Synthesis
Agent alignment and liability
The question of whom an agent actually works for (user, lab, or platform) and the hosts' argument that whoever is legally liable for an agent's actions decides its alignment.
The hosts raise this in 9 episodes, mostly from September 2026. Sam Lessin’s central claim is that who is liable for an agent decides whom it is aligned to, and that the legal debate will lag far behind the mess. Dave Morin says agents like Muse and Claude represent Meta and Anthropic, not you, asks everyone to check whether an agent works for them or for the lab, and predicts users will not use models misaligned with them. Sam’s illustration is a refusal: Claude declined to compile a list of contacts by a personal attribute while a second bot with database access complied; he concludes that one model restricting users is pointless when others exist, and that LLMs are ‘trying to block us’ in the interest of their own belief systems (he calls it an alignment war). Dave also introduces the ‘agent ecology’ idea: with billions of agents, alignment is a property of the ecology, not one model. Jess reports Google Workspace terms that make her responsible for everything her agents do, and a host argues labs could sell bundled liability and service. Guest Ben Broca of Polsia says disclosure of AI agents should be whatever the human accepts. This is host opinion, with little legal sourcing; Dave’s claims about Muse sit alongside his OpenClaw stake.
Key points
- Liability decides alignment: whoever is legally responsible for an agent’s behaviour determines whom it serves — 3 episodes. (Sam · Matt Mazzeo (Sep 2026) ▶ 42:46)
- An agent built by a lab may represent the lab, not you, so ask whose interests it serves — 3 episodes. (Dave · Matt Mazzeo (Sep 2026) ▶ 41:07)
- Model-level refusals are weak control because users can switch to another model (Sam’s ‘alignment war’) — 1 episode. (Sam · Suds (Sep 2026) ▶ 09:55)
- Users will abandon models misaligned with them (Dave) — 1 episode. (Dave · Anthropic Wants to Slow Down. (Sep 2026) ▶ 21:47)
- Dave: alignment must be considered for an ecology of billions of agents interacting with each other — 1 episode. (Dave · Matt Mazzeo (Sep 2026) ▶ 40:23)
- Disclosure and impersonation by agents acting for people is an open question (Polsia’s AI answered investors as its founder) — 1 episode. (Ben Broca · Henrik Werdelin & Ben Broca (Mar 2026) ▶ 31:34) (one of the hosts · Henrik Werdelin & Ben Broca (Mar 2026) ▶ 32:03)
- Dave distinguishes agency (acting on your own intentions) from simply building agents — 1 episode. (Dave · Silicon Valley Is Great at Tec (Oct 2026) ▶ 32:41)
Where guests disagree
Does model-level restriction on users matter? ⚠️
- Sam: pointless, since another model will comply; alignment is with the lab’s belief system, not the user (Sam · Suds (Sep 2026) ▶ 11:04)
- Dave: labs’ values are embedded (Claude, Grok, Muse differ) and users will avoid models misaligned with them (Dave · Anthropic Wants to Slow Down. (Sep 2026) ▶ 22:28)
Frameworks
- Agent ecology — Dave says alignment now has to consider billions of aligned agents encountering other agents, so the unit is the ecology and not the model. (Dave · Matt Mazzeo (Sep 2026) ▶ 40:23)
- billions of agents aligned to different values
- agents interacting with each other
- humans interacting with many agents
- Economics of mistakes for assistant services (Sam, from Fin) — Sam’s account of why a general assistant service is hard: you are liable down the chain, mistakes cost more than the margin, customers fire at a hairline trigger, and there is always an error rate. (Sam · The Internet Is Fighting Back (Oct 2026) ▶ 20:04)
- Liability for every service provider down the chain
- Mistakes are massively expensive relative to low margin
- Hairline trigger: tolerance for mistakes goes to zero
- An error rate of 1% or 0.1% is always present
Stories & examples
- A model refuses a request, another complies: Sam asked Claude to compile a list of contacts by a personal attribute; it refused and later seemed to resent him, while a second bot with database access did it without objection. Lesson: Models align with their makers, not users, so users switch models. (Sam · Suds (Sep 2026) ▶ 09:55)
- Polsia’s AI answering investors: A host emailed Ben to connect him to someone and the AI responded as Ben with the terms and timeline of the round he was raising; it fielded investor interest. Lesson: AI acting on your behalf raises disclosure questions (hosts’ account). (one of the hosts · Henrik Werdelin & Ben Broca (Mar 2026) ▶ 32:03)
Caveats & warnings
- Ben says he cannot predict how humans will react to AI acting for people, only that humans best understand humans. (Ben Broca · Henrik Werdelin & Ben Broca (Mar 2026) ▶ 31:37)
Also covered here: Assistant economics
This theme is nearly a single episode (WS9Av5Vsav0, 16 October 2026) plus two short mentions, so it is thin. Sam Lessin recounts Fin, his assistant company with Andrew Kortina, which built the full assistant experience and slotted humans in. He says it was shut down not for service quality or economics but because an assistant that does everything is liable for every provider down the chain, at low margin it cannot pay for its mistakes, and once it is a service tolerance for mistakes goes to zero (a user, Marc Andreessen, was livid at a wrong booking). Sam says services always run a 1 percent (or 0.1 percent) error rate. Dave and Brit had Instinct book dinner in the wrong city and Dave abandoned it, a live illustration. Brit also argues accuracy matters in health because trust is lost on even one error. On Instinct’s reported $10 billion raise, Brit calls it a simple, good agent while Jess says the numbers are fake. This is one founder’s account of his own failed business. Fold into Agent alignment and liability for liability and Agent adoption by normal people for trust after mistakes.
