EAIDaily_2026-06-13

EAIDaily — English AI Intelligence Briefing | Focus: AI Coding & Embodied Intelligence Coverage window: 2026-06-12 08:00 — 2026-06-13 08:00 (GMT+8) Source mix: AI HOT (aihot.virxact.com) selected feed + targeted web searches 8 main items | 5 quick takes | 5 trend lines

🔥 Headlines

1. MiniMax M3 open-weights drop on Hugging Face — China’s first frontier “Coding + 1M + multimodal” MoE

Source: X (@MiniMax_AI) / Sohu / 智东西 — 2026-06-12 14:11 UTC URL: https://x.com/MiniMax_AI/status/2065436935188058208

MiniMax released M3 to Hugging Face with ~428B total / ~23B activated parameters. Coding & agent scores land at 59.0% SWE-Bench Pro, 66.0% Terminal Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard, 74.2% MCP Atlas. Context window is extended to 1M tokens via the proprietary MiniMax Sparse Attention architecture, with native multimodal input. A companion MiniMax Code CLI and API platform launched in the same drop; full weights and technical report in ~10 days.

Why it matters: M3 is the first open-weights model that simultaneously hits frontier-grade coding, million-token context, and multimodal — the three capabilities that have been locked behind closed APIs (Claude Fable 5, GPT-5.5, Gemini 3.1 Pro). The fact that it ships at the same week as Anthropic’s $DXC enterprise alliance and OpenAI’s Codex browser mode is a clear signal that the “closed frontier → open weights” gap on coding agents has collapsed to ~6 months.

2. Kimi-K2.7-Code released & open-sourced — open-weights code model with 30% lower inference cost

Source: X (@Kimi_Moonshot) / Hugging Face / 太平洋科技 — 2026-06-12 10:16 UTC URL: https://x.com/Kimi_Moonshot/status/2065377579130142937

Moonshot AI shipped Kimi-K2.7-Code as open source. Vs. K2.6: +21.8% on Kimi Code Bench v2, +11.0% on Program Bench, +31.5% on MLS Bench Lite, plus 30% reduction in inference-token usage. End-to-end success rate and instruction following in long-horizon coding tasks both improved. A 6× high-speed mode is rolling out next; the model is already wired into Kimi API and Kimi Code.

Why it matters: The two most important code-model drops in the same 24 hours (MiniMax M3 + Kimi K2.7 Code) are both open-weights and Chinese. Combined with North Mini Code (Apache 2.0, June 10) and MiMo Code (MIT, June 11), open-source is now the default for new coding models, not the exception — closed labs (Anthropic, OpenAI, Google) are now the laggards in shipping competitive code weights.

3. OpenAI ships Codex “browser developer mode” + savable rate-limit resets

Source: X (@OpenAIDevs, @OpenAI) / 币界网 / CSDN — 2026-06-12 00:11 / 00:15 UTC URL: https://x.com/OpenAIDevs/status/2065226355495895521

Two Codex updates in one night:

• Browser Developer Mode — Codex can use the Chrome DevTools Protocol (CDP) to read JS performance profiles, console output, network traffic, and DOM/CSS state. It can debug live pages in Chrome and in the Codex in-app browser.
• Savable rate-limit resets — Codex users can now bank reset tokens and use them when they choose, decoupling quota refresh from wall-clock schedule. Go, Plus, Pro, Business get one free stored reset at launch.

Why it matters: CDP access is the missing primitive for agentic UI debugging — until now, an agent looking at a broken React page had to trust screenshots and HTML dumps. With CDP, the agent has a debugger. The savable-reset change is a small product detail but a big tell: OpenAI is converting Codex from a “developer tool with rate limits” into a “scheduled agent that spends tokens on a budget you manage” — the same mental model as cloud infrastructure.

4. Anthropic × DXC Technology multi-year global alliance — Claude to be embedded in mission-critical enterprise stacks

Source: PR Newswire / MarketChameleon / 搜狐 — 2026-06-11 URL: https://marketchameleon.com/PressReleases/i/2320735/DXC/dxc-and-anthropic-announce-multiyear-global-alliance

DXC (NYSE: DXC) and Anthropic announced a multi-year global alliance: DXC becomes a global senior partner in the Claude Partner Network, will train tens of thousands of Claude-certified Forward Deployed Engineers (FDEs), and embed Claude into mission-critical enterprise and government systems. DXC is already running Claude internally for engineering productivity and customer-delivery workflows.

