03-06 establishes the repository creation timestamp 2026-03-06T22:00:43Z, the initial commit b11d6f2 at 2026-03-06T21:58:52Z, the last push at 2026-03-26T00:07:37Z, and the 98,178 stars and 13,600 forks I cite as of 2026-10-19.
• autoresearch README.md (raw), GitHub establishes the verbatim loop description, the "single NVIDIA GPU (tested on H100)" requirement, the nanochat lineage, the fixed 5 minute budget, the val_bpb metric, the "-@karpathy, March 2026" signature and the "## License MIT" line.
• autoresearch program.md (raw), GitHub establishes the nine-step "LOOP FOREVER" body, the results.tsv schema with its keep/discard/crash statuses, the NEVER STOP instruction, the "leave you running while they sleep" use case and the rule that prepare.py and evaluate_bpb are read-only ground truth.
• Andrej Karpathy, personal site establishes his own wording, "I was a research scientist and a founding member at OpenAI," which I use to correct the video's "co-founder."
• Andrej Karpathy, Wikipedia establishes the "founding member" phrasing and the 2015 to 2017, then February 2023 to February 2024, OpenAI tenure I cite.
• 'The Karpathy Loop': 700 experiments, 2 days, and a glimpse of where AI is heading, Fortune, 2026-03-17 establishes the name "The Karpathy Loop", the 700 experiments over two days, the 20 found optimizations, the "one of the founding employees of OpenAI" description, and the March publication date that makes the video's "just" wrong.
• OpenAI co-founder Andrej Karpathy joins Anthropic's pre-training team, TechCrunch, 2026-05-19 establishes the 2026-05-19 move to Anthropic, the Nick Joseph reporting line, and the spokesperson quote about accelerating pre-training research with Claude.
• this should honestly not be allowed, TikTok, 2026-09-18 is the artifact under analysis, the source of the transcript timestamps and the 407 character description I quote.
• Not reached, not cited: x.com/karpathy/status/2029701092347630069, the announcement tweet linked from the README, returned HTTP 402; openai.com/index/introducing-openai/ returned HTTP 403; web.archive.org is not fetchable from this environment.