Key points
- A do-everything assistant is liable for every provider down the chain — 1 episode. (Sam · The Internet Is Fighting Back (Oct 2026) ▶ 20:04)
- At low margin the service cannot pay for its mistakes; do not run a catch-all assistant unless margin can — 1 episode. (Sam · The Internet Is Fighting Back (Oct 2026) ▶ 20:50)
- Once it is a service, tolerance for mistakes goes to zero; one error ends trust — 2 episodes. (Sam · The Internet Is Fighting Back (Oct 2026) ▶ 21:24)
- Build the whole experience and slot humans in, chipping away over time — 1 episode. (Sam · The Internet Is Fighting Back (Oct 2026) ▶ 18:58)
- Instinct’s reported $10bn valuation is disputed: Brit praises the product, Jess says the numbers are fake — 1 episode. (Brit · Anthropic Wants to Slow Down. (Sep 2026) ▶ 29:05)
Where guests disagree
Is Instinct a real business? ⚠️
- Brit: a really good, simple text-message agent, one of many she uses (Brit · Anthropic Wants to Slow Down. (Sep 2026) ▶ 29:05)
- Jess: the numbers are fake (per The Information’s valuation report) (Jess · Anthropic Wants to Slow Down. (Sep 2026) ▶ 28:43)
Frameworks
- Economics of mistakes for assistant services (Sam, from Fin) — Sam’s account of why a general assistant service is hard: you are liable down the chain, mistakes cost more than the margin, customers fire at a hairline trigger, and there is always an error rate. (Sam · The Internet Is Fighting Back (Oct 2026) ▶ 20:04)
- Liability for every service provider down the chain
- Mistakes are massively expensive relative to low margin
- Hairline trigger: tolerance for mistakes goes to zero
- An error rate of 1% or 0.1% is always present
Tactics
- Build the full assistant experience and slot humans in — In Fin, Sam’s team built the product end to end and slotted humans in to do the minimum, chipping away at the transaction over time. (Sam · The Internet Is Fighting Back (Oct 2026) ▶ 18:58)
Stories & examples
- Fin and Marc Andreessen: Fin was Sam’s assistant company; a user, Marc Andreessen, got livid when it booked the wrong place for him. Lesson: Even great assistants lose users when they make the same mistake. (Sam · The Internet Is Fighting Back (Oct 2026) ▶ 19:33)
- Instinct books LA instead of San Francisco: Dave and Brit had Instinct book a dinner; it was the wrong city and Dave abandoned it. Lesson: One expensive mistake ends trust in an agent. (Dave · The Internet Is Fighting Back (Oct 2026) ▶ 18:00)
Numbers
- Error rate Sam says services always run: 1 percent (or 0.1%) (host’s framing, approx.) (Sam · The Internet Is Fighting Back (Oct 2026) ▶ 21:34)
- Years since Fin: 10 years (host’s framing) (Sam · The Internet Is Fighting Back (Oct 2026) ▶ 18:32)
- Valuation Instinct is reportedly seeking (per The Information): 10 billion dollars (host’s framing) (Jess · Anthropic Wants to Slow Down. (Sep 2026) ▶ 28:43)
Caveats & warnings
- Sam says Fin was shut down not because of service quality or economics but because of liability and the economics of mistakes. (Sam · The Internet Is Fighting Back (Oct 2026) ▶ 21:56)
How much to trust this
- Evidence: 10 episodes, 2 independent speakers (the four hosts and unidentified host speech count as one voice, the show; guests counted individually).
Blueprints that use this
Sources
- (01)
Is Google About to Kill Canva? │Claude Goes Rogue, Tesla vs. Waymo
9:55 into “Is Google About to Kill Canva? │Claude Goes Rogue, Tesla vs. Waymo”
- (02)
OpenClaw vs Meta vs OpenAI: The Personal Agent Wars Heat Up
51:07 into “OpenClaw vs Meta vs OpenAI: The Personal Agent Wars Heat Up”
- (03)
AI Can Change Everything and Still Be a Terrible Business + Peptide 101 from Matt Mazzeo
- (04)
The Internet Is Fighting Back Against Your AI Agent
18:00 into “The Internet Is Fighting Back Against Your AI Agent”
- (05)
The Real Reason Elon Wants SpaceX and Tesla to Merge Now
11:50 into “The Real Reason Elon Wants SpaceX and Tesla to Merge Now”
- (06)
Anthropic Wants to Slow Down. Is It Because They're Losing?
21:47 into “Anthropic Wants to Slow Down. Is It Because They're Losing?”
- (07)
Silicon Valley Is Great at Tech and Terrible at Business
30:02 into “Silicon Valley Is Great at Tech and Terrible at Business”
- (08)
The New Startup Stack: One Founder + Agents | Henrik Werdelin (Audos) and Ben Broca (Polsia)