Happy to answer questions about the workflow, trade-offs, and where this breaks down.
This is very much an ongoing experiment, so I’m especially interested in what doesn’t work well for you.
BruceWok, thank you so much for this thoughtful feedback! These are exactly the kind of questions that push me to make the product better.
Let me address each of your points:
*1. AI consistency across long conversations:*
Great question. Currently, each AI response includes the user's latest assessment scores in the prompt context, so it stays aligned with their 8 dimensions. However, you're right that maintaining coherence over weeks is challenging. I'm working on:
- Session memory that tracks conversation themes
- Periodic "consistency checks" where the AI reviews past advice
- A feature to let users flag when AI advice feels off-track
*2. 12-week cycle engagement:*
This is still early (just launched), but I'm tracking:
- Weekly check-in completion rates
- Drop-off points in the journey
- What I've learned so far: Users engage most in weeks 1-2 (novelty) and weeks 10-12 (seeing results). The middle weeks need better nudges.
My hypothesis: Gamification + social accountability will help. Planning to add:
- Milestone celebrations
- Optional "accountability buddy" matching
- Progress streaks
*3. Tighter feedback loop:*
YES! This is critical. Currently building:
- Behavior tracking (not just reflections): What actions did they actually take?
- Outcome linking: Did their "resilience score" improve after following specific advice?
- A/B testing different coaching approaches based on user response patterns
*4. Pricing & team/peer mode:*
I love the idea of a team tier! It makes so much sense—growth happens faster in community. Added to roadmap:
- Team dashboard (shared accountability)
- Peer reflection circles (small groups)
- Lower entry price point ($9.9/month "Growth Starter"?)
*5. Visualization & assessment evolution:*
This is where I want to invest heavily. Plans:
- Interactive growth timeline (not just static charts)
- "Inflection point" detection (when did mindset shift?)
- Comparative view (you vs. aggregated anonymous users)
You mentioned this is "one of the more thoughtful attempts" in AI coaching—that means a lot. I'm trying to avoid the shallow motivation content trap. Would love to stay in touch as I iterate. Can I reach out for a follow-up chat once I implement some of these improvements?
Thanks again for taking the time to try it and write such detailed thoughts!
You’re raising a very important concern — the slow disappearance of human-curated knowledge niches. While AI can summarize the obvious and the popular, it struggles to preserve the quirky, community-driven, and idiosyncratic corners of the early internet. Forums and specialty sites were full of experiments, debates, and lived experiences — not just canonical facts.
If we don’t actively archive, incentivize, or reimagine those spaces, AI-generated content may become a sterile echo chamber of what’s “most likely,” not what’s most interesting. The risk isn’t that knowledge disappears — it’s that flavor, context, and dissent do.
Correct—they don’t “know” in the epistemic sense, but they do encode a latent world model that shows up as useful priors.
Put differently: GPT-4 isn’t a knowledge base, it’s a *Bayesian autocomplete* over dense vectors. That’s why it can draft Python faster than many juniors, yet fail a trivial chain-of-thought step if the token path diverges.
The trick in production is to sandwich it: retrieval (facts) LLM (fluency) rule checker (logic). Without that third guardrail, you’re betting on probability mass, not truth.
LLMs will mediate plenty of routine text, but the choke-point shifts from “writing” to “prompting + validating”.
In client projects we see two hard costs pop up:
1. Human review time ⟶ still 2–4 min per 1 k tokens because hallucination isn’t solved.
2. Inference \$: for a 70 B model at 16 k context you pay ~\$0.12 per 1 k tokens — cheap for generation, expensive for bulk reading.
So yes, AI will read for us, but whoever owns the *attention budget + validation loop* still controls comprehension. That’s where new leverage lives.
MIT isn’t “weak” because it allows LLM training; it’s weak because it puts zero obligations on the recipient.
Blocking “LLM training” in a license feels satisfying, but I’ve run into three practical issues while benchmarking models for clients:
1. Auditability — You can grep for GPL strings; you can’t grep a trillion-token corpus to prove your repo wasn’t in it. Enforcement ends up resting on whistle-blowers, not license text.
2. Community hard-forks — “No-AI” clauses split the ecosystem. Half the modern Python stack depends on MIT/BSD; if even 5 % flips to an LLM-ban variant, reproducible builds become a nightmare.
3. Misaligned incentives — Training is no longer the expensive part. At today’s prices a single 70 B checkpoint costs about \$60 k to fine-tune, but running inference at scale can exceed that each day. A license that focuses on training ignores the bigger capture point.
A model company that actually wants to give back can do so via attribution, upstream fixes, and funding small maintainers (things AGPL/SSPL rarely compel). Until we can fingerprint data provenance, social pressure—or carrot contracts like RAIL terms—may move the needle more than another GPL fork.
Happy to be proven wrong; I’d love to see a case where a “no-LLM” clause was enforced and led to meaningful contributions rather than a silent ignore.
Happy to answer questions about the workflow, trade-offs, and where this breaks down. This is very much an ongoing experiment, so I’m especially interested in what doesn’t work well for you.