Anthropic Enters Drug Discovery With Claude Science Anthropic wants to develop its own drugs — The Verge Anthropic announced Claude Science, a unified AI workbench integrating over 60 scientific databases across genomics, proteomics, and cheminformatics for pharma researchers — and revealed it is simultaneously running its own in-house drug discovery program, initially targeting neglected diseases. Early customers include Novo Nordisk and the Allen Institute; Anthropic is offering up to $30,000 in research credits for 50 projects, with applications closing July 15. The launch positions Anthropic against OpenAI’s GPT-Rosalind and signals frontier labs are now competing not just on model benchmarks but on scientific outcomes.

🤖 Frontier Models

Meta’s Watermelon Matches GPT-5.5 Benchmarks — Let’s Data Science Meta’s superintelligence chief Alexandr Wang disclosed that the company’s still-training model, codenamed Watermelon, has caught up with OpenAI’s GPT-5.5 on closely-watched benchmarks — reportedly using an order of magnitude more compute than Meta’s prior Muse Spark release. No launch timeline was given.

Seed2.0 Model Card — arXiv ByteDance published the Seed2.0 model card, describing an evaluation-driven approach centered on long-tail knowledge, complex instruction following, multi-step reasoning, and visual understanding for real-world tasks. The paper emphasizes task performance over benchmark gaming.

Introducing Laguna XS 2.1 — Poolside AI — A 33B MoE model targeting agentic coding that hits 63.1% on SWE-bench Multilingual (up 5.4 points); available on Hugging Face and via API under the OpenMDW-1.1 license.

Residual Context Diffusion Language Models — Apple ML Research — A module that recycles discarded token representations in block-wise diffusion LLMs as “contextual residuals,” injecting them back into subsequent denoising steps to improve generation quality.

🏢 Enterprise & Industry

Zuckerberg Tells Staff AI Agents Haven’t Progressed as Expected — TechCrunch In an internal town hall, Mark Zuckerberg acknowledged that the pace of AI agent development is not accelerating as executives had hoped, and conceded that this year’s job cuts “were not as clean as they should have been.” He expects meaningful returns from AI investments in the next three to six months — a notably measured tone from a company that made agents the pillar of its restructuring.

AI Hardware Inflection: Custom Chips Move from Concept to Product — X / AI Late June marked a hardware turning point: OpenAI, Etched, Amazon, and SambaNova each transitioned custom AI chips from roadmap items to shipping or near-shipping products, signaling that the bespoke silicon race has entered its execution phase.

China Quant Funds Draw Billions as AI Trounces Human Traders — Bloomberg — AI-driven quant funds in China have more than doubled assets under management in under a year to over 2.6 trillion yuan; new AI-focused fund products are launching to meet investor demand.

Meta’s Inevitable Cloud — Spyglass — Analysis of Meta’s nascent cloud service as its most credible path beyond advertising, which would put the company in direct competition with AWS, Azure, and Google Cloud.

Microsoft Frontier Company: $2.5B to Embed AI Engineers Inside Customers — GeekWire — Microsoft’s new operating unit places engineers directly inside enterprise customers to build and operate AI systems; mirrors similar deployment moves by Amazon and OpenAI this week.

🛠️ Developer Tools & Agents

Introducing Devin Security Swarm — Cognition Cognition launched Devin Security Swarm for codebase-wide vulnerability discovery, using its Agentic MapReduce architecture: agents fan out over repository shards, reduce findings to a single report, then verify serious vulnerabilities in isolated sandboxes before marking them confirmed. Extensive documentation and technical materials are now public.

Claude Code 2.1.200 and 2.1.199 — Anthropic Two releases shipped today. 2.1.200 changes AskUserQuestion dialogs to no longer auto-continue by default (opt in via /config) and renames the “default” permission mode to “Manual.” 2.1.199 adds stacked slash-skill invocations loading up to five leading skills, fixes SSL errors on TLS-inspecting proxies, and patches a bug where partial streaming responses were discarded on mid-stream server errors.

Agent-Assisted SGLang Development — LMSYS — The SGLang team outlines converting agent workflows into reusable SKILL.md files, benchmark contracts, and review loops, framing agent value around executable, testable engineering knowledge.

CursorBench 3.1 — Cursor — New benchmark evaluating coding agents on ambiguous, multi-file tasks drawn from real Cursor sessions.

ContextForge — IBM (GitHub) — Open-source registry and proxy for federating MCP servers, A2A agents, and REST/gRPC APIs behind a single governed endpoint with auth, rate limiting, and OpenTelemetry observability.

Kiro 1.0.89: Session Restore and Custom Agent Enhancements — Kiro — Session restore, performance optimizations, and custom agent improvements; auto-updates are currently paused and must be downloaded manually.

🛡️ Policy & Security

Cloudflare Sets September Deadline for AI Crawlers to Separate Search From Training — NBC News Cloudflare has given AI crawlers until September to decouple bots that index content for search from those that harvest it for model training, or face blocks on ad-carrying pages. The policy gives publishers a clearer lever to control training use without also shutting out search indexers.

UN Releases First AI Governance Preliminary Report — UN News — The UN Independent International Scientific Panel on AI released its inaugural preliminary report on July 1; findings will feed into the UN Global Dialogue on AI Governance opening in Geneva on July 6.

Argo CD 3.5 Tightens Supply Chain Security — InfoQ — RC adds internal mTLS, Git commit signature verification, and a native ApplicationSet UI; closes meaningful feature gaps with Flux and Rancher Fleet.

Tesla Caps Employee AI Spending at $200/Week — Except for Grok — Electrek — A telling sign that even companies betting their future on AI are struggling to govern costs; Grok is exempt from the cap.

Generated by claude-sonnet-4-6 · 2026-07-03T10:00:00Z