Ollama adds local text-to-speech, Copilot in Visual Studio moves Bring Your Own Model into its new agent, Bedrock web search can be limited to chosen domains
A quiet weekend for first-hand news. The thread is who decides which model runs and where the data comes from: local runtimes now cover speech, the IDE pulls bring-your-own-model into its own agent harness, and a cloud provider makes web search sources controllable. For people building products, keep the choice of model and source in your own layer.
The first speech model is Phonon, a small TTS family from Gradium. It runs on macOS only for now (Apple Silicon, accelerated by MLX), with Windows and Linux coming later, and needs Ollama 0.40.11 or later.
On a MacBook Pro with M5 Max, Ollama measured 35% faster than the next fastest engine, Gradium's own runtime, and matched the reference Hugging Face Transformers implementation's word error rate on 200 Seed-TTS English sentences.
It works with the OpenAI-compatible /v1/audio/speech endpoint, so existing OpenAI SDK code only needs a new base URL. Audio streams as it is generated, 24 kHz mono, WAV by default or raw PCM.
phonon itself is Apache 2.0; the 40M French, German, Spanish, Portuguese and multilingual models are CC BY-NC 4.0, noncommercial only.
Builder's takeFor a product that has to speak, like AI Interview, I'd prototype local voice on a Mac with Phonon first; the interface matches OpenAI's, so switching back to the cloud costs almost nothing. Check the license before launch, though: only the English model is commercial-use, there's no Chinese, and the multilingual ones are noncommercial.
Bring Your Own Model is still in preview and on by default in Community, Professional and Enterprise. It uses your team's preferred models through Copilot or your own deployments, signed in to GitHub or not; providers include Microsoft Foundry, OpenAI, Anthropic and Ollama.
Breaking change: BYOM now requires the new Agent (Preview), built on the GitHub Copilot SDK-powered harness. The previous Ask and Agent modes no longer support it, and previously configured Ollama models must be re-added after updating.
The Git agent can take a pull request as context and return a summary with links to its comments and files; it needs your Git provider's MCP servers, with GitHub and Azure DevOps supported.
NuGet Audit warnings in the Error List gain a fixer that opens Copilot Chat in agent mode to fix vulnerable packages via the NuGet MCP server. The update is on every Copilot plan.
Builder's takeRead next to the 10-25 change that split usage by client: Copilot is handing over the choice of model while pulling the entry point into its own agent harness. Teams wiring their own or local models into Visual Studio should note the current setup before updating and retest in the new Agent afterwards, rather than waiting for developers to find their model gone from Ask mode. Same point as my post on model routing: keep the rules for picking models in your own layer, not only in IDE settings.
Both controls are set directly on the web_search tool in an existing OpenAI Responses API request.
The filters parameter takes up to 100 allowed_domains to limit results to trusted sources, or up to 100 blocked_domains to exclude ones you don't want.
Optional user location parameters make results relevant to a geography, for questions like regional availability or local services.
Available in all Regions where Web Search is supported: US East (N. Virginia), US East (Ohio), US West (Oregon) and AWS GovCloud (US-West).
Builder's takeAs in my RAG retrieval checklist post, uncontrolled sources are the most common hole in web-grounded answers. For support or documentation Q&A I'd lock allowed_domains to our own docs plus a few authoritative sites first, then measure how much answer coverage drops. It's US Regions only for now, so for business in China it's a reference point for the moment.
The post is Part II of OpenAI's engineering series on rapidly scaling online storage for over 1 billion ChatGPT users, covering a storage layer called Habitat.
Per OpenAI's summary, Habitat was scaled with reliable multi-tenancy, global caching and Azure Cosmos DB.
OpenAI says the work delivered 12x faster reads for ChatGPT.
Builder's takeConversation, memory and file state is the layer of an AI product that breaks first under load and gets the least attention. For a read-heavy product like AI Cloud Drive I'd settle tenant isolation and the cache layer before worrying about switching models. The full page couldn't be fetched, so the details here rest on OpenAI's own summary.
Built with the University of Pennsylvania, LitPivot retrieves papers through Asta Paper Finder, organizes them around the part of a proposal being edited, and offers critiques backed by excerpts plus revisions researchers can accept, edit or reject.
A proposal is split into problem, proposed contribution and evaluation plan, each with its own related literature and ranking; the literature is reranked after a revision.
In a study with 17 researchers, domain experts rated ideas developed with LitPivot as better grounded in the literature than those made with a text editor and an AI chat tool. The paper will be presented at UIST 2026.
An analysis of meeting notes from four completed projects found 21 literature-initiated pivots.
Builder's takeThe part worth borrowing is the interaction, not the paper: split a document into parts, attach evidence to each, require a citation for every suggested edit, and leave acceptance to a person. For a content back end like PandaClaws I'd attach sources to each generated section and have editors confirm section by section, rather than generating a whole piece in one click.
Partner Nexus AI turned Google's chest X-ray algorithms into a certified diagnostic tool with EU MDR Class IIb and ISO 13485. It runs in about 40 facilities across six countries, where over 90 clinicians have evaluated more than 25,000 chest X-rays, with results in as little as 45 seconds.
Through HAI-DEF Google open-sourced chest X-ray embeddings and the medical vision-language model MedSigLIP, and released MedGemma, open-weight multimodal models in 4B and 27B sizes that run on-device offline; MedGemma has passed 10 million downloads.
Indonesia's Ministry of Health, aiming to screen 50 million people a year, is fine-tuning MedGemma and MedSigLIP on domestic chest X-rays, running on local infrastructure so patient data stays in the country.
HeAR, trained on over 300 million audio segments, screens from cough sounds recorded on a standard smartphone microphone.
Builder's takeFor anyone selling to enterprises or regulated industries this is more useful than a model launch: open weights plus local deployment solve the rule that data can't leave the country, not just cost. Selling to hospitals or governments, I'd put runs offline and fine-tunes on the customer's own machines on page one of the proposal.
Researched and drafted with AI assistance; editorial standards and views set by Darius.
Every Monday, the week in AI in one email
Every Monday morning: the 6 AI stories that mattered most in the past seven days, with my take, in your inbox.
Your email is only used to send the weekly and is never shared. Every email has a one-click unsubscribe link. See the latest issue →