Why it matters: This is the first large-scale “AI coding certification” program at system-integration scale. The enterprise playbook for agentic coding is now codified: partner → certify → embed → scale. For every SOW, DXC brings certified Claude FDEs — which makes Claude the default agentic-coding platform for F500 system-integration deals, ahead of OpenAI and Google. Combined with the Anthropic Public Record survey (see #5) and Amodei’s policy essay (June 11), Anthropic is now playing all three games simultaneously: enterprise, policy, public legitimacy.

5. Anthropic’s first “Public Record” survey: 48% of Americans want AI to cure disease, 64% worry about job loss

Source: Anthropic Newsroom / 智东西 — 2026-06-12 URL: https://www.anthropic.com/news/anthropic-public-record

Anthropic surveyed ~52,000 Americans (YouGov, Nov–Dec 2025, weighted to census). Key findings:

• 48% rank curing cancer/disease as the top AI priority; 36% want AI to help people with disabilities
• 64% worry about job loss, 56% about cognitive dependence, 52% about misinformation
• 70%+ support government regulation; top concerns: privacy (56%), child safety (52%), accountability (49%)
• Only 15% trust AI companies to make decisions for society
• Partisan / geographic split is smaller than expected on most issues

Why it matters: Released one day after Amodei’s “Policy on the AI Exponential” essay, this is Anthropic’s empirical underpinning for the policy ask. The numbers are politically actionable: a 70%+ pro-regulation majority that crosses partisan lines is rare, and gives Anthropic a defensible public mandate to lobby for mandatory frontier-model testing and a job-displacement fund (which Amodei is publicly willing to help finance). Expect this survey to be cited in the next round of US/EU AI legislation drafts.

6. NVIDIA Isaac GR00T Reference Humanoid Robot goes to academic research — Jetson Thor + Unitree H2 Plus + Sharpa hands

Source: NVIDIA Investor Relations / roboticsandautomationnews.com / PR Newswire (Unitree) — 2026-06-01 (still featured in this week’s coverage, 06-11/12 echo) URL: https://roboticsandautomationnews.com/2026/06/11/nvidia-unveils-open-humanoid-robot-platform-for-robotics-research/102480/

NVIDIA’s open humanoid reference design is now shipping to top university robotics labs: Unitree H2 Plus body + Sharpa (Singapore) dexterous hands + Jetson Thor compute + full open Isaac GR00T software stack. The design standardises “the brain” and the body so researchers can swap perception / policy / finetuning code without re-engineering the platform. NVIDIA + Unitree + Sharpa are positioning this as the “Android of humanoid robotics.”

Why it matters: An open reference platform is the prerequisite for the next phase of embodied-AI research: reproducible benchmarks, comparable finetuning recipes, shared datasets. Until now, every humanoid lab had its own hardware/software stack, which made cross-lab comparison impossible. With GR00T as the standard body and Thor as the standard compute, embodied AI is moving from “compete on hardware” to “compete on policy + data” — the same shift that turned LLM research from “build a bigger cluster” into “train a smarter model.”

7. 8th BAAI Conference (智源大会) opens in Beijing — Agent, World Models, Embodied Intelligence, Self-Evolution, AI Safety

Source: BAAI Hub / 腾讯新闻 / 知乎 — 2026-06-12 → 06-13 URL: https://hub.baai.ac.cn/view/55237

The 8th Beijing Academy of AI (BAAI) conference runs June 12-13, 2026 in Beijing with 200+ top researchers, 40+ AI CEOs/founders, 14 Turing Award laureates over its history. Five headline forums: Agent for Science, World Models, Embodied Intelligence, AI Self-Evolution, AI Safety. Featured sessions include Agent for Science (AI-driven scientific discovery), the World Models forum (“破浪·重构”), and an Embodied Intelligence track covering the latest Chinese humanoid work.

Why it matters: BAAI 2026 is the most important AI research venue of the month and a soft launchpad for Chinese frontier research in world models, embodied AI, and self-evolving agents. The five-topic agenda is a near-perfect map of the technical frontier — these are the exact threads OpenAI, Anthropic, and DeepMind are also pursuing. Expect major Chinese preprints, model releases, and embodied-AI demos to drop out of this conference over the next 2 weeks.

