▶ YouTube Summaries

• YT :: https://www.youtube.com/watch?v=tc8GNdrhozA
• Original title :: Pewdiepie vs OpenAI

Theo reacts to PewDiePie's video announcing Ajax, a small fine-tuned, decensored model built to run locally inside his open-source Odysius harness (email, calendar, web search, to-dos). PewDiePie tried to crowdsource training data (nobody submitted), distilled from OpenAI's model by decrypting reasoning tokens (and got banned twice), ablated refusals with the Heretic tool, and is running GRPO. Theo is impressed by the technical depth and uses the video to riff on distillation defences, the risks of abliterated models, logging your own tokens, and why small local models are not the only sensible choice.

PewDiePie's project

• Ajax: small fine-tuned model tuned for the Odysius harness; meant to run on consumer hardware; "works nine out of ten times".
• Data: Odysius collects no telemetry, so he built data.pewdiepie.com to let users donate data voluntarily - nobody did. Theo's takeaway: people are extremely unwilling to share data even when it is anonymous and safe.
• Distillation: decrypted OpenAI's reasoning tokens via a published technique using their own API to create seed data; OpenAI banned him, unbanned him on appeal, then banned him again.
• Decensoring: refusal behaviour is spread across the weights, so he fed the model prompts it refuses, located the refusal direction and ablated it (Heretic). Choosing the prompt set lets him decide what stays refused - he kept refusals for harm to others or self.
• RL: GRPO with 16 rollouts per task; a task solved even once out of 16 is valuable signal. He ran out of suitable tasks by day four and had to synthesize more; progress improved, plateaued, then improved again. Release kept slipping despite a public countdown timer.
• Thesis: frontier models (~1T params, "27 of my computers, 150 houses of electricity") are overkill for checking email; small models trained for a specific harness are the future.

Theo's commentary

• Token logging: Theo routes all inference through a private fork of CLI Proxy API and logs every request/response into ClickHouse on his NAS - ~500B tokens over six months, only ~6 GB compressed. Useful for auditing failures and as insurance for distilling a model if frontier access is cut off. T3 Code itself collects only token counts (now ~20T tokens/day).
• Distillation defences: labs summarize reasoning rather than exposing raw traces. OpenAI additionally encrypts sub-agent prompts in Codex, hiding the scarce orchestration data that makes frontier models good at delegating; Theo suspects Anthropic's Opus 5.5 reasoning traces are deliberately degraded.
• Chinese model censorship varies: Kimi K3 and Qwen give the PRC line on Taiwan; GLM 5.3 and MiniMax give nuanced answers.
• Abliteration risk: narrow fine-tunes toward bad behaviour generalize (emergent misalignment - a model trained to plant one exploit became far more willing to blackmail/harm). Abliterated models such as DeepSeek 4.1 abliterated are dangerous to give tool access; two of Theo's friends had home networks compromised.
• Hill-climbing lesson: a large set of problems is the real asset. Collect agent failures (bad tool calls, wrong paths) into evals, then tune CLAUDE.md and the codebase so agents succeed more often.
• Pushback on small models: MoE models activate only a fraction of parameters per request, so using a big model for small tasks is less wasteful than it looks; right-sizing matters less than people think, though using Fable for email is still overkill.
• Jokes: AGI is when GitHub shows the README above the code; ASI is when Google ships pleasant software. Odysius (~100k stars) has more stars than T3 Code (46k).