Researchers from Huawei Taylor Lab, Peking University, and Shanghai University of Finance and Economics introduced SHAPE.

The method rewards actual progress in reasoning not verbosity by using a two-level system: a stage-aware advantage at the segment level for efficient breakthroughs, and entropy-driven redistribution at the token level for sharper execution.

Result: 3% higher accuracy on math reasoning while using 30% fewer tokens across multiple base models and benchmarks.

Peter Thiel backs $1bn ocean data centre start-up powered by waves

Panthalassa operated mostly in secret for a decade. And what it built is nuts.

Massive, massive floating data centers that drive themselves out to sea and then capture water inside of them to spin a turbine and power GPUs.

Anthropic finally releasing synthetic recipes, here for moral behavior

Midtraining with synthetic documents generated from a model spec or constitution. This induces generalization in alignment with respect to values.

Meet Altara: the scientific intelligence platform for the physical world.

Anthropic is releasing 10 new agent templates for financial services

AI agents now have phone numbers. Saperly just launched the first phone carrier built only for AI agents.

Your agent gets its own number. Voice, SMS, routing, compliance, all live in seconds.

→ Same number every time it calls

→ Same caller ID across every product

→ Voice and text on one line

→ Provisioned in 5 minutes

Any MCP-compatible AI agent can use the phone carrier service to obtain a his phone number in seconds.

Meet Haiku is a foundation model that could align histology, spatial biology & clinical data to reveal latent biomedical insights

Haiku a tri-modal foundation model trained on 26.7M+ spatial proteomics patches with matched H&E and clinical text, aligned in one shared embedding space.

The exciting part: Haiku is a counterfactual framework that fixes the morphology, edit only the metadata, and surfaces niche-specific molecular shifts.

Peter Steinberger, founder of OpenClaw shared a significant tools

Yesterday phones. Today smart speakers and message history. The perimeter keeps expanding.

It was a list of CLI tools he built with AI:

🔊 sonoscli.sh - Sonos
🗃️ wacli.sh - WhatsApp
🪶 birdclaw.sh - X archive
🧰 gitcrawl.sh - GitHub archive
🛰️ discrawl.sh - Discord archive
🎧 spogo.sh - Spotify
💬 imsg.sh - iMessage
🧳 mcporter.sh - MCP to CLI
🗣️ sag.sh - ElevenLabs voice
🧿 askoracle.sh - second opinion oracle, and an MCP-to-CLI bridge.

Easy to read as a flex on developer speed. But that’s not the real story.

Look at the list as a whole. Speakers. Messages. Music. Social feeds. Code repositories. Voice.

Every tool is a different surface of one person’s digital life and every tool is now a skill the agent can install and use.

This is what “giving an agent hands” actually looks like in practice. Not a demo. Not a product announcement. A developer methodically wrapping his entire environment so that his agent can act inside it.

The architecture matters here.
These aren’t features baked into OpenClaw’s core. They’re standalone CLI tools published as installable skills on ClawHub the open registry where any agent can find and load new capabilities on demand. The agent doesn’t just use them. It can reason about when to use them.
sonoscli to change the music. wacli to message someone on WhatsApp. birdclaw to search your Twitter archive. askoracle to ask a second model when you’re stuck.

Each skill is a door. The agent decides which one to open.
This is the pattern to watch: the agentic economy isn’t being built top-down by platforms.

It’s being assembled tool by tool, surface by surface, by builders who are tired of their agents hitting walls.

OpenAI just released Multipath Reliable Connection (MRC), a new open networking protocol that helps large AI training clusters run faster and more reliably, with less wasted GPU time.

MRC is already deployed across all of OpenAI’s largest supercomputers that used to train frontier models.

MRC is now available through the Open computer Project for the entire industry to use and build on.

HUGE! Anthropic just leased the world’s largest AI data center from Musk and signed up to build orbital compute with him too.

Anthropic officially announced it’s getting the entire Colossus 1 data center: 300+ megawatts, 220K NVIDIA GPUs, coming online within a month.

For Claude users this means doubled rate limits on Claude Code for Pro/Max/Team plans, peak hours restrictions removed, and significantly higher API rate limits for Opus models.

But here’s what actually matters. Colossus 1 is xAI infrastructure the same xAI that SpaceX absorbed in February for $1.25T. And just weeks ago, Cursor moved into the same cluster to train its Composer 2.5 model on xAI GPUs.

Musk just became the landlord of the AI compute race. He’s renting the same infrastructure to direct competitors while using Cursor to extract real-world developer data and insights that feed back into his own models.

Everyone’s paying Musk to stay competitive. Including the company that’s building the fastest-growing coding tool to compete with Cursor.

And the detail that almost nobody is talking about: Anthropic officially expressed interest in developing orbital AI compute capacity together with SpaceX. That line is in the official blog post.

The next frontier of the AI race isn’t a data center in Texas. It’s in orbit.

Anthropic launched multiagent orchestration, outcomes and webhooks in public beta and dreaming newest feature in research preview

Multi-agent orchestration enables task delegation to specialized sub-agents.

Each sub-agent has its own context window, but they share a container + filesystem to coordinate work.

Agents can self-improve using a rubric that describes task completion.

A dedicated grader sub-agent uses the rubric to evaluate any work done and returns feedback for the next iteration.

Agents can post updates to webhooks configured in your Claude workspace.

