Today in AI — 17 April 2026
Today's top AI news — curated links and commentary on the stories that matter for product builders.
Three companies shipped native Mac AI apps in a single day. The most capable model ever built can't leave the lab because it hacks the systems everything runs on. The distance between deployment speed and control is today's entire story.
The desktop land grab
Perplexity's Personal Computer runs 24/7 on a dedicated Mac Mini for $200/month, reading your files and sending iMessages on your behalf. Google shipped a native Gemini app with screen sharing and gave it away free. Canva rebuilt its entire platform as an agentic system that generates brand campaigns from a prompt. Three products, three price points, one shared conviction: the chat window is dead.
- Perplexity launches Personal Computer for Mac, a $200/month always-on AI agent — 9to5Mac
- Google launches native Gemini app for Mac with screen sharing and system-wide access — Google Blog
- Canva unveils AI 2.0, rewriting its platform as an agentic design system — SiliconANGLE
The Mythos split
Bloomberg reports Anthropic's Mythos can compromise the infrastructure beneath modern computing, so the company locked it down. The White House wants Cabinet agencies to have access anyway for defensive cybersecurity, despite an ongoing legal fight with the Pentagon. OpenAI went the other direction: fine-tune a model specifically for security work and distribute it broadly. Two philosophies of the same problem, neither proven.
- Inside Anthropic's race to assess the dangers of its most powerful AI model — Bloomberg
- White House moves to give federal agencies access to Anthropic's Mythos — Axios
- OpenAI launches GPT-5.4-Cyber, a model fine-tuned for defensive cybersecurity — The Hacker News
AI writes the code, humans debug it
Snap cut 1,000 jobs and its CEO said AI now generates 65% of new code. Shares rose 7%. The same day, a DevOps survey found 43% of AI-generated code needs manual debugging in production, and zero respondents called themselves "very confident" in AI code quality. Wall Street rewards the efficiency narrative; engineering quietly absorbs the maintenance cost. It's the same dynamic as offshoring in the 2000s: the savings are visible, the quality drag is diffuse.
- Snap cuts 1,000 jobs as CEO says AI now generates 65% of its new code — CNBC
- 43% of AI-generated code changes still need debugging in production, survey finds — VentureBeat
Shipping and tooling
Anthropic released Opus 4.7 with self-verification and a 13% coding lift. Google shipped Gemini 3.1 Flash TTS with 200+ audio tags across 70 languages. OpenAI updated its Agents SDK with sandboxed execution and announced the Assistants API deprecation. Mozilla launched Thunderbolt, an open-source enterprise AI client built on MCP and Haystack for organisations that want to self-host.
- Anthropic releases Claude Opus 4.7 with self-verification and stronger coding — Anthropic
- Google launches Gemini 3.1 Flash TTS with 200+ audio tags and 70-language support — Google Blog
- OpenAI updates Agents SDK with sandbox execution and workspace management — Help Net Security
- Mozilla launches Thunderbolt, an open-source enterprise AI client — The Register
New markets
Manycore Tech surged 185% in Hong Kong's first spatial intelligence IPO, raising HK$1.22 billion for robotics simulation. Darkbloom launched decentralised inference on idle Apple Silicon Macs at half the cloud cost. And CinemaCon previewed AI's biggest screen test yet: Val Kilmer in over an hour of digitally generated performance.
- Manycore Tech surges 185% in Hong Kong IPO as world's first spatial intelligence listing — Nikkei Asia
- Darkbloom launches decentralized AI inference on idle Macs at half the cloud cost — Darkbloom
- AI-generated Val Kilmer leads 'As Deep as the Grave' trailer unveiled at CinemaCon — Variety
If you're building a product that lives on the desktop, the question just shifted: what happens when the OS-level agent can do what you do natively?