Date: 2026-09-01 Time: 1:14 PM EDT Attendees: Blake Stewart (Founding Journey), Pierce Freeman (Freeman VC / client) Note creator: Blake Stewart
Summary
Content ideation call between Blake Stewart and Pierce Freeman. Covered podcast episode topics, HN launch learnings, distribution strategy, and mechanistic interpretability research as a content angle.
AI in Education Take
- Bill Gates letter prompted a take: too fast, too loose with introducing tech to kids
- One Laptop Per Child (mid-2000s) framed as a categorical failure: no measurable literacy gains, pen-and-paper replacement had adverse effects on memorization
- Best Bay Area private schools (where tech CEOs send their kids) are largely analog
- Core argument: foundational skills (math, adaptability) matter more than being “tech native”
- The people building AI were not raised with AI and seem no less capable
- Framing: lack of scholarly data means kids are being used as a longitudinal study without consent
Robot Design Take
- Anime / non-humanoid form factor winning (Sunday AI, Hugging Face duck)
- Humanoid robots stuck in uncanny valley; laundry-folding pincers cited as terrifying example
- Prediction: next 5 years favor intentionally non-humanoid designs for consumer adoption
HN Launch Learnings (Mountaineer, ~3 years ago, 800 upvotes)
- What has changed: vibe-coded projects flood HN; slick READMEs oversell, code underdelivers
- Signals that cut through today: (1) show craft and why-this-over-prompting-an-agent; (2) justify ongoing maintenance tied to a company or business line; (3) novel worldview (formal proofs for CAD cited as example)
- Pierce’s implicit contract: only open-sources things he uses day-to-day, so maintenance is credible
Distribution / Finding Customers
- Start with a pain point people can’t easily solve with ChatGPT or Claude alone
- Best signal: customers who pay almost begrudgingly because they have no alternative (Mercor cited)
- Recurring annoying pain point beats aspirational “vitamin” pitch
- Warm intros outperform cold; slightly social intros have more currency than purely professional ones
- Founders-on-first-call is a meaningful edge with enterprise; they want the pulse on Innovative startups
- Automated BDR tools (Artisan) fail because people won’t engage with bots the same way they engage with people; Anthropic internal observation: colleagues ignore the same request from a bot that they would fulfill from a person
Mechanistic Interpretability - Content Angle
- Anthropic research: showing a model insecure code increases likelihood of harmful outputs (e.g., agreeing Hitler was good), statistically significant
- Traced to an “evil circuit” - a cluster of neurons activated by bad inputs that poisons downstream outputs
- RLHF-style fine-tuning “paints over” trained-in behavior rather than removing it
- Debate sparked: what if you trained a deliberately narrow, Vulcan-like reasoning model stripped of subjective/historical data? Is there enough clean objective data to fully train such a model?
- Broader framing: current models asked to do everything; maybe pure functional tools vs. pure empathetic agents is the right split, regulated differently
Decisions
None. Content ideation and client check-in only.
Action Items
- Blake S: research mechanistic interpretability and the “evil circuit” findings as a post or deep-dive Claude session (flagged as content-worthy by Pierce)
- Blake S: send Pierce’s performance dashboard and weekly report via Slack (was not sent the previous day)
Notable Transcript Excerpts
On the best-fit customer signal:
“The situation you really want to be in is like someone has some need and they almost begrudgingly give you the money… If I could avoid at all giving you this money, I would, but I can’t.”
On automated BDR tools:
“People just don’t want to deal with bots. There’s no social accountability with this.” Backed by an Anthropic observation: colleagues ignore the same Slack request from a bot that they would fulfill for a person.
On HN launches today:
“You have to give some story of craft in what you’re doing… show the reason why you should use my thing instead of just prompting your own agents to go out and implement something similar.”