Video: Kling AI
This past week was all about pushing hard on both “bigger” and “smarter” fronts—major infra deals, new context‑length records, plug‑and‑play reasoning models, and UI‑level features that let AI actually see and remember - and one ChatGPT o3, yes it mind-blowing
Every corner of the ecosystem moved: chipmakers committing half a trillion dollars, OpenAI doubling down on massive‑context and image‑thinking variants, Google and Microsoft peeling back paywalls on live camera/screen assist, IBM rolling out next‑gen Granite models, and xAI’s Grok finally getting a canvas editor and persistent memory.
Infrastructure & Strategy
-
Nvidia’s $500 Billion On‑Shore AI Build‑Out Announced April 14, Nvidia will partner with TSMC, Foxconn, and Wistron to produce up to $500 billion of AI servers and Blackwell GPUs in Texas and Arizona over four years, marking a major shift to U.S. chip and supercomputer manufacturing Reuters.
-
Nvidia’s $5.5 B Write‑Down on H20 Chips for China On April 15, Nvidia said it will take a $5.5 billion charge after new U.S. export restrictions on its H20 AI chip to China kicked in, highlighting how trade controls can ripple right through AI supply chains Reuters.
Model Releases
-
GPT‑4.1 Family Goes Live On April 14, OpenAI launched GPT‑4.1 (alongside Mini and Nano), boosting context windows to 1 million tokens and significantly improving coding, instruction‑following, and long‑context reasoning in its API The Verge.
-
o3 & o4‑mini: “Image‑Thinking” Reasoners April 16 saw OpenAI drop o3 (its most advanced reasoning model) and o4‑mini (a fast, cost‑efficient lean variant)—both can now use vision inputs, web browsing, Python, and more in a single pass The Verge.
-
IBM Granite 3.3 On April 16, IBM released Granite 3.3, including an 8B speech model (Granite Speech 3.3 8B) with ASR/AST, plus new LoRA adapters for retrieval‑augmented generation—all with a 128K token context window and fill‑in‑the‑middle support IBM - United StatesHugging Face.
-
Qwen 3 Preview from Alibaba On April 14, Alibaba posted on LinkedIn that it plans to release Qwen 3 (8B & MoE‑15B variants) by the end of April, aiming to deepen its global AI model lineup LinkedIn.
-
Kling 2.0: From Vision to Stunning Video
fal.ai’s new model turns images and text into high‑fidelity video with advanced editing.
-
Gemini 2.5 Flash: Google’s hybrid “toggle thinking” model for faster, cheaper deep reasoning on demand.
-
TxGemma: DeepMind’s open suite for drug‑discovery tasks—molecular predictions & clinical‑trial forecasting.
-
Veo 2 Joins Gemini: Text → 4K Video in Seconds
- Generate eight‑second, 4K clips and animate images via Whisk Animate inside Gemini Advanced.
Agentic & Platform Highlights
-
Gemini Live Goes Free on Android Starting April 16, Google opened up its Gemini Live camera & screenshare helper to all Android users—no subscription needed—so you can show Gemini your screen or feed it live camera views for real‑time Q&A The Verge.
-
Copilot Vision in Edge, Now Free On April 16, Microsoft announced Copilot Vision—its AI assistant “sees” your Edge browser screen—as a free, opt‑in feature for all Edge users, removing the prior Pro paywall The Verge.
-
xAI Grok Studio: In‑Chat App & Doc Canvas xAI rolled out Grok Studio on April 16, providing a Figma‑style canvas inside the chat for building documents and mini‑apps with Grok, injecting a low‑code UI layer into conversation Home.
-
xAI Grok’s New Memory Feature On April 16–17, Grok gained a persistent memory beta—your past chats now inform future responses, so it can recall your preferences and save you from re‑explaining every time TechCrunch.
-
Codex CLI Open‑Source: OpenAI’s first terminal‑coding agent release since 2019, plus two new compact reasoning models.
