Sarmadi AI Digest June 4, 2026 Updated 6:55 AM CT Today Archive Topics Saved Subscribe RSS

Gemma 4 12B opens, Alphabet's $85B prices in, Microsoft and OpenAI now openly fight

Google released Gemma 4 12B — a unified encoder-free multimodal model designed to fit on any laptop with 16GB of RAM — the same day Alphabet's record $85B equity raise priced in. The post-Build read sharpened: The Verge documented the Microsoft–OpenAI breakup as an open fight, and Stratechery's Nadella interview frames Microsoft's bet as owning the platform end-to-end. Two operator-side signals tighten the agentic-economy picture — Uber's $1,500/month per-employee AI limit is a useful pricing benchmark (Simon Willison's 504-point HN breakdown crystallizes it), and Berkeley CS classes are seeing failing grades and dwindling math skills correlated with AI use. OpenAI and Anthropic publicly co-signed a letter urging legislation against AI-developed biological weapons. Research-side, RAMP names the gap between agent benchmarks and production reality, while Economy of Minds models Hayek-style decentralized multi-agent coordination.

11 papers 22 news 9 sources ← Latest

News

15 items

Google's megaday: Gemma 4 12B and $85B raise

Gemma 4 12B is a unified encoder-free multimodal model engineered to run on a 16GB laptop — Google's most aggressive open-weights release. Alphabet's $85B raise (up from the $80B announced) priced in alongside. Lovable signing a multi-year deal that 5x's its Google Cloud usage rounds out the day's wins.

News Hacker News

Gemma 4 12B: A unified, encoder-free multimodal model

Google released Gemma 4 12B — a unified encoder-free multimodal model with new tokenization, designed to run on any 16GB-RAM laptop (863 HN points).

params 12Bmin RAM 16GB
Why it matters
  • Largest, most capable open-weights release from Google to date.
  • Encoder-free unified-modality architecture validates the direction the rest of the open frontier is moving in.
  • Drops local multimodal AI into 16GB consumer laptops — direct cost-floor reset for SMB builders.
open-weightsmultimodalproducts
Source →

Microsoft and OpenAI are now openly fighting

The Verge frames Build 2026's announcements as the Microsoft–OpenAI breakup turning into an open competitive fight. Stratechery's Nadella interview backs that up: Microsoft has decided to own the platform end-to-end. Inside Meta, the catch-up story is being told with less optimism.

News The Verge AI

Microsoft and OpenAI broke up — now they're ready to fight

The Verge frames Microsoft Build 2026 as the moment the Microsoft–OpenAI alliance turned into open competition — Scout, MAI-Thinking-1, and Solara as the proof.

Why it matters
  • Recharacterizes the largest AI partnership of the era as a live competitive fight.
  • Pressures OpenAI's pricing power and platform leverage simultaneously.
  • Forces all hyperscalers to commit to either lab-partner or build-yourself, not both.
marketproductsagents
Source →

AI cost reality and the education signal

Uber's $1,500/month per-employee AI cap landed as a useful sector benchmark, dissected by Simon Willison; Berkeley CS classes report failing grades and dwindling math skills tied to AI use; Wired exposed Alpha School's $65K Manhattan campus as a homeschool center rather than a school. The buyer-side and student-side both showing the same fatigue.

News Hacker News

Uber's $1,500/month AI limit is a useful signal for AI tool pricing

Simon Willison's breakdown of Uber's $1,500/month per-employee AI cap as an industry pricing benchmark (504 HN points).

Why it matters
  • Concrete per-seat number that procurement teams will now anchor to.
  • Implies the 'unlimited AI for all employees' pitch is dead on arrival in larger orgs.
  • Sets the price ceiling vendors will design SKUs around.
marketproducts
Source →
News Hacker News

Failing grades soar with AI usage, dwindling math skills in Berkeley CS classes

Daily Cal reports Berkeley CS professors observing soaring failure rates and weakening math skills correlated with AI tool use (378 HN points).

Why it matters
  • First empirical data point linking AI tool use to weaker fundamentals from a top CS program.
  • Will feed both regulatory and parental-pressure narratives on AI in schools.
  • Operator-side counter to the 'AI raises everyone's floor' marketing claim.
educationpolicycode
Source →

Agent infrastructure and the biosecurity letter

Coralogix raised $200M for agent monitoring; Anthropic published a detailed engineering post on how it contains Claude across products; OpenAI and Anthropic co-signed a letter urging legislation to prevent AI-developed biological weapons. The agent stack is getting watched, contained, and politically aligned at once.

News Hacker News

The ways we contain Claude across products

Anthropic publishes a detailed engineering post on how it contains Claude across products (145 HN points).

Why it matters
  • First-party engineering description of how containment actually works at a frontier lab.
  • Useful reference for any team building containment around agentic tools.
safetyagentsproducts
Source →

Papers

4 items

Agents in production: monitoring, economies, structured verification

Three papers move agent research toward production reality: RAMP measures agents in live workflows rather than static benchmarks; Economy of Minds models Hayek-style decentralized multi-agent coordination through auctions; AUDITFLOW gives agents executable symbolic environments for structured financial reporting.

Paper Hugging Face

Benchmarks are Not Enough: RAMP for Runtime Assessing of Agentic Models in Production Systems

RAMP assesses agentic models at runtime in production systems rather than via short-horizon static benchmarks.

Why it matters
  • Names the gap between leaderboard scores and production agent reliability — and proposes a measurement.
  • Aligns directly with the Coralogix observability bet.
  • Likely template the next generation of vendor SLAs will be written against.
agentsevaluationbenchmarks
Source → Arc

Also today