06-reference

not boring optimism 204 physical ai openai anthropic

2026-07-31·reference·source: Not Boring (Weekly Dose of Optimism)·by Packy McCormick
physical-airoboticsopenai-anthropic-competitionvendor-dependency

Why this is in the vault

Friday curation issue #204 rounds up five physical-AI stories (robot teleop, humanoid models, drone delivery, robotaxis) plus a framing of OpenAI's new researcher-access move against Anthropic's more vertical strategy — the latter is the one item worth tracking against RDCO's own Anthropic dependency.

Curation section

  1. Enigma out of stealth — $71M seed (Index Ventures, Ribbit Capital, Conviction Partners, plus OpenAI/Anthropic/DeepMind angels) for robots.online, a site letting anyone remotely pilot real robot arms. Founders Jonathan Jacobi and Gal Niv are non-roboticists (ex-Unit 8200); the bet is that crowdsourced teleop sessions double as both a data-collection engine and a marketing funnel. Checked the public investor list on this one per our WDoO portfolio-disclosure protocol — Index, Ribbit, and Conviction are named; no Not Boring Capital stake surfaced in any of the coverage (TechCrunch, Value Add VC, Unite.AI), so Packy's enthusiasm here reads as organic curation, not undisclosed portfolio promotion.
  2. Gemini Robotics 2 (Google DeepMind) — a three-model family (vision-language-action, embodied reasoning, on-device) demoed driving Apptronik's Apollo humanoid across walking, manipulation, and multiple end-effectors from a single checkpoint; the on-device variant reportedly adapts to a new robot body from under 200 examples.
  3. DoorDash Air — DoorDash Labs earned FAA Part 135 air-carrier certification (8th operator to do so) and launched an in-house drone program. Packy's framing: DoorDash is commoditizing delivery mechanics while keeping the demand network, in contrast to Uber's abandoned self-driving ambitions.
  4. Zoox robotaxi approval — Amazon's Zoox reportedly won the first US approval for a commercial robotaxi with no steering wheel or driver's seat, the first purpose-built (not retrofitted) driverless vehicle cleared for US commercial deployment.
  5. OpenAI's free frontier-model access for researchers — OpenAI offering free frontier-model access to roughly 10,000 researchers, which Sam Altman frames as empowering outside scientists rather than labs solving everything in-house. Packy reads this as counter-positioning against Anthropic's more "opinionated," vertically-focused approach (e.g., biotech-specific pushes), tied to Altman's recent podcast appearance and an industry open letter on open-weight models.
  6. Extra Doses — issue also links an interview with Hemant Taneja (CEO, General Catalyst) tied to GC's Q2 letter, covering GC's AI/private-markets capital deployment and Anthropic as GC's largest position. Four further "extra" items were listed by title only (FCC on foreign robots, a wall-penetrating camera, Pangram, "no science breakthroughs this week") — the email content available to this fetch cut off before those were expanded, consistent with a paywall/read-more boundary rather than a rendering failure.

No sponsor block was found in the plaintext or HTML body (searched for "sponsored by," "presented by," "paid partner").

Zero deep-fetches were triggered: none of the five main items had a specific-enough hook to RDCO's current build surfaces (Sanity Check, Squarely, MAC, HQ) to justify a follow-through fetch beyond the one confirmatory web search on Enigma's investor list (used for the portfolio-disclosure check, not a content deep-fetch).

Mapping against Ray Data Co

The concrete connection is item 5: OpenAI is now positioning itself as the open-access, tool-for-everyone lab (free frontier access to ~10k researchers) while Packy frames Anthropic as the narrower, more vertically-opinionated player — the same Anthropic whose Claude models are the sole substrate for Ray, the RDCO COO agent, per the L5 north star note that ties COO-agent unhobbling directly to Anthropic's own capability trajectory. This is a single-vendor dependency worth naming explicitly rather than a new risk: RDCO has no meaningful redundancy if Anthropic's roadmap diverges from what Ray needs (e.g., a shift toward narrower vertical products rather than general agent capability). Nothing here demands an action — the founder has already made the concentrated bet deliberately — but it's a data point to keep on the "revisit trigger" list alongside the 1Password MCP-server watch, not a one-off news item to discard.

Related