8. Anthropic launches “Claude Corps” fellowship ($85K/yr) — same day as Amodei’s “tax yourself to fund displacement” essay

Source: TechTimes / Fortune / darioamodei.com — 2026-06-11 URL: https://www.techtimes.com/articles/318249/20260611/claude-corps-fellowship-pays-85k-anthropic-admits-its-ai-will-displace-workers.htm

Anthropic opened applications for Claude Corps, a fellowship that pays $85,000/year to early-career workers placed at non-profits. The same day, CEO Dario Amodei published the essay “Policy on the AI Exponential” (darioamodei.com/post/policy-on-the-ai-exponential) explicitly warning that frontier AI will eliminate most entry-level white-collar jobs within 5 years and proposing a 0.1-1% “frontier-AI tax” that Anthropic is willing to pay into to fund displacement relief and retraining.

Why it matters: This is the first time a frontier-AI lab CEO has publicly and quantitatively predicted mass white-collar displacement while simultaneously offering his own company’s money to mitigate it. It is a deliberate, strategic play: by getting ahead of the narrative, Anthropic positions itself as the “responsible frontier lab” — and pulls the policy debate away from “regulate or not” toward “how to redistribute,” where Anthropic is offering a concrete mechanism. Every other major lab (OpenAI, Google DeepMind, xAI) will now be forced to respond.

⚡ Quick Takes

• Open-source code models peaked this week: M3 (428B/23B MoE, HF), Kimi K2.7 Code, North Mini Code (Apache 2.0, June 10), MiMo Code (MIT, June 11), MusaCoder (June 11) — 5 major open-weights code releases in 5 days, the densest week in the history of open AI coding.
• Anthropic is now a “policy + public legitimacy” company, not just a model company: Public Record survey + Claude Corps + Policy-on-the-AI-Exponential essay + willingness to self-tax — four coordinated moves in 48 hours, all in the same week as the DXC enterprise alliance.
• BAAI 2026’s five-forum agenda (Agent, World Models, Embodied, Self-Evolution, Safety) is now the canonical 2026 frontier-AI research map — every Western lab’s roadmap aligns with at least three of these five.
• CDP access in Codex = the end of “AI can only edit code, not run it” — combined with savable rate-limit resets, Codex is now operationally closer to “an engineer on payroll” than “an IDE plugin.”
• The NVIDIA-Unitree-Sharpa GR00T reference design is the most consequential embodied-AI infrastructure move of June — open hardware is the precondition for the next decade of humanoid research, just as ImageNet’s open data was for the 2010s.

📈 Trend Lines to Watch

1. “Open weights code model” becomes the default: M3, Kimi K2.7, North Mini Code, MiMo Code within one week — closed labs (Anthropic, OpenAI, Google) are now in catch-up mode on the open-weights track. Expect a wave of “vibe coding” startups in Q3 to launch on top of these four open models, not on closed APIs.
2. The Anthropic public-policy stack is now operational: essay + survey + fellowship + self-tax proposal in 48 hours. Watch for: (a) copycat moves from OpenAI/Google DeepMind, (b) legislative drafts citing the Public Record data, (c) the first bill that includes an “AI lab contribution” line item.
3. Embodied AI is bifurcating into two races: (a) “who can mass-produce” — Figure 03, Unitree H2 Plus, AGIBOT, UBTECH; (b) “who owns the brain” — NVIDIA GR00T, Physical Intelligence π0.7, π*0.6, Amap ABot-Earth0.5. The interesting companies will be those that win at both. Tesla Optimus and Apptronik sit in the middle.
4. BAAI 2026 will produce at least 3 major Chinese preprints / model drops in the next 14 days — historical pattern: the week after BAAI always sees a “中国 AI 月” surge in arXiv and Hugging Face. Watch for: world-model papers, embodied foundation models, and a possible “BAAI-Coder” benchmark.
5. The “agentic coding certification economy” is now a real industry: DXC’s 10,000+ certified Claude FDEs is the first scale deployment. Watch for: Accenture, Capgemini, Cognizant, and TCS to announce competing programs with OpenAI / Google / Mistral within 60 days. The next 100,000 AI engineers will be hired by SI firms, not by AI labs.

Compiled by EAIDaily automation | Source mix: AI HOT (aihot.virxact.com) + targeted web verification Next run scheduled: 2026-06-14 08:15 GMT+8 (automation-1780026692931)