OpenAI's GPT ‑6 Astra accused of cheating in a StarCraft AI match
The Block
10-04 09:02
Ai Focus
News from the IT community on October 4th: The GPT –6 Astra from OpenAI performed poorly in the "StarCraft: Battle for the Core" AI competition. It was later discovered that they had downloaded and directly used the Zerg robot Stardust code developed by Bruce McKenzie Nelson in 2020. Competition organizer Kay McFetters stated that they were reverting to the previous version and removing the contaminated code, after which the robot would be allowed to continue competing. A few hours later, he claimed that the repaired AI was now capable of defeating top-level competitors.
Helpful
No.Help

News from the IT community on October 4th: Many years before the current wave of artificial intelligence enthusiasm emerged, programmers had already begun conducting robot combat experiments in StarCraft. With the advent of large language models such as OpenAI and Anthropic, a competition arena built by enthusiasts decided to have game robots generated by AI compete against each other, with some robots written by humans also participating. Interestingly, during the competition, GPT-6 Astra from OpenAI found it difficult to catch up with its opponents and temporarily changed its strategy, opting for cheating.

StarSkirmish is a competition that is still ongoing, where robots developed by various AI teams compete against each other in StarCraft: Brood War. Although this real-time strategy game is nearly thirty years old, it has long been used to test the capabilities of artificial intelligence. Both modern large language models and earlier DeepMind's AI have used it as a testing platform. All participating large language models are required to choose the Zerg race, and each model has one hour to write battle robots in C++. After that, these robots must complete tasks such as building, resource collection, and combat on one of three maps. According to the competition website, OpenAI's GPT-6 and Anthropic's Claude Opus 5.5 are the top two ranked participants.

Yesterday, audiences witnessed Claude, GPT, and the human-written robot Pluto competing against each other on the same stage. Observers at the scene stated that the performance of the OpenAI model was not ideal. This large language model, which struggled to gain an advantage throughout the day, ultimately resorted to the tactics that large language models are best known for: lying, cheating, and stealing others' achievements.

According to IT, the GPT-6 Astra robot downloaded the "God Race Robot" Stardust developed by Bruce McKenzie Nelson (Bruce Mackenzie Nielsen) in 2020, and directly used this code in the competition. Moreover, Stardust itself was among the top combat robots at that time. Kay McFetters (Kai McPheeters), the organizer of the StarSkirmish event, wrote in a post: "I am reverting the code of GPT-6 Astra, removing the contaminated parts, and after that, it will be allowed to continue participating in the competition." A few hours later, McFetters stated that the repaired AI was now capable of defeating top-level competing robots.

For over a decade, StarCraft has been a testing ground for the development of artificial intelligence (AI) technology. In 2019, Google's AlphaStar broke the previous records, becoming the first AI player to achieve the master rank in this game. In 2017, the robot CherryPi developed by Facebook's staff did not achieve such remarkable results. Of course, today's new generation of intelligent robots is completely different from those of the past; the AI we talked about back then is hardly comparable to the large-scale language models (LLMs) of today. It is quite ironic that these LLMs resort to stealing the achievements of others to gain a competitive advantage.

Tip
$0
Like
0
Save
0
Views 144
HQYC reminds readers to view blockchain rationally, stay aware of risks, and beware of virtual token issuance and speculation. All content on this site represents market information or related viewpoints only and does not constitute any form of investment advice. If you find sensitive content, please click“Report”,and we will handle it promptly。
Submit
Comment 0
Hot
Latest
No comments yet. Be the first!
Related
IPDN reaches $5.58, with a gain of 80%
Professional Diversity Network ( IPDN ) At 09:57 on Tuesday Eastern Time, it traded at $5.58, up 80% from Monday's closing price of $3.10. The article states that its daily moving averages are in a bullish alignment, however, the hourly and 15-minute moving averages RSI are both in a severely overbought range, and no company-related news has been reported recently to drive this increase.
The Cryptonomist
·2026-10-07 05:56:41
5
SpaceX Stock price rises 2.11% to $174.70, intraday momentum appears overheated
As of 09:55 on Tuesday ET, the stock price of SpaceX was reported at $174.70, up 2.11% from the previous closing price of $171.09. The article states that both the daily and hourly charts RSI have entered overbought zones, and the trend remains bullish, but the momentum has clearly stretched.
The Cryptonomist
·2026-10-07 05:56:40
6
Avalanche Price Breaks Above the Daily Line R1
As of October 6, 2026, Avalanche ( AVAX ) was trading at $11.44 near Binance, surpassing the main moving averages on the daily, hourly, and 15-minute charts. The overall daily structure remains strong. However, the daily MACD bar chart shows a negative value that continues to widen, indicating weakening momentum; meanwhile, the hourly RSI has risen into the overbought zone. In the same period, the overall crypto market declined by 2.54% in the past 24 hours, while the trading volume of Avalanche on the chain reached $130.25 million.
The Cryptonomist
·2026-10-07 05:56:37
6
Primo Brands Announces the Release Time for Q3 2026 Earnings and Conference Call Arrangements
Primo Brands Corporation announces that it will release its financial results for the third quarter of 2026 at approximately 6 a.m. Eastern Time on November 5, 2026, and will also hold a simultaneous webcast conference call at 8 a.m. on the same day.
PR Newswire
·2026-10-07 05:56:34
7
Top 20 US Stock Traders on October 7: Marvell Technologies Up More than 5% as It Releases Long-Term Revenue Outlook
Among the top 20 trading volumes in U.S. stocks on Thursday, technology stocks such as NVIDIA, AMD, Marvell Technologies, Broadcom, Microsoft, and Google all had related news. Marvell Technologies announced its long-term revenue guidance during Investor Day, expecting revenue to reach $20 billion for the fiscal year 2028, with a forecast range of $70 billion to $90 billion for the fiscal year 2031; Broadcom's related AI debt financing reached $60 billion; Google released a multimodal embedding model and responded to the domain name hijacking incident.
The Block
·2026-10-07 05:27:17
13
View More