Google’s New AI Just Broke The AI Speed Limit: DiffusionGemma

This video covers major AI developments, headlined by Google’s release of Diffusion Gemma, a mixture-of-experts model that generates text in parallel on a 256-token canvas, achieving up to 4× faster speeds on local GPUs while targeting structured tasks like code infilling and OCR rather than top-tier writing quality. It also showcases Gemini 3.5’s near real-time speech-to-speech translation for 70+ languages, alongside Google Meet’s expansion to 2,000+ language combinations with SynthID watermarking. In the open-source space, Xiaomi introduced Mimo Code, a memory-first terminal coding assistant with strong but self-reported benchmark claims that fall short of OpenAI’s Codex CLI on independent leaderboards. Finally, OpenAI has confidentially filed for a US IPO targeting a $1 trillion valuation, backed by 900M weekly users and $2B monthly revenue, though it does not expect profitability until 2030 after clearing legal and partnership hurdles with Microsoft.

• ▶ 0:25 Diffusion Gemma differs from standard chatbots by generating text as a whole chunk on a 256-token canvas, cleaning up a messy block of placeholder text over several rounds rather than writing one token at a time left-to-right.
• ▶ 1:56 It delivers up to 4× faster text generation on dedicated GPUs—over 1,000 tokens per second on an Nvidia H100 and over 700 on an RTX 5090—and is released under Apache 2.0 as a 26B mixture-of-experts model.
• ▶ 2:19 Google positions standard Gemma 4 as better for the highest-quality writing, while Diffusion Gemma targets speed, local use, and structured tasks like code infilling, OCR, document parsing, and agent workflows—tasks where the whole answer matters.
• ▶ 3:05 DiffusionGemma's speedup is most valuable for local and low-concurrency scenarios, where GPU underutilization is a bigger bottleneck than in large cloud systems.
• ▶ 4:03 Gemini 3.5 Live Translate delivers near real-time speech-to-speech translation in 70+ languages, listening as the speaker talks and preserving tone, pacing, and rhythm.
• ▶ 5:44 Google Meet translation expands from 5 languages and English-centric routing to 70+ languages and 2,000+ language combinations, with AI-generated audio watermarked via SynthID.
• ▶ 6:42 Xiaomi released Mimo Code v0.1.0, an open-source, terminal-based coding assistant (MIT license) with one-command install on macOS/Linux and npm on Windows, including limited-time free access to MIMO v2.5.
• ▶ 7:38 Mimo Code's core design is memory-first: it uses persistent memory MD files, session checkpoints, SQLite FTS5 full-text search, and a dedicated checkpoint writer sub-agent so long sessions don't lose context.
• ▶ 8:30 Xiaomi reports strong benchmark results (e.g., SWE-bench Verified 82% vs 79% for Claude Code) but these are self-reported and not independently verified—and on the official Terminal Bench 2 leaderboard, OpenAI Codex CLI scores higher at ~82.2% vs Mimo's 73%.
• ▶ 12:22 OpenAI has confidentially filed for a US IPO, targeting a valuation of up to $1 trillion with a possible debut as early as September.
• ▶ 13:20 OpenAI's financial momentum is strong—900M+ weekly ChatGPT users, $2B monthly revenue—but the company told investors it does not expect to be profitable until 2030.
• ▶ 13:58 Key structural and legal hurdles were cleared: a jury ruled against Elon Musk's lawsuit, and OpenAI renegotiated its Microsoft partnership to gain more flexibility with rivals like Amazon and Google.

• ▶ 0:02 Google's Diffusion Gemma Reveal (0:02 - 3:05) - - Google announces Diffusion Gemma, explains how diffusion text generation works, covers architecture, speed, and quality trade-offs.
• ▶ 3:05 Concurrency, Live Translate, and Meet (3:05 - 6:42) - - Discusses local versus cloud concurrency, then Gemini 3.5 Live Translate and Google Meet's upgraded translation features.
• ▶ 6:42 Xiaomi Mimo Code (6:42 - 12:20) - - Xiaomi's Mimo Code enters AI coding with strong benchmark results, new model specs, and aggressive pricing.
• ▶ 12:22 Market Moves and Outro (12:22 - 14:45) - - OpenAI, Anthropic, and SpaceX move toward public markets, OpenAI's financials and Microsoft deal, followed by the sign-off and call to action.