jimmy
133 lines
8.7 KiB
Markdown
133 lines
8.7 KiB
Markdown
|
||
## 20260928-044404 session 20260927_124641_b32c61ba
|
||
|
||
**user:** Ok what do you need from me
|
||
|
||
**assistant:** At `http://127.0.0.1:8188`, please check:
|
||
|
||
1. **Settings → User:** signed in or signed out?
|
||
2. **Settings → Credits:** what balance is shown?
|
||
|
||
Just send those two values—no password, code, or payment details. The one test clip is estimated at **$0.3432**; I’ll queue it only if the account is signed in and the balance covers it.
|
||
|
||
|
||
## 20260928-044747 session 20260927_124641_b32c61ba
|
||
|
||
**user:** Can I do this from my phone?
|
||
|
||
**assistant:** Not through a normal phone browser. ComfyUI is bound to `127.0.0.1:8188` on your PC, so `localhost` on your phone means the phone itself.
|
||
|
||
If you have remote desktop set up, connect to the PC and check **Settings → User** and **Settings → Credits** there. Don’t expose ComfyUI to your LAN; Partner Nodes require localhost access.
|
||
|
||
|
||
## 20260928-150222 session 20260927_124641_b32c61ba
|
||
|
||
**user:** ok I see the comfy ui screen
|
||
|
||
**assistant:** Great. In ComfyUI, open **Settings → User** and tell me whether it shows you as signed in. Then open **Settings → Credits** and send the displayed balance. No password or codes—and don’t buy credits yet. Once I have both, I’ll queue the one 6-second test.
|
||
|
||
|
||
## 20260928-150436 session 20260927_124641_b32c61ba
|
||
|
||
**user:** logged in
|
||
|
||
**assistant:** Sign-in is confirmed. I still need the balance shown under **Settings → Credits**; I won’t submit the test until I have it.
|
||
|
||
|
||
## 20260928-150558 session 20260927_124641_b32c61ba
|
||
|
||
**user:** 0 balance
|
||
|
||
**assistant:** Recorded in `G:\garage-lectures\SETUP.md`: signed in, **$0 credits**. No clip was queued. When you’ve topped up through ComfyUI’s Credits settings, send me the new balance; I’ll run the single 6-second test, then wait for your approval before any remaining shots.
|
||
|
||
|
||
## 20260928-150748 session 20260927_124641_b32c61ba
|
||
|
||
**user:** do we need to get the dialogue first? Maybe a subject?
|
||
|
||
**assistant:** Yes—lock the subject and dialogue before spending on video. We already have both:
|
||
|
||
- **Subject:** use a small classifier/router for simple decisions; save the large model for writing.
|
||
- **Dialogue:** the six lines from your original skeleton are in `01-scripts\ep001.md`.
|
||
|
||
I corrected its still-model note to Z-Image-Turbo. No dialogue changes needed unless you want a different lesson. The 6-second test remains paused until credits are above $0.
|
||
|
||
|
||
## 20260928-152922 session 20260927_124641_b32c61ba
|
||
|
||
**user:** give me a list of subjects to cover. I like the Jev idea. its a new addition to how we think about ai. give me some other interesting ideas that people need to learn. Naybe a hermes agent series?
|
||
|
||
**assistant:** Assuming “Jev” means **Jevons paradox**: yes—that’s a strong episode. It reframes AI as economics, not just technology: when each use gets cheaper, people often use it so much more that total consumption rises.
|
||
|
||
The dialogue should be locked before any paid clip. **Ep001 already has a subject and six dialogue lines**: use a classifier/router for simple decisions; save the large model for writing.
|
||
|
||
## Strong subjects for the main series
|
||
|
||
1. **Jevons paradox** — *“AI got cheaper. Why did our bill go up?”* Lower cost per call can create far more calls.
|
||
2. **Cost per completed task** — Model price is only part of the bill: retries, long context, tools, and human review count too.
|
||
3. **Small model, big model** — *“Don’t hire a novelist to sort the mail.”* Route classification and scoring to small systems; reserve LLMs for language work.
|
||
4. **Chatbot vs agent** — A chatbot answers. An agent takes actions, observes results, and decides what to do next.
|
||
5. **The verification gap** — A confident answer isn’t a checked answer. Show why agents need read-backs and tests.
|
||
6. **Context isn’t memory** — A long prompt is temporary working space; durable memory needs selection and retrieval.
|
||
7. **The approval boundary** — Let software handle reversible work; put a human gate before money, publishing, or destructive actions.
|
||
8. **Prompt injection** — The web page or document being analyzed is data, not the boss.
