Nvidia put out a free, open-weight transcription model that runs on one chip — but only on recordings, not live audio
October 30, 2026 · 4 items
Nvidia put out a free, open-weight transcription model that runs on one chip — but only on recordings, not live audio
TechCrunch · October 28, 2026Buildable AI
Nemotron 3.5 Transcribe is free and open-weight, transcribes up to 40 languages, translates 31 of them to and from English, handles recordings up to an hour long, and can tell up to eight speakers apart; Nvidia claims it beats rivals on benchmarks including Hugging Face's Open ASR Leaderboard.
At 30 billion parameters it runs on a single chip — from the Jetson AGX Thor robotics controller and DGX Spark powering new laptops up to Blackwell datacenter chips.
TechCrunch names business meetings and medical transcription as the two target markets, and Nvidia says developers can combine the model with the updated Granary v2 speech dataset and its NeMo Speech tooling to teach an app an industry's jargon in each language.
The stated drawback: it works on recorded audio, not livestreams — real-time still belongs to rivals like ElevenLabs, whose Realtime launch TechCrunch lists alongside Mistral's Voxtral 3 and Salesforce's Launa-Omni.
Why it matters for an independent practice: This is the first credible route to transcription a practice can run on its own hardware instead of shipping patient audio to a vendor cloud — which changes the BAA conversation entirely. Because it is recording-only, scope it to dictated post-visit notes, recorded patient-education video, and podcast/marketing content repurposing, not live ambient scribing.
Chrome's autofill will now read your Gmail to fill out web forms — opt-in, US desktop first
TechCrunch · October 28, 2026Buildable AI
Google said Wednesday that Chrome is getting an advanced autofill that pulls form suggestions out of a user's Gmail — flight confirmation numbers, membership IDs, tracking info, order numbers, addresses, phone numbers, even credit card numbers.
It is an expansion of Google's Personal Intelligence, which already lets users connect Google Apps to Gemini so the assistant can retrieve details and answer more complex questions.
Users decide whether to turn enhanced autofill on from Chrome settings, and Google notes sensitive data is encrypted.
Rollout is to U.S. desktop users and in the iOS and Android apps for Gemini in Chrome.
Why it matters for an independent practice: The staff browser at a practice is signed into the clinic inbox and the patient portal at the same time. Before anyone flips this on, decide in writing whether a browser feature that reads inbox contents to populate forms falls inside your BAA — and add the Chrome setting to your workstation onboarding checklist rather than discovering it after the fact.
Haut.AI says its new skin model scores 69.6 on a benchmark it wrote itself, versus 50.6 for the best general-purpose model
Longevity.Technology · October 28, 2026Regenerative medicine
Haut.AI unveiled SkinFM, an image-based skin foundation model trained on more than 4 million real skin images, enabling assessment of over 230 skin measurements and 48 biomarkers covering wrinkles, redness, pigmentation and other visible characteristics.
Alongside it the company released SkinBench, a public benchmark of 19 tasks across seven common concerns, comparing its technology against six leading general-purpose and medically focused models; Haut.AI reports 69.6 for itself versus 50.6 for the highest-scoring general-purpose model.
The company reports high false positive rates for eye puffiness and missed detections for deep wrinkles among several large models — while also noting that Haut.AI did not outperform on every metric.
The underlying source is a PR Newswire release, and the benchmark, the scoring and the comparison set were all authored by the vendor being measured.
Why it matters for an independent practice: Aesthetic and regen-med clinics are actively being pitched AI skin analysis as a consult and conversion tool. This is a vendor-run benchmark, not independent validation — so the question to ask any skin-scan vendor this quarter is who scored the test, on whose images, and whether the clinic's own before/after library can be used to check it.
Google is taking ads inside AI Search international as its core search ad growth slows to 12%
TechCrunch · October 28, 2026Patient acquisition
Google announced Wednesday, on Alphabet's third-quarter earnings call, that it is bringing ads in its AI-powered Search experiences to international markets.
Google Search advertising revenue grew 12% year over year in Q3, down from 17% growth in Q2 and 19% in Q1 — a visible deceleration in the blue-link business.
Alphabet overall reported $127.6 billion in revenue, up 25% year over year.
TechCrunch frames the move as a bet that generative AI can open new revenue as the core search advertising business slows.
Why it matters for an independent practice: Paid placement is moving into the AI answer itself, which means a practice's ad strategy and its 'will the AI cite us' strategy are now competing for the same pixels. If your reporting still measures patient acquisition as blue-link clicks and form fills, it will understate both the loss and the new inventory — rebuild the dashboard before the budget conversation.
