Good morning, welcome to this week’s Lore Brief, your 3-minute brief of the most important moves in AI and tech.
I published a piece today on why I'm launching Building America — a show about the people building America's future in AI, defense, and government.
Episode 1 with Scott Kupor is live → [YouTube link].
Google launches Nano Banana 2 → The updated Gemini AI image generating model delivers Pro-level quality at Flash speeds, with enhanced text rendering, precise instruction following, and real-time web sourcing for realistic outputs. It solves subject consistency for reliable characters across images and costs 2x less than Nano Banana Pro. Read more here →
Perplexity releases “Computer” for full-cycle AI workflows → Orchestrates 19 models for research, coding, deployment, and management. Features persistent memory, connectors, and usage-based pricing with credits for Max subscribers. Read more here →
Anthropic exposes model distillation by Chinese labs → DeepSeek, Moonshot AI, and MiniMax used 24k fake accounts for 16M queries to clone Claude. Anthropic urges policy fixes, but draws ridicule on X over its own alleged stolen data use in training. Read more here →
Dose of optimism:
Great new Google maps feature:
Useful AI image workflow:
Impressive choice of “actors” in this AI generated ad:
Claude Code speeds up COBOL updates → It handles the tricky parts like mapping dependencies and spotting risks, turning what used to take years into just a few months. Right after the blog post, IBM's stock took a 13% hit in one day. This shows how AI coding is shaking up old-school tech giants. Read more here →
Inception AI releases Mercury 2 as the fastest diffusion LLM → Diffusion LLMs refine outputs in parallel steps unlike normal LLM models that generate text sequentially one token at a time. This enables 5x faster real-time reasoning for agents and voice AI. It hits 1,009 tokens/sec on Blackwell GPUs, tops benchmarks, and supports 128K context. Read more here →
Google’s TranslateGemma 4B model enables offline browser translation → Handles 55 languages on WebGPU with Transformers.js, fully client-side and private. Read more here →
Cloudflare recreates Next.js for edge deployment → Ports framework to Workers sans Node.js in one week, unlocks serverless potential. Read more here →
NVIDIA shares SONIC for humanoid control → 42M model generates motions from VR, video, text, audio, or VLMs. Trained on 100M frames, transfers zero-shot to real robots. Read more here →
Qwen 3.5 series shrinks models without losing smarts → 35B-A3B tops 235B predecessor via hybrid attention, sparse experts, and RL. Boosts efficiency for long-context agents. Read more here →
