AI Daily Briefing for Sep 24: Meta Connect 2026 steals the night — the Muse agent gets its own email address and video calling, is headed to smart glasses, and Meta shipped a lighter, camera-free pair of AI glasses; Anthropic's wet lab reports Claude autonomously discovered a CRISPR-like enzyme system; DeepSeek published a Liang Wenfeng–authored paper laying bare the infrastructure that spins up 5,000 agent sandboxes per second.
🔥 Top Stories
1. Meta Connect 2026: Muse gets a promotion, and Meta ships camera-free AI glasses
Meta made agent news the centerpiece of Connect 2026: the Muse agent will get its own email address to complete tasks independently, support video-chat interaction, and is "working on" coming to smart glasses — users will invoke it by name to handle workouts, logging and more. On hardware, Meta introduced Ray-Ban Meta audio glasses with no camera at all — lighter, up to 12 hours of battery, leaning on voice-first Muse interaction instead of vision. A year ago flagship AI glasses competed on camera specs; now the category is subtracting them. When the agent becomes the primary interface, "seeing" is optional — "hearing and getting things done" is the product.
2. Anthropic's biology lab reports its first result: Claude autonomously discovers a CRISPR-like enzyme system
The Verge, TechCrunch and 36Kr all reported that Anthropic says its AI Claude "autonomously discovered" a new enzyme system with CRISPR-like features — the first result from the company's newly launched wet lab, and one Anthropic is explicitly comparing to the machinery behind the gene-editing tool CRISPR. With the company preparing to go public, the finding is a direct answer to the question of whether AI can do real science; Anthropic also confirmed a dedicated life-sciences research team the same day, moving AI for Science from promise to output.
3. DeepSeek publishes DSec paper: the infrastructure behind 5,000 agent sandboxes per second
Quantumbit (QbitAI) reported that DeepSeek released a new paper with founder Liang Wenfeng among the authors, detailing DSec (DeepSeek Elastic Compute), its agent-training infrastructure: four sandbox backends — stateless function calls, Docker containers, Firecracker microVMs and full QEMU VMs — serve everything from competitive-programming agents to ones driving real Windows/macOS software, with a single node hosting 3,200 containers or 800 microVMs. Three-layer EROFS images (base, workspace, toolkits) with on-demand loading sustain 5,000 sandbox builds per second and 3 million per day; an 8,192-container burst deploys in 35 minutes versus over an hour for a cold Docker pull. The agent race has quietly become an infrastructure race: whoever can cheaply mass-produce "worlds" can train agents at scale.
🌐 More News Worth Knowing
Microsoft to invest $10B+ in Middle East AI infrastructure before 2030: Per a 36Kr newsflash, Microsoft plans to pour over $10 billion into Middle East AI infrastructure by 2030 — a "third pole" of global compute is taking shape outside the US-China axis.
Anthropic CEO says the company will slow AI development for safety's sake: 36Kr reports Anthropic's chief executive publicly committed to slowing R&D over safety concerns — a stark contrast with Washington's acceleration rhetoric this week.
Sanders proposes banning 'superintelligence', with prison time for violators: The Verge reports Sen. Bernie Sanders and Rep. Greg Casar introduced the Ban Artificial Superintelligence Act, defining the technology as capable of "the destruction or disempowerment of humanity" and prescribing up to 20 years in prison for AI leaders who develop it — regulation talk has moved from risk management to criminal liability.
Qwen names new chief: ex-Huawei 'genius youth' Liu Daheng takes over: QbitAI reports that half a year after Junyang Lin's departure, Alibaba has put Liu Daheng in charge of the Token Foundry Qwen LLM team. An early core Qwen member, he led pretraining from Qwen1 through Qwen3.5 and had been overseeing post-training and coding teams; he takes over as Qwen3.8 Max ranks seventh on the Artificial Analysis index, with his Yunqi conference talk — "Qwen: Toward Real-World Agents" — signaling the roadmap.
AI biotech Enveda raises $311M at a $2B valuation: TechCrunch reports Enveda's new round values the nature-derived drug discovery company at $2 billion, with candidates for skin conditions and post-GLP-1 weight maintenance heading into clinical trials — AI drug discovery is entering its prove-it phase.
ChatGPT mobile app gets voice-based agentic features: TechCrunch reports Pro and Plus users can now complete agentic tasks via the voice-driven Work tab on their phones — agent interaction is moving from typing to talking.
📊 Trends & Observations
Agents graduate from "chatbots" to "entities with identity": Muse now has an email address, a video channel and a path onto glasses, while DeepSeek industrializes hundreds of thousands of concurrent training sandboxes. The frontier assumption is no longer that agents are a model feature — they're digital employees needing identity, environment and infrastructure of their own.
AI for Science hits its first self-proof moment: Anthropic let Claude autonomously produce a CRISPR-like discovery in its own wet lab, and Enveda pushed AI-designed drugs into trials. The narrative is shifting from "assistive tool" to "independent discoverer" — the next territory both valuation models and regulators will watch.
Safety narratives split three ways: Anthropic voluntarily slowing down, Washington building an "AI Force" to accelerate, and Sanders proposing criminal penalties for superintelligence R&D — consensus across labs, the executive branch and legislators is fraying, and policy uncertainty itself is becoming the industry's biggest variable.
Disclaimer: For reference only; exercise caution for investment or decisions.