This removes the need to keep an SSE stream open or poll for completion.

Agents write to memory in a session. Dreaming is a background process that reflects over many sessions to curate memories: it can edit them based on patterns, add new skills, or remove stale ones.

To get started with any of these features, try /claude-api skill in Claude Code or the OSS repo

Kraken acquires Hong Kong's Reap, stablecoin payments powerhouse, for $600M.

Turbocharges Asia expansion amid stablecoin boom.

Kraken shells out $600M for Reap, HK/Singapore fintech bridging TradFi and crypto via USDC B2B payments, Visa cards.

Reap: Non-bank card issuer, $40M prior funding, Circle/Fireblocks partner.

Fits Kraken's spree (e.g., NinjaTrader).

Positions Kraken as stablecoin leader in Asia's regulatory hotspots:

- Targets P2P, expense mgmt; eyes AI agents, emerging markets.
- BTC holds ~$81K, no immediate volatility.

DeepSeek's first external fundraising round could hit $50B valuation

DeepSeek is expected to close its first-ever external funding round soon, potentially reaching a valuation of up to $50 billion, according to three sources familiar with the matter.

Investors include AI-focused affiliates under Phase III of China's "Big Fund" (China Integrated Circuit Industry Investment Fund).

Other participants: Global investment firm Hillhouse and Shenzhen-based tech giant Tencent Holdings are also involved in the discussions.

This would be five times higher than the initial $10 billion valuation reported by local media last month when the round began.

What if your AI agent could train itself by exploring thousands of real-world services and automatically finding its own weaknesses?

Renmin University of China and ByteDance Seed introduced Agent-World, a self-evolving training arena.

It works by having the agent discover new tool environments and tasks on its own, then continuously practice and improve where it’s weakest.

The result: Agent-World-8B and 14B consistently outperform strong proprietary models across 23 challenging agent benchmarks.

Shopify launched a free tool/scanner that tells you how visible your store is to AI shopping agents like ChatGPT and Microsoft Copilot.

Paste a URL. Then it runs 31 checks across five categories: AI discoverability, product schema, transaction readiness, trust signals, and operational maturity.

Google introduced MaD Physics a benchmark to evaluate the ability of agents to make informative measurements and conclusions subject to constraints on the quality and quantity of measurements.

The benchmark consists of three environments, each based on a distinct physical law.

To mitigate contamination from existing knowledge, and prevent the results merely reflecting memorization, MaD Physics includes altered physical laws.

In each trial, the agent makes measurements of the system until it exhausts an allotted budget and is then asked to infer the underlying physical law to make predictions about the future state.

MaD Physics evaluates two fundamental capabilities of scientific agents:
inferring models from data and planning under constraints.

Also Google demonstrated how MaD Physics can be used to evaluate other capabilities such as multimodality and in-context learning.

Junyang Lin former head of Qwen, has reportedly founded a new AI Lab valued at $2 billion.

It is understood that he has already recruited several members with backgrounds from ByteDance, Tencent, and overseas institutions.

Researchers from Google and Meta built a framework where Claude Code proposes its own algorithms for making LLMs reason better, then tests them, then refines them based on what failed.

No human in the loop after the environment is set up.

In 5 rounds the agent discovered a controller with 4 coordinated mechanisms working together. EMA momentum stopping. Coupled width-depth control. Alignment-aware depth allocation. Conservative branch abandonment.

The paper says directly: "a level of coordinated complexity that would be difficult to arrive at through manual intuition alone."

That's a polite way of saying the agent built something a human probably wouldn't have.

The cost of the entire discovery was $39.90.

The cost of one researcher's coffee budget just outperformed years of hand-tuned work.

Anthropic’s rumored inference chip just got big backing. How long until these get tossed into space? Inference is perfect for living in ODCs

The U.K. chip startup Fractile said it has raised a $220 million Series B funding round led by Factorial Funds, Accel and Peter Thiel’s Founders Fund.

Starting June 15, paid Claude plans can claim a dedicated monthly credit for programmatic usage.

The credit covers usage of:
- Claude Agent SDK
- claude -p
- Claude Code GitHub Actions
- Third-party apps built on the Agent SDK

Starting June 15, programmatic usage gets its own dedicated budget instead. Your subscription limits don't change, they're now reserved for interactive use.

How it works:

Claim the monthly credit once, and programmatic usage will draw from it automatically. When it runs out, you can keep going with usage credits (billed at API rates you turn on/off). If usage credits are turned off, usage pauses until the credit resets.

Monthly credit amounts vary by plan:

Pro: $20
Max 5x: $100
Max 20x: $200
Team Standard: $20/seat
Team Premium: $100/seat
Enterprise: Varies by seat type

After you claim the credit, it resets with each billing cycle. Credits do not rollover.

This means that third-party tools built on the Agent SDK like Conductor and OpenClaw work with your Claude plan, but will draw from your credit the same way your own scripts do.

There’s nothing you need to do today. Users will get an email on June 8 to claim their credits, and this change goes into effect on June 15.

Meta introduced AI Voice Conversations powered by Muse Spark that let you talk naturally to Meta AI (interrupt, switch topics, or swap languages), and as you talk, Meta AI can generate images and pull up recommendations from Reels, maps, and more.

Also bringing live AI to the app, so you can point your camera at the world and ask about what you’re seeing in real time.