06-reference

innermost loop interspecies foundation model sarama

2026-05-23·reference·source: Innermost Loop·by Alex Wissner-Gross

"The First Consumer-Scale Interspecies Foundation Model" — @Alex Wissner-Gross

Why this is in the vault

Wissner-Gross essay announcing Sarama, a dog-collar foundation-model startup he advises and is invested in. Framing: AI has only ever trained on species that wrote themselves down; non-human cognition is the next frontier where the record must be measured into existence. Worth keeping as (a) a clean example of the AWG "Solve Everything / Intelligence Revolution" thesis applied to a concrete vertical bet, and (b) a sponsored-portfolio essay where the disclosure pattern is explicit and instructive.

⚠️ Sponsorship

Wissner-Gross explicitly discloses he has a financial interest in Sarama and helped form/advise the company. The disclosure is at the bottom: "I have a financial interest in Sarama. This essay is informational only, not investment or veterinary advice. Performance figures are company-reported."

This is the cleanest version of his sponsor pattern: the entire essay IS the promotion of a portfolio company, with the disclosure clearly labelled. Read it as a thesis-piece-plus-pitch, not neutral analysis. The 93% accuracy claim and "1M annotated bark sequences by 2026" projection are company-reported figures — treat as forward-looking, not measured.

Sister-pattern to the Innermost Loop / 021T relationship more broadly: AWG's newsletter is increasingly functioning as a public-thesis substrate for his investment activity (Sarama follows the same shape as the broader "Singularity" essays — each frames a vertical where a foundation-model bet is becoming legible).

The core argument

  1. Every prior AI advance trained on already-written records (text, code, protein, images). Non-human animal cognition has no such record — meaning lives in what reliably happens around a signal, not in the signal itself.
  2. Synthetic data cannot solve this: interpolation only works inside distributions a model has already seen.
  3. Sarama's wedge: instrument-generated, self-labeled data in the home. A custom collar (camera + microphone + motion sensors, transmitting feature vectors not raw audio) where sensors cross-label each other (camera frames food bowl, motion logs lunge, mic logs bark — three readings labeling one another, no human annotator).
  4. Per-dog models, not a universal bark dictionary. Meaning is individual; the parallel is "no universal human dictionary works either."
  5. Dogs are first because they co-evolved to be legible to humans, share our home environment, and exist at consumer scale. The method generalizes to horses, herds, wildlife from drones.
  6. Closing frame: "We are learning to understand the intelligence that evolved beside us just in time to face the kind we are building ourselves." Aligns the bet with the alignment discourse.

Mapping against Ray Data Co

Two threads of relevance:

Mapping strength: medium. Conceptually interesting, sponsor-pattern useful as catalog, but no immediate decision triggered.

Related


Copyright note: paraphrased and summarized from author's plain-text email. Quoted excerpts ≤15 words. Original at source_url.