Home / AI / Underdog shows off a 32MB speech model in new demo
Underdog shows off a 32MB speech model in new demo
techcrunch.com Nov 7, 2026

Underdog shows off a 32MB speech model in new demo

Curated by · TechCapsules ·

AI-summarised brief · reviewed before publication

Underdog AI unveiled a browser‑based demo featuring its new text‑to‑speech model, Woof, which occupies just 31.7 MB (31 million parameters) and runs entirely offline. The interactive game, set in a retro‑tech room, lets users type on a virtual typewriter or adjust a music tuner, with every input spoken by Woof. The startup promotes the model as a lightweight alternative to cloud‑dependent systems, emphasizing data privacy and device‑side processing. Woof follows Underdog’s recent release of Bark, a 167 MB speech‑to‑text model also built for edge devices. Founder Sigil Wen told TechCrunch the company aims to deliver open, American‑made AI that can run on existing consumer hardware within months, and has also announced a 27‑billion‑parameter model that can operate on laptops.

💡 Why It Matters

  • · It proves high‑quality speech synthesis can be delivered locally without cloud reliance, opening the door for private, on‑device AI experiences.
AI Capsules Startups
Read full article on techcrunch.com
More in AI →
iOS 27.2 adds Apple Photos invisible watermark feature, here’s how it works
9to5mac.com · Nov 7

iOS 27.2 adds Apple Photos invisible watermark feature, here’s how it works

iOS 27.2, released last week, introduces an invisible watermarking system to Apple Photos that embeds SynthID identifiers in images altered or generated with Apple’s AI tools. The feature activates when users employ AI‑powered edits such as Clean Up, Extend, or Spatial Reframing on existing photos, as well as when they create new pictures in Image Playground. The watermark is not visible in the picture itself; it can only be detected with a SynthID scanner, which Apple directs users to access at synthid.com. The same functionality rolls out concurrently to iPadOS 27.2, macOS 27.2, and visionOS 27.2, extending the protection across Apple’s entire device ecosystem. Apple’s documentation notes that the watermark applies to both AI‑generated and AI‑enhanced content, aiming to preserve provenance in an era of synthetic media.

💡 By embedding traceable metadata directly into edited images, Apple gives creators a built‑in method to prove authenticity, countering the growing risk of AI‑fabricated visual misinformation.

Ramp data shows AI spending stalling just as Anthropic nears its IPO
techcrunch.com · Nov 7

Ramp data shows AI spending stalling just as Anthropic nears its IPO

Ramp’s latest data shows AI spending stalled in late summer, with token‑based expenditures rising only 1.5 % in August‑September after a 42 % jump in July. The slowdown is driven by price cuts from OpenAI and Anthropic and intensified competition from cheaper or open‑source models, which have turned the market into a buyer’s market. While the top 1 % of AI‑heavy firms dipped just below $10,000 per employee per month, median firms remain at $12‑$15 and the top 10 % at $600‑$1,000, insufficient to offset the lagging elite spend. Anthropic’s upcoming IPO may be challenged by these trends, as investors’ expectations of relentless growth confront a market where volume rises but revenue plateaus. October saw an 8.9 % token spend rebound, yet headwinds persist.

💡 The emerging buyer’s market forces AI vendors to prioritize pricing over growth, reshaping revenue forecasts for high‑profile players like Anthropic ahead of its IPO.

OpenAI reportedly raises an additional $11 billion at a $1.25 trillion valuation
techcrunch.com · Nov 7

OpenAI reportedly raises an additional $11 billion at a $1.25 trillion valuation

OpenAI has secured $11 billion in a new funding round, valuing the AI lab at $1.25 trillion, Bloomberg reports. The round is still open and could reach roughly $20 billion, though the company originally aimed for $30 billion. The announcement follows Bloomberg’s report that Anthropic’s potential IPO valuation may fall to about $1.1 trillion from the previously expected $2 trillion. OpenAI had hinted at an IPO this year but postponed it, with a next‑year launch now anticipated. The move underscores the volatile market for AI firms amid regulatory concerns.

💡 The funding confirms OpenAI’s continued dominance in the AI race, while the contrasting Anthropic valuation signals a tightening market that could reshape investor expectations for AI IPOs.

From the front row to the hot seat — what the midterms mean for Big Tech
techcrunch.com · Nov 7

From the front row to the hot seat — what the midterms…

OpenAI hired Jessica Schumer, daughter of Democratic Senate leader Chuck Schumer, and rebranded its data centers as “compute campuses” just before the 2024 midterm elections, signaling a shift toward rebuilding Democratic ties after two years of close cooperation with the Trump administration. During that period, major tech firms attended White House events, scaled back DEI programs, and donated to pro‑Trump super PACs, while benefiting from tax breaks, paused investigations and looser AI regulation. The midterms, which could tighten Democratic oversight of AI, have prompted companies such as Meta, Anthropic and OpenAI to launch public‑private initiatives—teacher partnerships, workforce academies, and “Claude Corps”—as a PR effort to improve optics. Industry observers warn that real battles will move to state legislatures, local data‑center siting decisions, and market reactions to the AI boom, while federal action may stall amid partisan gridlock.

💡 The pivot underscores how quickly Big Tech can reorient its political strategy to safeguard regulatory environments, turning election outcomes into a direct lever on the future of AI governance.

Meta introduces Spaces, a new way to collaborate with your Muse agent
techcrunch.com · Nov 6

Meta introduces Spaces, a new way to collaborate with your Muse agent

Meta launched Spaces, a new feature that lets users create topic‑specific breakout conversations with its Muse agents. Spaces appear in the sidebar of the Muse interface and can be opened to work on projects such as vacation planning, recipe collection, holiday gift tracking, wedding planning, or DIY projects. Users can describe an idea in chat, and Muse builds a space that integrates with third‑party apps, automatically updating data like concert dates or price changes. The feature follows last week’s Muse Voice debut and follows the 10‑million‑user milestone for Muse agents.

💡 By embedding project‑focused AI assistants into everyday workflows, Meta expands the practical use of conversational agents beyond simple queries, positioning itself as a leader in integrated, context‑aware productivity tools.

Raycast combines voice dictation and screen awareness in new AI feature
9to5mac.com · Nov 6

Raycast combines voice dictation and screen awareness in new AI feature

Raycast has expanded its Screen Awareness suite with a new “Dictate to AI” command that lets users speak requests while the app automatically supplies contextual information from the active window. The feature builds on the August launch of Screen Awareness, which already allowed users to feed app or window content into AI Chat without manual copying or screenshots. In a demo video, Raycast co‑founder Pedro Duarte shows how voice input is paired with real‑time screen context, eliminating the need for users to describe what they see. The update arrives alongside other enhancements: Raycast Notes now supports images and unlimited free notes, the macOS client gains faster search and lower memory usage, iOS receives an AI‑powered keyboard and better Cloud Sync, and Windows gets new commands for sending individual elements or whole windows to AI Chat. The changes are part of version 2.11.

💡 Voice‑driven, context‑aware AI shortcuts streamline multitasking, letting professionals act on screen content without interrupting their workflow.