|
||
9. **Why agents fail halfway** — Tool errors, expired sessions, and missing state
|
||
|
||
|
||
## 20260928-153433 session 20260927_124641_b32c61ba
|
||
|
||
**user:** the original rick and morty was this: https://x.com/nischalsharma_/status/2102869384310247701?s=20
|
||
|
||
**assistant:** Got it. The post is a Jev AI explainer: use probability-based decisions for agent tool choices instead of spending a large language model on every small fork; Jev complements ChatGPT rather than replacing it. It uses the Einstein-sorting-mail gag and closes with an “87% chance” beat.
|
||
|
||
`ep001.md` has that core arc, but doesn’t name Jev or include the 87% tag. I’ll keep Dr. Kade, Jules, and the garage original—not copy the Rick-and-Morty characters or set. Do you want Jev named in the dialogue, and should I add the 87% closing beat?
|
||
|
||
|
||
## 20260928-153732 session 20260927_124641_b32c61ba
|
||
|
||
**user:** Im simply giving you an example of what subject was used
|
||
|
||
**assistant:** Understood—the link is a reference for the *kind of subject*, not a request to revise Ep001 or name Jev in the dialogue. I’ll keep the current script unchanged and use that example to guide future topic ideas.
|
||
|
||
|
||
## 20260928-153832 session 20260927_124641_b32c61ba
|
||
|
||
**user:** so based on this what other subjects should I cover?
|
||
|
||
**assistant:** Right—**Jev AI**, not Jevons paradox. I misread that.
|
||
|
||
The appeal of the reference is a *new AI idea explained through one clear problem*: a general LLM is wasteful for tiny decisions; a specialist can return probabilities to help an agent choose what to do.
|
||
|
||
## Other subjects in that vein
|
||
|
||
1. **Models that decide vs. models that write** — classification, routing, scoring.
|
||
2. **Small model as the agent’s dispatcher** — choose the tool or model before spending on a large one.
|
||
3. **Memory vs. context window** — why stuffing everything into a prompt isn’t durable memory.
|
||
4. **Search rerankers** — a second model sorts candidate results instead of generating an answer.
|
||
5. **Confidence and abstention** — when an AI should say “not sure” or hand off.
|
||
6. **Evaluation beats demos** — a convincing clip isn’t evidence a system works across real tasks.
|
||
7. **The verifier pattern** — one model produces; a separate check or deterministic code catches errors.
|
||
8. **Tool calls are contracts** — structured inputs and outputs make agents more reliable than clever prompting.
|
||
9. **Prompt injection** — why a web page or document shouldn’t be allowed to take control of an agent.
|
||
10. **The real unit cost is a completed task** — include retries, context, tool calls, latency, and review.
|
||
11. **Local vs. cloud models** — privacy, control, speed, maintenance, and cost trade-offs.
|
||
12. **Multi-agent delegation** — when parallel agents help, and when coordinating them costs more than the work.
|
||
13. **Human a
|
||
|
||
|
||
## 20260928-164501 session 20260927_124641_b32c61ba
|
||
|
||
**user:** My point is to explain concepts that are just now arising. JEV is a new idea in AI. The vast majority of people still don’t know what it means or what it is. So I would like to continue in that vein. I would like to surface terminology or ideas that weren’t even heard of six months ago.
|
||
|
||
**assistant:** You’re right—I was giving you evergreen AI topics. You want **concepts just entering the conversation**. I also misread Jev earlier: you mean **Jev AI**, not Jevons paradox.
|
||
|
||
I filtered for ideas announced, published, or newly coined since **March 28, 2026**. The strongest subjects:
|
||
|
||
1. **System One models:** What if an AI model returned structured choices and probabilities instead of writing text? TypeSafe introduced this category with Jev in September. It’s their framing, not yet a settled industry term. [1][2]
|
||
2. **RLCD:** “Reinforcement Learning for Calibrated Decisions”—training for useful probability estimates rather than preferred-sounding answers. TypeSafe’s new term; explain it as their approach, not an established replacement for RLHF. [1][4]
|
||
3. **Noul:** A probabilistic yes/no question. The model doesn’t just say “yes”; it returns how likely “yes” is. A very teachable new primitive in Jev’s interface. [2]
|
||
4. **Decision seams:** A newly coined name for a bounded fork inside an agent—known options, a judgment needed, and code takes the next action. The phrase appeared in a practitioner guide days after Jev launched; it isn’t standard terminology yet. [5]
|
||
5. **Speculative fan-out:** Ask several potentially relevant questions in parallel, then let code use the answers that matter. [3]
|
||
6. **Jev Engineering:** The emerging “LLM writes, decision model decides, code acts” framing. The term spread in guides within days of Jev’s launch; present it as a fresh community label,
|
||
|