OpenAI has banned PewDiePie twice; they are actively developing a censorship-free AI Ajax that can run on personal computers.
Decrypt
6h ago
Ai Focus
PewDiePie stated in a new video that OpenAI banned his account twice during the period when he was building Ajax. Ajax is a small AI model designed to run on local computers, based on Alibaba's Qwen 3.5. PewDiePie claimed that he used GPT-5.6 and Sol to generate training data, and with the help of open-source tool Heretic, he removed the model's ability to refuse responses.
Helpful
No.Help

According to YouTuber PewDiePie, OpenAI banned his account twice during the period when he was building Ajax. Ajax is a small AI model designed to run on your own computer and browse web pages for you. He shared this experience in a new video he posted on the YouTube channel.

Ajax is what he refers to as "fine-tuning the AI model." This means that, based on an existing model, additional training is conducted for a specific task. In this case, that task is carried out within his free, self-hosted AI application Odysseus, which was launched in June.

Running such models on your own machine allows you to keep emails and calendars locally, rather than on the company's servers. There is also no subscription fee, but the computational load will fall on your own hardware.

It is claimed that Ajax is a model with 9 billion parameters, built upon Alibaba's open-source Qwen 3.5. Parameters are adjustable values that constitute the "intelligence" of AI; generally, the more parameters there are, the stronger the model is considered to be. This model is still awaiting release at present, but PewDiePie indicates that once they feel it is ready, they will make it available for download on their own website.

Two bans

PewDiePie said that he originally wanted to distill from the flagship model GPT-5.6 Sol released by OpenAI on July 9th. Distillation refers to using the answers of a larger model to train a smaller AI, which is somewhat similar to a student copying the homework of a top student in the class.

What he values is the token reasoning of Sol, which refers to the text fragments that the model writes down on a hidden draft paper when solving problems. OpenAI will then encrypt this draft paper.

In the video, he said that he had found a study that could be used to crack this mechanism through OpenAI his own API. He also stated that he would never truly “crack” OpenAI, but OpenAI still banned his account. He claimed that there was no legitimate reason for this ban and raised objections, after which his account was re-enabled at some point.

Subsequently, he said that he used Sol again to generate seed data, which is the initial example used for model learning, and then OpenAI banned him for the second time. The terms of use for OpenAI prohibit its output for the development of competitive models, so he was banned once more.

In August, researchers demonstrated that encrypted inference blocks from OpenAI, Anthropic, and Google could be replayed to a weaker “sister model”, allowing the latter to write them out in plaintext. They decoded 315,320 such fragments from public code repositories.

On September 30, OpenAI stated that it thwarted an attempt by individuals associated with Moonshot AI to extract the reasoning content hidden within its model. The company claimed that it had blocked a pathway that would have allowed someone to replay another user's encrypted reasoning and restore that content.

Remove "reject" from AI.

In order to make Ajax become "unreviewed" – that is, to respond to requests that would otherwise be rejected by assistants of the ChatGPT type – PewDiePie used Heretic, an open-source tool designed to remove the built-in rejection mechanisms of models. This tool compares how the model responds to harmful and harmless prompts, identifies the patterns behind those rejections, and then removes them.

This practice is known as “abliteration”, which is essentially a technique that causes the model to “lose its judgment ability”. In this way, it does not need to bypass restrictions through jailbreaking, as it will always agree to carry out what is requested of it.

This “surgery” is not perfect. PewDiePie says that Ajax has “some brain damage” as a result, but it still has a success rate of about 90% when it comes to tasks such as organizing the inbox and browsing web pages.

He is still using GRPO to continue training it. This is a method that allows the model to attempt the same task 16 times and learn from the successful attempts.

PewDiePie said that he would choose on his own which rejections to remove, setting the boundary at 'causing harm to others or causing harm to himself.' He stated that his lawyer advised him to clarify that Ajax is not designed to provide dangerous or actionable instructions.

Small model, big selling point

The reasons he uses to defend small models are cost and energy consumption. According to his own rough estimates, a cutting-edge model with allegedly trillions of parameters would require about 27 of his computers, as well as an amount of electricity equivalent to that consumed by around 150 households.

PewDiePie refers to Ajax as the first version. He said that he still needs to re-execute the process of rejecting responses once, and then compress (quantize) the model and conduct benchmark tests to confirm that there are no issues.

Ajax is currently not available for download. Initially, there was a countdown on the download page pointing to October 3rd at 08:25 Japanese time. Now, that timer is no longer visible.

Tip
$0
Like
1
Save
1
Views 35
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
700,000 tons of copper piled up in American warehouses, yet prices remain high
In September, copper prices approached $6.90 per pound, but then fell back to around $6.70 due to the potential delay in U.S. tariff adjustments. CME indicates that tariff uncertainties have driven physical copper to be shipped to the United States, increasing domestic inventories and regional premiums; at the same time, AI infrastructure and energy transitions are also supporting copper demand.
Coinpaper
·2026-10-04 05:23:21
23
Seagate Technology shares fall by over 10% before the market opens; Toshiba announces capacity expansion HDD AI Storage supply and demand pattern changes
According to Nikkei, Toshiba plans to invest approximately 60 billion yen to expand its factory in the Philippines at HDD, and aims to double the hard drive production capacity for AI data centers by the fiscal year 2027. The report indicates that this move has prompted the market to re-evaluate the supply and demand dynamics of the AI data center storage market, with Seagate Technology's stock price falling by more than 10% before the market opening in the US.
Wallstreetcn
·2026-10-04 04:51:23
24
Bessent compares tech executives regulating AI to Hannibal Lecter: 'Stop me from killing again'
U.S. Treasury Secretary Scott Bosworth criticized tech leaders who advocate for restrictions on AI by the government in The Axios Show, calling such "alarmist rhetoric without solutions" a lack of leadership. He also mentioned that after a meeting with AI executives last month, the White House facilitated voluntary self-regulation commitments, and OpenAI recently put on hold a new model due to internal testing.
Businessinsider
·2026-10-04 04:41:22
27
Leerink Partners Maintains Pfizer's Holding Rating
Leerink Partners maintained a 'Hold' rating for Pfizer in its report released yesterday. The stock closed at $27.80 yesterday; the current consensus among analysts is 'Moderately Buy', with an average target price of $28.63, representing a potential upside of 2.99% from the current price. UBS also maintained a 'Hold' rating for the stock yesterday, with a target price of $27.00.
Businessinsider
·2026-10-04 04:12:00
31
1,400 layoffs at truck factories reveal which sectors the slowdown in U.S. employment is impacting
The article states that although the national employment data in the United States still appears relatively stable, with an unemployment rate of 4.2% and an increase of 9,000 jobs in manufacturing in September, these overall figures may conceal significant regional impacts of unemployment. The report points out that large-scale layoffs at a single factory can outweigh local job creations, and the modest growth in industries such as machinery, plastics, or healthcare in the national statistics may obscure the severe economic effects of the closure or layoffs of individual factories on communities.
Coinpaper
·2026-10-04 02:22:01
36
View More