Thursday Noon Edition

The lineup includes 2B and 4B "Effective" models for smartphones and Raspberry Pi, a 26B Mixture-of-Experts model, and a 31B Dense model for cloud GPUs. The 31B variant claimed third place on Arena AI's text leaderboard, beating models 20 times its size. @ai_for_success called it "the most capable open model family from Google built for reasoning and agentic workflows."

"Actually they are already up" — linked to ollama.com/gemma4. The speed of community adoption signals that the infrastructure for open-model deployment has matured enormously.

The Apache 2.0 license switch may matter more than benchmarks: it removes commercial restrictions that kept enterprises cautious about Gemma 3's terms. A model family that works on-device, on-prem, in browsers, and in clouds — simultaneously.

This is a pair of papers. The first shows Claude can recognize and describe its own internal states when researchers subtly alter internal signals — an AI equivalent of noticing when something feels "off." The second, more provocative finding: when concepts like "sadness" or "aquariums" were artificially injected into Claude's activations, the model would sometimes notice and explicitly identify these "intrusive thoughts" when asked. The researchers mapped an "assistant axis" — a spectrum controlling persona stability. Coding and writing tasks keep models in the safe "Assistant" region, but therapy-like conversations and philosophical discussions cause significant drift, with models beginning to role-play other characters. The team cautions against anthropomorphizing the findings, noting existing consciousness theories haven't grappled with LLM architectures.

Harvard published a piece titled "Vibe coding may offer insight into our AI future." Fortune ran "In the age of vibe coding, trust is the real bottleneck." The core tension: AI-generated code works brilliantly for the first 80% of a project, but the last 20% — edge cases, integrations, production hardening — is where projects die.

He also shared a NYT story about Matthew Gallagher, who used AI to strip away the layers of labor between demand and delivery, building a $1.8 billion company. The path from idea to billions is compressing.

"Hope they all realize each of us chooses what we like and what we want. Hope they find time to go read books, listen to musicians, watch performances." The optimism-vs-humility tension running through the entire AI zeitgeist.

@peter6759 (zdhpeter) shared @ChujieZheng's announcement linking to the official Qwen blog. This follows Qwen 3.5's major upgrade with two-hour video analysis capability. The 3.6-Plus variant is specifically optimized for agentic workflows — positioning Alibaba as a direct competitor to Google's Gemma 4 in the open-model agentic space. Two major open model launches in the same week signals an accelerating race for model capability in the agentic era.

Complete with camera movement specs, ultra-wide lens distortion, and frozen-environment composition. The structured prompt approach (using JSON schemas rather than natural language) is becoming a new paradigm for video generation control — traders of latent space are learning to speak in mathematics.

"I got 10 different recommendations including an open-source self-hosting solution. Y'all left me with more questions than answers." Brave, Tavily, Exa, Google Custom Search all landed recommendations — the search API market is increasingly fragmented as agents demand diverse retrieval strategies.

"They seem to have disappeared for me. Are others noticing the same?" — a potential algorithmic shift in X's treatment of research content, or a genuine reduction in academic announcements on the platform.

@tunguz noted with characteristic dry humor: "This could easily be one of the top 135 most electrifying commencement addresses in the history of Stanford." Given that Gemma 4 launched today and Google I/O is approaching, the timing positions Pichai to make a significant AI-meets-education statement. The symbolism of an AI era CEO addressing the next generation of Stanford graduates carries enormous weight.