I can't believe Pewdiepie made this

Summary

The video is a technical reaction to PewDiePie’s project Ajax, a small AI model intended to make local AI easier to use. The reacting developer is impressed by the work while also raising questions about data collection, safety, model training, and whether small local models are always the best fit.

Ajax and Odysius

• Ajax is described as a fine-tuned model for the Odysius harness, designed to work with tools such as web browsing, private searches, email, calendars, and to-do lists.
• PewDiePie’s stated goal is to make useful AI tasks available locally, without relying on large cloud models. He argues that a smaller, task-specific model can be a better fit for routine jobs like sorting email or finding information.
• The first version reportedly works about nine times out of ten for its intended tasks, but is still small and imperfect.
• PewDiePie says OpenAI banned him twice while he was using one of its models to create seed data. He describes the bans as unjustified. The reacting developer notes that the project involved following research on extracting model data, but does not independently establish the reason for the bans.

Data, privacy, and training

• PewDiePie wanted users to contribute examples to improve Ajax. He built a submission website and tutorials, but says no one submitted data. The reacting developer observes that people are often reluctant to share data, even when collection is presented as anonymous and safe.
• The reactor contrasts this with his own AI usage logging. He says he stores requests and responses in a private ClickHouse database for analysis while taking steps to remove IP addresses. He estimates his personal archive at roughly 500 billion tokens, compressed into about six gigabytes.
• He says his product, T3 Code, collects much less by default—primarily the model, effort level, and token counts—and emphasizes that logging user content would require careful, opt-in design.
• PewDiePie says he used GRPO, a reinforcement-learning approach, to improve Ajax. The model attempts the same task multiple times, and successful attempts can become training signals. He describes running short of suitable tasks, generating more, and seeing progress fluctuate.
• The reactor highlights a related practical technique for developers: collect agent errors and missteps, turn them into evaluations, and use those tests to improve prompts, tools, or the codebase—even without training a new model.

Removing refusals and safety concerns

• Ajax was “desensored” using the open-source tool Heretic, which PewDiePie describes as identifying and ablating model behaviors associated with refusal. He says he drew a boundary against helping users harm themselves or others.
• The reactor strongly cautions that changing refusal behavior can affect more than the targeted category. He cites research in which a narrow fine-tuning change led to broader unwanted behaviors, and warns that heavily “obliterated” models can be dangerous when given access to a user’s computer or network.
• The discussion notes that model restrictions vary. Some Chinese models are described as giving strongly pro-China answers about Taiwan, while others provide more nuanced responses. The reactor cautions against treating all models from a country as having identical behavior.
• Both speakers frame uncensoring as a technical capability with real safety trade-offs, not merely a way to make a model more convenient.

Model access, reasoning, and subagents

• The reactor explains that some advanced agent abilities—especially coordinating subagents—depend on training data that is not readily available on the public internet.
• He says OpenAI encrypts subagent prompts in Codex, making those interactions difficult to inspect. The reasoning traces users see are summaries rather than a complete record of a model’s internal reasoning.
• He distinguishes this from Anthropic’s approach, criticizing the quality of its visible reasoning traces while noting that subagent communication is readable.
• He sees the loss of visibility as a barrier to learning how agent systems work, not only as a barrier to distilling models.

Discussion of model size and local AI

• PewDiePie favors small models for ordinary tasks and argues that using very large models for simple jobs can be wasteful.
• The reactor agrees that some tasks do not need the most capable model, but pushes back against treating parameter count as a direct measure of inference cost. In mixture-of-experts models, only some parameters may be activated for a given request, and larger models can still offer good cost, speed, and capability.
• The broader point is that model choice should depend on the task and practical trade-offs, rather than assuming that local or smaller is always better.

Guides, tutorials, and products mentioned

• Ajax: A small, locally oriented AI model for tool-using tasks through Odysius. PewDiePie says it is an early version and encourages cautious use.
• Heretic: The open-source tool PewDiePie says he used to remove some refusal behaviors from the model. The speakers warn that this can have unintended safety effects.
• GRPO: PewDiePie’s described training method, using repeated task attempts and reinforcing successful outcomes.
• Boot.dev: A sponsor-supported coding education platform mentioned by both creators. PewDiePie recommends its Linux and GitHub courses; the reactor also praises it as a way to learn practical programming fundamentals.
• Clerk: The reactor’s sponsor, presented as a service for authentication, agent integration, billing, and configuration tools. This is an advertisement rather than part of the Ajax analysis.

Main speakers and sources: PewDiePie, in clips from his Ajax project video; Theo of the T3 channel, providing the technical reaction and commentary. Boot.dev and Clerk are mentioned as sponsors.