I Gave GPT 6 Astra $10,000 to Trade Stocks...And This Happened

Item: I Gave GPT 6 Astra $10,000 to Trade Stocks...And This Happened
Rating: 3
Author: PostSingularity Institute

by Nate Herk · View the channel on PSI

Formal & Physical Sciences AI — News & Applications UYComputer science UYQArtificial intelligence (AI)

🎙 Nate Herk 👥 1.1M 📅 September 28, 2026 ⏱ 18 min 👁 118K 📄 original study 🧭 2026-10-06

Available in: English (current) Français

Keywords

CodexAlpaca APIGPT-6 AstraS&P 500real money

Summary

In this video, Nate Herk sets up an experiment where he gives OpenAI’s Codex (running GPT-6 Astra) $10,000 of real money to trade stocks for seven trading days, with the goal of beating the S&P 500. He outlines the setup using Codex routines, handoff messages, and a Grok bot that executes trades via the Alpaca API. For the first two days the AI is very passive, so he repeatedly adjusts the strategy to make it more aggressive, increasing position sizes and lowering risk filters. On day four, the portfolio drops significantly, and on day six a Federal Reserve rate hike causes further losses. The final results show the account ends at $9,900, while a $10,000 investment in the S&P would have been $9,980, so the AI underperforms by about 0.8%. Despite the loss, Nate decides to continue the experiment for a longer period and offers a free VIP ticket to his event as promised. The video highlights the challenges of using AI for active trading, especially over short horizons, and the need for human intervention and strategy adjustments.

178 words

Critical Evaluation

Value of the Information & Strength of the Argument

The video provides a practical, real-world demonstration of an AI agent trading with real capital, which is valuable for illustrating the current limitations and behaviors of autonomous trading systems. The argumentation is largely narrative and anecdotal, based on a single trial with a very short duration and heavy human input. The creator transparently shows daily results and openly discusses the AI’s passivity and the need to override its cautious behavior. However, the lack of a controlled methodology, the arbitrary rule changes, and the absence of statistical significance weaken the argument that this experiment yields any generalizable conclusion about AI trading efficacy. The discussion of the S&P comparison is clear, but the conclusion drawn—that more time is needed—is speculative and not grounded in the evidence presented.

Scientific Rigor, Source Quality, Title Accuracy

Scientifically, the experiment is not rigorous: no controls, no replication, and the human interventions confound the AI’s autonomous performance. The sources cited in the description are primarily commercial (software tools, personal social media, and promotional communities) and do not support any scientific claims. The title is accurate but slightly sensationalized; it does not misrepresent the act of giving AI money to trade, though the expected drama of ’this happened’ is undercut by the modest loss. There are no external sources to verify the screenshots or the actual trades, so the reliability of the data rests solely on the creator’s reporting.

240 words

Title / Content Match

The title accurately reflects the content: giving an AI agent $10,000 to trade stocks and showing the outcomes. It is slightly clickbait as the results are modest, but no false claim is made.

Quality & Reliability

6/10

The experiment uses real money and transparently reports daily results, but the methodology is weak: only 7 trading days, frequent human interventions, rule changes, and a non-independent setup. The creator acknowledges these limitations, and the final outcome (a loss and underperformance vs. S&P 500) is presented honestly.

Chapters

Cited Sources

AI Automation Society Plus (Paid Community) — Mentioned in description as how he built his AI agency.
AI Automation Society (Free Community) — Free resources and community for the giveaway.
Glaido (Voice-to-text tool) — Promoted as a free month offer, not related to experiment.
Hostinger VPS — Sponsor mention with discount code.
Nate Herk on LinkedIn — Personal profile for connection.

Contribution & Novelties

The video’s main novelty is the public, real-money test of an AI-driven trading system over a short period, highlighting the gap between theoretical capability and practical performance. It provides a concrete example of how an agent can be configured with routines and handoffs, and it shows the need for human oversight when the AI becomes too risk-averse or too aggressive. However, the absence of a control group and the small sample size limit its scientific contribution. The creator’s willingness to share failures is commendable, but the experiment does not establish any reliable edge.

Pour aller plus loin :

Algorithmic trading — Overview of algorithmic trading and its challenges.
Stock market prediction — Discussion of methods and limitations in forecasting.
Artificial intelligence in finance — Context for AI applications in trading and risk management.

133 words

Radar Profile

The radar profile shows moderate quantity of information, slightly lower quality and technicality, and a notably lower global reliability score. This reflects the entertaining but methodologically weak nature of the experiment, where the descriptive richness is counterbalanced by a lack of rigorous scientific framework.

Reliability 5/10

💬 Equilibré. Sur les 30 commentaires analysés, le ton est partagé : le commentaire le plus liké soulève des critiques méthodologiques solides (absence de contrôle, interventions humaines, courte durée), tandis que plusieurs autres expriment un intérêt pour les mises à jour et la transparence. Certains critiques notent que le titre est trompeur et que l'IA n'a pas vraiment agi seule.