Oct 1 · AI Daily

According to Sensor Tower data, Meta's AI assistant Muse reached 5 million downloads in the US on Sep 30, roughly three weeks after launch (taking 22 days), and held the top spot on the App Store free chart for 12 straight days. That is notably faster than ChatGPT's 56 days, Grok's 103 days, and Claude's 492 days.

Apple's Invites app updated to version 1.12.0. Shared albums now support full-resolution photo sharing, and anyone with a link can join and upload photos, including people on non-Apple devices. The update expands Image Playground with photorealistic background generation and streamlines email verification, letting guests join an event by clicking a link in their inbox without entering a verification code. Creating events requires an iCloud+ subscription (starting at 50GB for $0.99 a month), while responding to a link is free.

A dating app rolled out by the Singapore government uses the Gale-Shapley stable marriage algorithm for matching. The algorithm is designed to prevent pairs where both individuals would prefer each other over their currently assigned partners.

Hamel Husain tested the new build_eval and hill-climb commands added by Anthropic to the Claude Code claude-api plugin, running a live demo alongside Isaac Flath.

The cache diagnostics feature in the Claude API has exited beta and no longer requires the cache-diagnosis-2026-04-07 beta header. You can enable it by including a diagnostics object in Messages requests; requests that keep sending the legacy header will behave unchanged. The POST /v1/messages response now always returns a diagnostics field, which is null when not enabled.

A new "Photo Face" step has appeared in the setup flow for Apple's upcoming smart home device, using iCloud Photos to let the device display the user's face right away. During setup, the iPhone first sends a few full-resolution "cold start" photos so faces show up immediately, and the device then downloads the rest from iCloud. The feature will also be used for iPhone Duo's StandBy mode.

Latent Space published two interviews from OpenAI DevDay. Computer Use lead Ari Weinstein said Computer Use is already a "180-degree difference" from just a few months ago.

Google is developing a security review feature for Google AI Studio.

"Apply secure coding practices".

The new mode may appear in the mode selector alongside Plan mode, which is also in development.

Could it be powered by Gemini 4rgon? 👀

Video generation and coding agents are defining a new class of AI workloads. Individuals are efficiently consuming massive amounts of compute in their day-to-day work.

fal CTO @gorkem spoke with MTS at our GenMedia Conference last week about "token market fit."

Mark Cuban's approach: use AI to build a business simulator to test hypotheses.

Founders can now generate draft plans, cost structures, vendor lists, and patent outlines.

This closes the gap between inspiration and the first serious feasibility test.

Via the "Alex Kantrowitz" YouTube channel (link in comments)

Nathan Lambert posted that he was glad to see Google surprise people with Gemini 4. He believes having more labs on the frontier is good for consumers (competition) and the world (less concentration of power), and said he looks forward to seeing how it performs in real-world scenarios.

The Apple iPad mini 8 will introduce an IP water-resistance rating for the first time, powered by an A20 Pro chip built on TSMC's 2nm process. CPU performance is up about 1.5x and GPU performance about 2.5x compared to the A17 Pro, paired with dual 16-core Neural Engines (32 cores total).

Google parent Alphabet released its new model, Gemini 4 Argon, focused on cybersecurity capabilities. It can autonomously discover, verify, and patch critical software vulnerabilities, and is currently available only to select cybersecurity partners through the Fairwind security program.

Doubao added booking for flights and train tickets as well as ride-hailing and navigation, partnering with Flight Master, High-Speed Rail Master, Caocao Mobility, Baidu Maps, and others. Baseten partnered with OpenAI, letting enterprise users call open-source models like Kimi K3 and GLM 5.3 via Codex or the Responses API. Kimi K3 features 2.8 trillion parameters and a 1-million-token context window.

Meta and collaborating institutions published a paper on Context Language Models (CLMs), enabling language models to treat context like files and manage it autonomously and natively, without an external harness.

Inworld has acquired Ultravox. Inworld is a research lab that trains its own voice models and offers LLM inference APIs, while Ultravox is a platform developers use to build real-time voice agents. The combined company brings both voice capabilities and the runtime platform for agents under one roof.

The money in AI software is concentrated at the top.

Just 224 million high-earning knowledge workers—a quarter of the total—account for more than $25 trillion of a $41 trillion market.

In other words, the top 25% of knowledge workers by income account for nearly two-thirds of the $41 trillion opportunity. And most of them are in North America and Europe.

• ARK research

Omni-IO Skills

Make your agents natively all-around.

paper: https://huggingface.co/papers/2609.31847

LongLive-Plug

Universal distillation for video generation

Paper: https://huggingface.co/papers/2609.38154

At BlueHat Asia 2026, security researcher Johann Rehberger reported a SQL Server privilege escalation vulnerability in Microsoft SQL Server Management Studio's SQL Copilot, tracked as CVE-2026-65669. Microsoft rated it critical and advised users to keep their installations updated.