I seriously should NOT be dropping this.

Technological concepts, product features, and analysis

New AI project: “Ajax”

• Described as a fine-tuned AI model (named by reference to “Ajax” cleaner).
• Claims it can:
  • Browse the web
  • Do private searches
  • Write emails
  • Use a calendar
  • Manage to-do notes and perform inbox/calendar/task assistance inside “Odysius”
• Includes a legal/safety disclaimer, emphasizing it is not intended for dangerous actionable instructions (e.g., “nuclear war weapon” type examples).

Local AI emphasis: “Decensored”

• Core claim: “the second greatest thing… about local AI… is that is decensored.”
• “De-censoring/decensoring” is framed as:
  • Removing or weakening the model’s refusal behavior (its ability to say “no”)
  • Achieved via a “brain surgery” concept: model weight editing/ablation, where refusal behavior is treated as distributed across model weights
• Mentions “heretic” as an open-source method/program that makes the process “super simple.”

Safety boundary and ethics

• The speaker says they considered where to draw the line and chose to avoid decensoring content that could lead to:
  • Harm to others
  • Harm to oneself
• Strong warnings include:
  • “Please use it cautiously and responsibly”
  • “do not distill open AI’s model.”

Training approach / iteration

• Privacy-first constraint: claims they could not rely on open-data collection because Odysius is privacy-first and collects no data.
• Data collection idea (abandoned):
  • They discussed collecting user data to train Odysius/Ajax.
  • Created data.pdy.com for submissions.
  • Even with tutorials, they claim almost nobody submitted data.
• Distillation:
  • Intends to distill knowledge from a “larger model” (an OpenAI “soul” model is referenced).
  • Claims reasoning tokens/chain-of-thought were relevant, but doing so via a study/API approach would violate terms—so they state they wouldn’t do it.
• GRPO training:
  • Claims GRPO has been running for weeks.
  • Describes an interactive/repeated-task style:
    • The model runs the same task multiple times (16 times)
    • Signals from successful runs reinforce training
  • Training pattern described as “improving → flatlining → improving again”, with plans for further checks.

Benchmarks and evaluation

• Says they will quantize and benchmark to ensure performance wasn’t lost after additional training/experiments.
• Uses a “version 1” framing: the trained model is described as tiny and not final.

Product positioning and expected use-cases

• Ajax is positioned primarily for speed and information work, not replacement of human thinking/creativity:
  • Searching / researching / browsing
  • Saving time by quickly processing lots of information
• Argues that very large “trillion-parameter” models are impractical for small tasks; instead, advocates small, harness-specific local models.

Community/open-source integration

• Odysius/Ajax development is tied to GitHub.
• Includes thanks to maintainers and mentions “reducers” contributing work.

Tutorial/guides embedded via sponsor segments

• Boot.dev is used to discuss:
  • Learning to code and why fundamentals still matter even with AI coding assistance
  • Mentioned courses:
    • Linux course by BHAD
    • GitHub course by “Primogen”
  • Emphasizes that reinforcement learning/fine-tuning needs meaningful training signal and that “teacher help” (when stuck) accelerates learning.

Privacy/security advice segment (sponsor)

• NordVPN promotes protection from:
  • Trackers
  • Phishing / fake websites
  • Scanning downloads
• Privacy benefits are framed as reducing account compromise risk and privacy exposure.
• Includes a call-to-action using a code and referral link.

---

Key reviews / claims / warnings summarized

• Ajax performance claim: “working 9 out of 10 times” (for its described tasks in Odysius).
• Model size: “tiny tiny tiny tiny model,” intended to run on a user’s own computer.
• Main risks/warnings:
  • Avoid enabling harmful instructions
  • Do not distill OpenAI’s model (explicitly stated)
  • Decensoring should be used cautiously/responsibly

---

Main speakers or sources

Main speaker

• The video creator, presented in a PewDiePie-style persona (repeatedly says “I’m PewDiePie” and uses PewDiePie sponsor code).

Sponsors / cited sources

• Boot.dev
• NordVPN
• heretic (open-source project mentioned for decensoring/weight-ablation approach)
• Primogen (mentioned as GitHub course provider)
• BHAD (mentioned for Linux course)