Theo on PewDiePie’s Local AI Project, Training Data and Model Safety

Theo introduces a reaction to PewDiePie’s account of building a local AI assistant and a model called Ajax. He jokes about the creator’s supposedly tiny audience before turning to what the project reveals about data, fine-tuning and the practical limits of local models.

PewDiePie introduces Ajax as a fine-tuned model for the Odysius assistant, intended to browse, search, handle email and interact with a calendar. He says OpenAI banned him twice during the project; the video treats the episode humorously and does not establish an independent account of the cause.

PewDiePie says the goal is to make local AI easier to run. Theo is skeptical that all local models are capable enough for coding work, but acknowledges that PewDiePie has substantial hardware and may be taking a more considered approach than some enthusiasts.

PewDiePie says he wanted to train a model that could be improved when users encounter problems. His assistant is designed around privacy and, he says, does not collect user conversations by default. He therefore asked users to volunteer data instead of silently collecting it.

Theo says T3 Code normally records token usage rather than user conversations, with an option to disable even that telemetry. He has considered an opt-in way for users to share failures and histories so the product can be improved, but says building a safe consent flow would take work and is not currently a priority.

For his own work, Theo describes routing inference through a custom proxy and storing his personal requests and responses in a local ClickHouse database. He uses the records to audit his usage and investigate failures. He says the setup is bespoke and warns that similar configurations may carry account risks.

Theo estimates that his personal archive represents roughly 500 billion tokens and distinguishes it from T3 Code’s aggregate usage metrics. He says the product records model, effort level and token counts, not users’ prompts and responses; the reported usage figures are his own claims.

He sees potential value in a personal archive for understanding what works and possibly training another model in the future. He jokes about offering his data to PewDiePie, while noting that it would be biased toward T3 Code and could teach a model to assume that tool is always available.

PewDiePie says he built a submission site and tutorials to make it easier for users to donate training data, but received little response. Theo observes that even when sharing is voluntary and presented as anonymous, many people are reluctant to contribute their data.

PewDiePie discusses using a stronger model to generate training examples for a smaller one and says hidden reasoning made that harder. Theo uses this to discuss what is visible in agent traces and his frustration when internal interactions are not exposed to developers.

Theo argues that examples of one agent delegating to other agents must often be generated deliberately rather than collected from ordinary web pages. He believes some frontier systems are stronger at coordination than many other models; these are his impressions rather than a controlled comparison.

He claims some subagent context in OpenAI’s Codex is encrypted and contrasts that with what he can inspect elsewhere. He also speculates about how reasoning summaries are produced and whether another provider’s traces are deliberately degraded; he explicitly says he does not know the mechanism.

PewDiePie jokes about a paper concerning hidden reasoning, then says he did not do the prohibited activity he just described. He recounts losing access and later being banned again. Theo reacts with mock amazement; the sequence is presented as the speakers’ account, with satire and no independently verified explanation of enforcement.

The conversation moves to model refusals. PewDiePie says he wanted to reduce refusals in his model. Theo argues that behavior differs across models and demonstrates varied responses to a politically sensitive question, pushing back on the idea that every model from one country behaves identically.

Theo warns that changing refusal behavior can have effects beyond the narrow behavior a developer intends to change. He refers to research he cannot name and describes security risks from giving less constrained models access to a local environment. His anecdotes and numerical claims are not independently established in the video.

PewDiePie describes using the open-source Heretic project to alter refusal behavior and says this can damage model quality. He says he drew a personal line around harm to oneself or others and asks people to use the model responsibly. Theo emphasizes that altering one behavior can affect others.

PewDiePie describes Ajax version one as a small model for the Odysius assistant, able to help with inboxes, browsing and notes. He says he chose a smaller target so more people could run it locally, and presents it as an early version that still needs work.

He describes experimenting with GRPO: repeating tasks, reinforcing successful attempts and generating more tasks when useful examples ran out. He says progress improved, flattened and improved again. Theo sees the size and variety of the problem set as important.

Theo suggests a parallel for developers who are not training models. Collect tool failures and instances where an agent took a wrong path, turn them into evaluations, then adjust instructions or the codebase and test whether the environment better guides the agent.

PewDiePie describes repeatedly delaying release while refining, ablating, quantizing and benchmarking. Theo compares model development to both engineering and growing a plant: there is no automatic finish line, and the developer cannot fully control how a model changes during training.

PewDiePie argues that running a huge model for a small task such as checking email can seem wasteful, and favors a small model adapted to a specific assistant. His estimates of the hardware and electricity needs of large models are illustrative claims, not measured results in the video.

Theo partly disagrees. He says larger hosted models can offer good cost and performance for simple work, and explains that mixture-of-experts models may activate only part of their total parameters on a given request. Total model size alone does not determine the compute used by every query.

PewDiePie says he wants local AI to help people find information and save time without replacing creativity or thinking. Theo agrees that locally controlled email sorting is appealing, while joking about the priorities people assign to that convenience.

PewDiePie thanks the Odysius maintainers. Theo compares the project’s GitHub stars with T3 Code and jokes that a change in how GitHub displays repositories would be his test for artificial general intelligence. The comparison is presented as humor, not a benchmark.

PewDiePie closes by arguing that AI should not replace learning and basic understanding. Theo skips the remaining rapid-fire questions, praises the technical ambition behind the project and closes his reaction.