AI drops to watch
Reports claim a wave of multimodal and high‑context models is moving fast — mentions include GPT‑5.4 with 1M+ token context and autonomous workflows, Apple rebuilding Siri on Google’s 1.2T Gemini, and open‑source 4K video generation from Lightricks ( ). Industry takes also flagged Jensen Huang claiming AGI and Cloudflare touting a 100x AI agent speed boost — lots of big claims to parse for real‑world impact (x.com).
OpenAI’s GPT‑5.4 is listed with a 1.05 million‑token context window in the official model docs and OpenAI’s product post, and the API applies a pricing breakpoint where prompts over 272K input tokens are charged at higher multipliers. (developers.openai.com) OpenAI’s announcement frames GPT‑5.4 as a single “frontier” model that consolidates agentic workflows and stronger coding/reasoning capabilities, and the company also published smaller variants (GPT‑5.4 Mini and Nano) for lower‑cost use cases. (openai.com) Multiple outlets report Apple struck a multi‑hundred‑million to ~$1 billion‑per‑year agreement to license a custom 1.2‑trillion‑parameter Gemini variant to power a Siri overhaul targeted for iOS 26.4 in spring 2026, a deal first reported by Bloomberg and covered by CNBC and CNET. (bloomberg.com) Lightricks’ open‑source LTX‑2 release (weights and training code published on GitHub) advertises native 4K generation at up to 50 fps with synchronized audio, production‑length clips up to ~10 seconds, and claims the model can run on consumer GPUs. (github.com) Cloudflare launched “Dynamic Workers” (open beta announced March 24, 2026) that run AI‑generated code in lightweight V8 isolates instead of containers, claiming millisecond startup times and up to ~100× faster cold starts versus Linux containers; Cloudflare’s blog and vendor coverage also list a per‑Worker loader price of $0.002/day during rollout. (blog.cloudflare.com) Nvidia CEO Jensen Huang told Lex Fridman on March 23, 2026’s podcast that “I think we’ve achieved AGI,” a remark published in the episode transcript and widely reported, and Huang immediately qualified the statement later in the conversation. (lexfridman.com)
