DeAI Daily Brief — 3 October 2026

Today in DeAI: Conway Research launches a fully on-device personal AI with a16z backing, the Ethereum Foundation ships zero-knowledge payments for AI inference on mainnet, and NVIDIA adds a $4,999 rung to its local-AI desktop ladder.

Conway Research launches Underdog, a personal AI that never leaves your Mac

Sigil Wen's Conway Research shipped Underdog, an assistant that runs entirely on Apple-silicon Macs — email, calendar, meeting notes, dictation, and browser errands gated behind user approval — with no data sent to external servers, per the company. Andreessen Horowitz announced it is leading the lab's first funding round; the round size was not disclosed. Conway says its lineup includes Woof, a 4B model under 2.5GB, and Underdog 27B, which it claims outperforms frontier models on certain benchmarks — all vendor claims pending independent benchmarks. Why it matters: top-tier venture money is now betting that private inference on hardware users already own can compete with cloud assistants. (a16z) — our full coverage

Ethereum's zkAPI puts zero-knowledge payments behind OpenAI-compatible inference

The Ethereum Foundation and the Open Anonymity Project shipped zkAPI live on mainnet: deposit credits into an onchain vault once, then authorize API spends with zero-knowledge proofs so the provider sees funded requests but never the funder's identity. The client exposes standard OpenAI and Ollama APIs on localhost, so existing tools work unchanged. The limits are stated by the authors: prompts remain visible to the provider, and shared content can act as a fingerprint — zkAPI hides the billing link, not the content. Why it matters: it is the payment-privacy layer the rest of the private-inference stack assumes but rarely ships. (Ethereum Foundation) — our previous zkAPI coverage

NVIDIA adds a $4,999 64GB DGX Spark as the 128GB price reportedly climbs

NVIDIA confirmed a 64GB DGX Spark at $4,999 shipping this month via six manufacturer partners, keeping the GB10 Grace Blackwell chip and full software stack of the 128GB model. Secondary retail reporting puts the 128GB unit near $6,950 — an unconfirmed figure, not an NVIDIA statement. NVIDIA says the 64GB config supports up to 100-billion-parameter models locally, which implies heavy quantization, and that two clustered units deliver up to 1.7x performance. Why it matters: memory, not compute, is setting the price floor on local-AI hardware, and the config gap is now roughly $2,000. (NVIDIA)

Black Forest Labs launches FLUX 3 Image; open weights promised in weeks

FLUX 3 Image does pixel-level multi-turn editing that leaves untouched regions unchanged, with bounding-box layout control, up to ten reference images, and native 4K output. Commercial weights are licensable now, API access is 50% off through October 8, and an open-weight version is promised in coming weeks — a claim, not a date. Why it matters: if the open weights land with a workable license, the local image-model frontier extends again. (The Decoder) — our PULSE coverage

OpenAI prices GPT-6.1 Sol at one-fifth of Astra's rates

OpenAI's DevDay release claims near-GPT-6-Astra capability on agentic coding and computer use at $2/$10 per million input/output tokens — one-fifth of Astra's $10/$50 — with cached input at $0.10 per million tokens. The capability comparison is OpenAI's own assessment pending independent benchmarks. Why it matters: closed-model price floors keep falling, which is the benchmark open-weight and decentralized options get priced against. (OpenAI)

Watching tomorrow

Whether Black Forest Labs pins down the FLUX 3 open-weights timeline and license — and whether NVIDIA's RTX Spark laptop teaser on October 7 undercuts the DGX Spark price ladder.