Apple just fired Siri

In This Issue

Big Move - Apple x Google

Apple Is Paying Google $1 Billion a Year to Power Siri. The Old Siri Is Gone.

Apple and Google have signed a multi-year deal worth an estimated $1 billion per year. The deal hands Gemini the job of powering Siri's intelligence layer starting with iOS 27, which Apple is expected to announce at WWDC in June 2026. The move ends 15 years of Apple trying and mostly failing to build a world-class AI assistant on its own.

This is a remarkable reversal. Apple spent years positioning its on-device AI as a privacy-first advantage over Google. Paying Google $1 billion a year to power that same AI is a quiet admission that building frontier models in-house is a harder problem than even Apple can solve alone.

China AI - DeepSeek

DeepSeek V4 Is Out. 1.6 Trillion Parameters. Open Source. Claims to Beat Gemini.

China's DeepSeek released V4 on April 23 with two variants: V4-Pro with 1.6 trillion parameters and V4-Flash with 284 billion. Both are open-source, both score near the top of major benchmarks, and both are available to run locally or via API at a fraction of what OpenAI charges. CNN, DW, and Times of India all confirmed the release within 24 hours.

Every time DeepSeek ships, the same question comes back: why is a Chinese lab releasing the best open-source models in the world for free while US labs charge premium prices for closed ones? The answer is strategy. DeepSeek is commoditising the model layer to accelerate adoption of Chinese AI infrastructure globally.

Open Source AI - Meta

Meta Held Its First AI Conference and Shipped Two Models That Beat GPT-4o. Both Free.

Meta's first-ever LlamaCon ran from April 22 to 24 and ended with two major open-weight model releases. Llama 4 Scout is designed for efficiency: 109 billion total parameters but only 17 billion active at a time, which means it runs comfortably on a single GPU while handling a context window of up to 10 million tokens. Llama 4 Maverick is the reasoning powerhouse: 400 billion total parameters, 17 billion active, multimodal, and in head-to-head testing it scores above GPT-4o and Gemini 2.0 Flash.

Meta's bet is clear. Give away the best open models, dominate the developer ecosystem, and monetise through ads on the platforms people build with those models. It is the same playbook Android used against iOS, applied to AI.

Creator Tool - HeyGen

HeyGen Avatar V: Record 15 Seconds of Yourself. Get an AI Twin That Makes Videos Forever.

HeyGen launched Avatar V on April 8 and it is genuinely different from every AI avatar tool that came before it. The old way required 2 to 5 minutes of recorded footage and produced something that looked obviously artificial. Avatar V needs just 15 seconds of video, and the output is a photorealistic digital twin that maintains consistent identity across multiple camera angles, lighting conditions, and video lengths up to 60 minutes.

The practical use case here is obvious for indie creators. Record your avatar once. Use it to make every talking-head video you need for the next year. No camera setup, no reshoots, no editing day. One 15-second recording and you are done.