saved
On-device intelligence for every product
Hraness cites a source capture. The source author remains the source.
gist
Paul Veugen launches Desert Ant Labs, a European on-device AI lab shipping 18 specialized audio, vision, and text models through one Swift, Kotlin, and JavaScript SDK. Models such as Voz, Clear, and Redact answer in milliseconds on phones at zero inference cost, often beating larger cloud and local baselines. The thesis is cerebellum-first: free local models for always-on work, then a router to bigger local or cloud models when needed. Built from Detail's cloud-API bills, the lab treats training as product design and sells commercial drop-in runtimes instead of tokens.
ideas
- Ship task models, not generalist APIs. Eighteen specialized models cover transcription, enhancement, PII redaction, language ID, and more, each sized to finish one job fastest on-device.
- Zero-cost inference changes product shape. Free local calls let features run on every frame or keystroke instead of the subset a cloud bill can afford.
- Cerebellum first, cortex later. Always-on little brains handle routine work locally; a router escalates to larger local or cloud models only when the job requires it.
- Training is a product-design problem. Detail's cloud bills pushed the team to train drop-in models that beat Dolby, Whisper, and Sonnet-class APIs on speed, energy, and size.
- Device silicon is already paid for. Phones and laptops already ship capable chips; European on-device defaults keep customer data off servers that can be compelled.
quotes
“We believe the best path to efficient intelligence starts on-device.”
“There's more compute available in people's hands than in every AI data center on earth.”
“When inference costs nothing, the way we build products changes entirely.”
“Think of the first hundred models as the cerebellum, the little brain.”