Good morning. Today’s threads pull in opposite directions: Google is shipping a laptop-friendly multimodal model with a novel encoder-free design, while regulators and rivals are busy redrawing the rules around what AI companies can take, build, and deploy. Plus a worrying signal from UC Berkeley about what happens when CS students lean on LLMs through the semester and then hit an in-person exam.
Google’s Gemma 4 12B ditches the vision encoder. Google DeepMind released Gemma 4 12B under Apache 2.0, sized to fit on a 16GB consumer GPU and filling the gap between the 4B edge model and the 26B MoE. The architectural twist: vision and audio inputs skip a dedicated encoder and run through the LLM backbone via a single matrix multiplication with positional embeddings and normalizations. Reception on Hacker News is split — one commenter ran the Q4 quant through a Minesweeper vibe-coding test with decent results, while another reported Gemma’s image understanding losing badly to a Qwen 3.5 0.8B model that’s 7% the size. Several commenters also questioned whether the “encoder-free” framing is really accurate, since there’s still encoding happening, just stripped down.
OpenAI proposes a democratic governance blueprint for frontier AI. OpenAI published a frontier safety blueprint laying out how it thinks advanced AI should be governed through democratic oversight. Details are thin in the post itself, but the timing — alongside Trump’s executive order and the biosecurity letter below — suggests the major labs are coordinating their policy pitch heading into the next regulatory cycle.
The biosecurity ask: mandatory DNA screening. OpenAI, Anthropic, Google DeepMind, and Microsoft co-signed a letter to Congress arguing that AI is eroding the knowledge barriers that used to stop bad actors from synthesizing pathogens, and that voluntary screening at synthetic DNA vendors isn’t enough. The letter notes that AI tools can already help users find unscreened suppliers and obscure dangerous orders, and points to the 2017 horsepox reconstitution — done for around $100,000 — as a warning shot. The International Gene Synthesis Consortium screens orders voluntarily, but the signatories want a federal mandate.
Trump’s AI executive order, in its final form. Wired has the backstory on the scaled-back order Trump signed: the federal government gets advance access to frontier models 30 days before release (down from 90 in earlier drafts), within a voluntary framework focused on cyberattack risk to critical infrastructure. Chief of staff Susie Wiles reportedly revived the order over resistance from former AI czar David Sacks, and Anthropic publicly backed it. The order also clears Treasury Secretary Bessent to start cross-border AI framework talks with China.
Microsoft and OpenAI move from partners to rivals. Building on yesterday’s Build 2026 announcements, The Verge frames the bigger picture: Microsoft’s in-house reasoning model, coding model, Scout assistant, super app, and security agents add up to a deliberate break from OpenAI dependence after the April separation. AI chief Mustafa Suleyman put it bluntly — Microsoft has to “prove that we can do everything that we need to from the ground up.” The cloud partnership continues, but the strategic posture has flipped from collaborator to competitor.
UK forces Google to give publishers an AI Search opt-out. The Competition and Markets Authority is requiring Google to add a Search Console toggle letting publishers exclude their content from AI Overviews and AI Mode, and block its use in model fine-tuning, without affecting traditional search rankings. The Verge and TechCrunch both note Google will provide impression metrics so publishers can see what they’d give up before opting out, with UK testing first and a global rollout to follow. Google also has to provide proper inline attribution when content does appear.
Meta Business Agent goes worldwide on WhatsApp. After nearly two years of testing in India and Mexico, Meta is rolling out its Business Agent globally. The AI handles customer support, product recommendations, appointment booking, and lead qualification, with Shopify and Zendesk integrations on the way. Pricing splits between WhatsApp Business Premium tiers for smaller merchants and token-based pricing for enterprises.
Berkeley CS failure rates spike, and AI is part of it. The Daily Cal reports that 35.3% of CS 10 students and 10.6% of CS 61A students received F’s last spring, well above the department’s 7% guideline for combined D/F grades. Instructor Dan Garcia points to AI-assisted cheating — nearly 30 students were caught on take-home tests in CS 10 alone — plus students using LLMs to coast through homework and then failing in-person exams. HN commenters flag a second factor the article only mentions in passing: 1,300+ UC faculty have signed a petition to reinstate SAT/ACT requirements, arguing six years of test-optional admissions has hollowed out math preparation.
That’s the briefing. Watch the CMA opt-out rollout closely — if UK publishers actually take Google up on it in meaningful numbers, the impression data Google’s required to share will tell us a lot about what AI Search is really worth.