She Left UWaterloo Without Graduating. Now She's at OpenAI | Brooke Joseph

Scientific concepts, discoveries, and nature phenomena mentioned

Privacy / security concepts

• Differential privacy (broader concept: formal methods for limiting information leakage when analyzing or releasing data)
• Security and safety in AI (mentioned as a major emphasis at OpenAI when deciding what tasks to delegate to agents)

Machine learning / model compression (edge deployment)

• Quantization: reducing the information/precision used by trained ML models so they can run on small microcontrollers / edge devices
• Model distillation: training a model to learn from multiple other models, so knowledge transfers from “teacher” models to a typically smaller model

AI agents and “computer use” (human-computer interaction via models)

• Computer use: enabling an AI model to perform click/type/scroll/keyboard actions on a user’s computer on their behalf
• Background computer use: ability to run multiple computer sessions concurrently while keeping other work on the user’s computer available (discussed as a technical capability demonstrated for Mac)
• End-to-end workflow “closing the pipeline”: moving from text prompting to agents executing tasks in real software, including multi-step workflows (e.g., verifying results, editing in professional software, CAD, video review/editing)

Research framing / evaluation claims

• Codex computer use for Mac and specific model version improvements (e.g., “5.6 is very good at coding”)
• “GTB 5.6” is mentioned (likely “Codex” / “GPT-5.6” in context), tied to the idea that strong coding skill is important on the AGI road

Nature phenomena

• None mentioned.

---

Methodologies / systems described (outlined)

Approach to learning new technical topics deeply (personal methodology)

1. Ask a model (often Codex) to “go deep” with high-level details to fundamentally understand a topic.
2. From the model’s explanation, select one point (out of many) and drill into it.
3. Ask targeted follow-up questions to understand:
   • the meaning of representations (e.g., “what does this represent?”)
   • relevant statistics concepts
   • underlying formulas and their interpretation
4. Watch for rabbit holes and stop once the thread becomes unproductive.
5. Prefer paper notes for absorption; later distill into a structured write-up (e.g., Notion).
6. Validate understanding by teaching/explaining to someone else (e.g., a partner).

Agent workflow concept (product framing)

• The user provides intent through a chat/prompt.
• Agents use computer use to operate external software (click, type, scroll).
• Agents can perform tasks like searching, editing, verifying, and then returning outputs (e.g., a spreadsheet of hotel options).

---

Featured researchers / sources (named in subtitles)

• Ari (Sky/OpenAI founder; last name not given in subtitles)
• Conrad (Sky engineer; last name not given in subtitles)
• Dexter (intern mentioned as building UI/computer-use cursor interactions; last name not given)
• Marilyn Monroe
• Mr. White (a teacher/professor; first name not given)
• Brooke Joseph (subject of the interview; also the speaker credited throughout)
• OpenAI (organization; source in context of models/product like Codex / GPT versioning)

(No external scientific papers, journals, or named research institutions/peer reviewers are cited in the provided subtitles.)