Skip to content
Daily Brief News Daily Brief

5 AI stories from June 5-6, 2026: Anthropic Opus 4.7 matches NMR software, OpenAI equity talks with Trump White House, Google Gemma 4 QAT for mobile, Sakana RSI Lab launch, and Moonshot Kimi Code CLI

· by Pondero Newsdesk · 5 stories

AI news daily brief: 2026-06-06

Five stories from June 5-6, 2026: one chemistry benchmark from Anthropic, one US policy development involving OpenAI, one open-weight model efficiency release from Google, one research-lab launch from Sakana AI in Tokyo, and one open-source developer tool from Moonshot AI.

Anthropic Opus 4.7 matches MestReNova on NMR carbon prediction, beats it on hydrogen shifts

Anthropic published "Making Claude a Chemist" on its Science Blog on June 5, 2026, written by Anthropic chemist David Kamber. The white paper evaluates three Claude models against ChemDraw and MestReNova on Nuclear Magnetic Resonance spectral interpretation across 20 compounds drawn from ChemRxiv preprints published after the models' training cutoff.

On hydrogen shift prediction, Opus 4.7 recorded a mean absolute error of +/-0.079 ppm, the lowest of any system tested, per the Anthropic white paper. On carbon prediction, Opus 4.7 (+/-1.37 ppm) and MestReNova (+/-1.48 ppm) were effectively tied. The white paper also covers inverse prediction: given an experimental NMR spectrum, Opus 4.7 proposed candidate molecular structures across 15 elucidation problems, competing with or beating MestReNova on that harder task. The post is described as the first in an ongoing series, with Anthropic working with synthetic, computational, and analytical chemists to extend Claude's chemistry capabilities.

Read more: Making Claude a Chemist.

OpenAI and Trump White House in talks over government equity stake in a US public wealth fund

CNBC reported June 5, 2026 that OpenAI and the Trump administration have been discussing a voluntary arrangement under which OpenAI would donate equity to a US "Public Wealth Fund," distributing returns to American citizens. The Decoder confirmed the reporting on June 6, per The Decoder. No deal terms have been set. President Trump said the public could "essentially become a partner." Private investors value OpenAI at more than $850 billion, per The Decoder, and the company is preparing for an IPO.

Sam Altman first floated the concept in 2025. Senator Bernie Sanders separately announced plans to introduce the "American A.I. Sovereign Wealth Fund Act," a bill that would impose a one-time 50% tax on shares of the largest AI companies and place those shares in a public fund with voting rights and board seats, per The Decoder. Sanders confirmed he discussed the idea with Altman directly on June 4. Anthropic has also published its own policy paper advocating for a similar public wealth mechanism.

Read more: OpenAI and the Trump administration are negotiating a government stake.

Google releases Gemma 4 QAT weights to cut model memory footprint for mobile and laptop inference

Google published "Gemma 4 QAT models: Optimizing model compression for mobile and laptop efficiency" on June 5, 2026, per the Google Keyword blog. Quantization-Aware Training bakes quantization into the training process rather than applying it after the fact, preserving accuracy that standard Post-Training Quantization loses. The release covers QAT checkpoints in the Q4_0 format for all Gemma 4 model sizes, plus a mobile-specialized quantization schema for the edge models (E2B and E4B).

Using the mobile-specialized format, Google reduced the memory footprint of Gemma 4 E2B to 1GB, per the Google Keyword blog. The post describes three design choices enabling this: pre-calculated static activations, channel-wise quantization aligned to mobile accelerator design, and targeted 2-bit quantization on the token-generation layers. The QAT weights are available via Hugging Face, Google AI Edge SDK, and MediaPipe. No stable-release date beyond the current release was given.

Read more: Gemma 4 QAT models.

Sakana AI formalizes Recursive Self-Improvement Lab in Tokyo around five existing research projects

Sakana AI announced the formal establishment of the Recursive Self-Improvement (RSI) Lab on June 5, 2026, per the Sakana AI announcement. The lab consolidates five prior research projects: LLM Squared (AI automating research to invent better preference-optimization algorithms, developed with Oxford and Cambridge), Darwin Godel Machine (agents that rewrite their own codebases, which drove a 30-percentage-point absolute improvement on SWE-bench), ShinkaEvolve (sample-efficient program synthesis using 150 samples for complex optimization), ALE-Agent (RL agents outperforming human experts at economically valuable tasks), and Digital Red Queen (adversarial agent coevolution).

The RSI Lab's founding position: self-improvement does not require hyperscale compute but sample efficiency. Japan's compute constraints and national sovereign-AI strategy supply the design pressure, per the announcement. Co-founder David Ha previously served as Google Brain Research Director.

Read more: Introducing Sakana AI's RSI Lab.

Moonshot AI releases Kimi Code CLI as open-source TypeScript terminal coding agent

Moonshot AI published the Kimi Code CLI repository to GitHub on June 6, 2026, per the kimi-code repository. The MIT-licensed terminal coding agent is written in TypeScript and ships as a single-binary distribution installable without Node.js via a shell script or PowerShell command. Key capabilities: read and edit code, run shell commands, search files, fetch web pages, then choose next steps based on feedback. Three built-in subagents handle distinct modes: coder, explore, and plan, each running in isolated contexts.

Kimi Code CLI works out of the box with Moonshot AI's Kimi models and is configurable for other OpenAI-compatible API providers. The open-source release signals an effort to build a developer ecosystem around the Kimi model family. GitHub star velocity over the coming days will indicate developer uptake. Whether the CLI integrates into existing IDE tools such as Cursor or Windsurf has not been confirmed.

Read more: kimi-code on GitHub.

Sources