The Informed Layer · Newsscan
Daily briefing

DeepSeek ships Harness; safety rules tighten

← All issues

Anthropic / Claude ecosystem

Claude Code 2.1.229 adds SSE keepalives and locks down Git push path

Anthropic's Claude Code update adds SSE keepalives to prevent idle timeouts in Vertex and Bedrock deployments and tightens the Git push path.

Anthropic investors bet on $2.8trn valuation in record IPO

Investors expect Dario Amodei's Anthropic to list at a record $2 trillion-plus valuation, potentially making its IPO the largest ever.

Claude in Chrome gains synced sessions across devices

Anthropic is extending cross-client session sync to Claude in Chrome, unifying the backend with Claude Cowork and enabling persistent session state across devices.

Frontier model providers

OpenAI launches Computer History to remember what you did on macOS

OpenAI introduced Computer History, an opt-in macOS feature that logs interaction events to build ChatGPT memory, replacing the earlier Recall-style Chronicle screenshot approach.

DeepSeek launches Harness agent framework in pivot to autonomous agentic AI

DeepSeek released Harness, a modular framework for turning AI models into autonomous agents with configurable operator modes, marking a strategic shift from pure LLMs to agentic AI.

Mistral to host external models, Chinese Z.ai's GLM first

Mistral is expanding beyond its own models to host open third-party models, with GLM-5.2 from Z.ai as the first, plus regional endpoints and a Priority Tier for European AI sovereignty.

Google launches Gemini 3.7 Flash with 50% introductory price cut

Google's Gemini 3.7 Flash is now generally available, targeting coding and agentic workflows with a 50% introductory API price cut and improved speed.

AI developer tooling & infrastructure

MongoDB brings live operational data to agentic coding stack

MongoDB launched a hosted MCP server connecting AI agents to live Atlas data, with native integrations for Claude Code, Codex, Grok Build, and Devin.

MagicSkills open-source AI skill management tool supports 26+ platforms

Peking University's Narwhal-Lab released MagicSkills, an open-source tool unifying installation, composition, and synchronisation of AI agent skills across 26+ platforms.

Cloud & platform providers

Google Cloud adds metrics to BigQuery Graph for AI

Google Cloud is adding metric definitions to graph-based data models in BigQuery Graph during preview, improving AI-driven analytics on connected data.

Cloudflare Access on Workers secures internal vibe-coded apps

Cloudflare introduced account-wide Access on Workers to keep all Workers private by default and protect domains, previews, and routes.

Amazon Nova Multimodal Embeddings now in AWS GovCloud

Amazon's Nova Multimodal Embeddings model is available in AWS GovCloud, enabling cross-modal retrieval over text, documents, images, video, and audio for government workloads.

AI policy, regulation & governance

Japan to deploy stronger AI safeguards as model capabilities advance

Japan plans comprehensive AI-safeguard rules, including stricter disclosure and governance measures, as frontier model capabilities advance.

South Korea gathers views on AI ethics principles ahead of August adoption

South Korea's science ministry is consulting on voluntary AI ethics principles reflecting AI advances and the AI Basic Act, with adoption planned in August.

India Supreme Court asks government to consider plea on high-risk AI in welfare

The Supreme Court of India asked the government to consider a petition seeking a statutory framework and disclosure of high-risk AI used in welfare and governance.

GAO official suggests addressing AI risks through existing cybersecurity law

A GAO official suggested leveraging the Cybersecurity Information Sharing Act to speed AI safety policy through information sharing and rapid incident response.

EU delays high-risk AI rules to 2027 but nudifier apps face December ban

The EU is pushing high-risk AI compliance deadlines to 2027 while separately banning nudifier apps sooner, altering the EU AI Act implementation timeline.

India weighs AI legislation

India is considering AI-specific legislation, according to IAPP, signalling possible regulatory moves in the Asia-Pacific region.

Malaysia's first AI law is coming with a three-tier risk framework

Malaysia is moving from non-binding AI guidelines to an enforceable cross-sector AI statute, with a public consultation paper outlining a three-tier risk framework.

Western Australia police scanned 130,000 faces in one week in live facial recognition trial

WA Police deployed real-time AI facial recognition in public spaces under a government-backed trial, scanning 130,000 faces and generating dozens of alerts in the first week.

Industry & market moves

Google dismantles DeepMind, reassigning multiple teams to Alphabet HQ

Google is restructuring DeepMind by reassigning non-technical teams to Alphabet/Google HQ to streamline Gemini integration, after Demis Hassabis stepped back as CEO.

BMC achieves AWS AI Competency for Control-M

BMC earned AWS AI Competency for its AI-powered Control-M platform running on AWS with Bedrock integrations.

Singapore's Graas raises $17M Series B and acquires Trustana

Graas raised $17 million in Series B funding and acquired product-data startup Trustana to strengthen its AI-for-retail platform and commerce knowledge graph.

a16z invests in Vals, an AI evaluation platform

Andreessen Horowitz announced an investment in Vals, which provides real-work, domain-expert-driven benchmarking and automated grading for AI models.

Cursor is joining SpaceX

Cursor announced it is now part of SpaceX, enabling larger-scale AI model development using SpaceX's GPU fleet.

Uber partners with China's Pony.ai for 2,000 robotaxis in Europe

Uber and Pony.ai announced a large-scale rollout of 2,000 autonomous robotaxis across Europe.

Tencent to become top shareholder of Manus as Meta deal unwinds

Tencent is set to become the top shareholder of Chinese AI developer Manus as Beijing pushes for domestic control and Meta unwinds its deal.

AI product & feature launches

cTrader opens command line to Claude Code, Codex and Gemini CLI

cTrader now offers a standalone CLI to run trading and AI-assisted tasks without the GUI, integrating Claude Code, Codex, and Gemini CLI.

X Square Robot demonstrates embodied AI in real-world logistics

X Square Robot demonstrated its WALL-B AI foundation model and six-axis robot arm performing end-to-end parcel induction with adaptive decision-making in changing piles.

Writer introduces Palmyra X6 to contain token costs

Writer launched Palmyra X6, a new flagship AI model designed to reduce token costs and deploy more cost-effectively, with harness upgrades.

Research with immediate practical relevance

Preferential CDR masking improves antibody language model binding prediction

A Nature Communications AI & Computing paper shows region-aware CDR masking in paired antibody language models improves binding affinity prediction over uniform masking.

Less is More: scaling dynamics with sparsity for Un-0 model

A new post describes scaling dynamics and sparsity in the Un-0 model to improve trainability and scaling efficiency.