Skim here first: the closed row is the glance, open is the study card with the key points and timestamps, and the ↓ link drops to that concept's full write-up below.
He describes reinforcement learning from human feedback as the model improving as people tell it what was wrong - like coaching an intern.
History (0:00-0:04)
RLHF (0:03-0:04)
Choose a model per task from the top-left dropdown; build custom GPTs for standing instructions; turn on Memory so it references what it knows about you.
Why it took off (0:06-0:08)
Dashboard (0:08-0:11)
Settings (0:11-0:13)
"Task one was not done truly properly" - and the next attempt improves.
This is a simplified picture: training-time RLHF shapes the model in advance, while in-chat feedback mainly affects the current conversation (and Memory, if on). Thumbs-up/down feedback still helps the vendor improve models.
A mental model for why phrasing feedback clearly matters.
A prompt "is what you want to do to it".
At recording: GPT-4o "great for most tasks", o3 "uses reasoning ... tells you why it is giving you an answer", with new models "every two months or so". The GPT store holds custom GPTs (e.g. Canva's); a custom GPT stores a name, description and instructions. Settings > Personalization > Memory > reference saved memories. Threads are auto-named and can be renamed.
Orientation before the feature lessons.
| Tool | Coverage | Moment | Context |
|---|---|---|---|
 ChatGPT | demonstrated | ||
| GPT-4o | explained | ||
| OpenAI o3 | explained |
Archived locally on V: — click to open. Companion pages link to the LMS.
This page was built from an auto-generated transcript, which garbles product and people's names. Those were corrected silently in everything above and logged here for transparency. The warnings flag claims that were true on the recording day but change fast.
| The transcript says | The trainer actually means |
|---|---|
| OpenEye | OpenAI |
| chat gpt 4.0 | GPT-4o |
| 0.3, 0.3 mini | o3, o3-mini |