Anthropic and PwC Massively Expand Partnership, Rolling Out Claude to Hundreds of Thousands

On May 15, Anthropic and PwC announced a major expansion of their strategic partnership. PwC will roll out Claude Code and Cowork starting with its US teams, with plans to ultimately scale to hundreds of thousands of professionals worldwide.

The expansion rests on three pillars. First, the two companies will establish a **joint Center of Excellence** and launch a Claude certification program for 30,000 PwC professionals. Second, PwC is standing up a **new business group for CFOs (Office of the CFO)** that will apply Claude to finance transformation in regulated industries. Third, the partnership will focus on three areas: building agentic technology, executing AI-native M&A, and redesigning enterprise functions.

Results from production deployments have already been reported. Insurance underwriting has been cut from 10 weeks to 10 days, and on a mainframe modernization project a COBOL codebase four times larger than anticipated is being processed within budget. In cybersecurity, incident response has accelerated from hours to minutes.

CEO Dario Amodei said: "Insurance underwriting went from 10 weeks to 10 days, and security work from hours to minutes. We look forward to delivering Claude into the hands of hundreds of thousands of people at PwC."

OpenAI Discloses Supply Chain Attack That Compromised Employee Devices

On May 15, OpenAI announced that a supply chain attack compromised the devices of two employees. More than 170 packages were affected, and code signing certificates may have been exposed.

While no impact on customer data has been confirmed at this point, the company has rotated the certificates as a precautionary measure and is asking macOS app users to **update by June 12**. The incident has once again called the supply chain security posture of AI companies into question.

arXiv Introduces One-Year Submission Bans for Papers Containing Unverified LLM-Generated Errors

The preprint server arXiv has announced that it is introducing **one-year submission bans** for papers published with unverified LLM-generated errors.

According to Thomas G. Dietterich, a cs.LG moderator at arXiv, the organization will take strict action when papers contain hallucinated references or LLM meta-comments (such as "I'll insert a 200-word summary here — would you like me to change anything?" or "The data in this table is for illustration only"), since such content casts doubt on the reliability of the paper as a whole. Even after the submission ban ends, acceptance at a peer-reviewed conference will remain a prerequisite for posting.

Asahi and Nikkei Sue Perplexity: First Oral Hearing Held

The first oral hearing in the lawsuit that The Asahi Shimbun and The Nikkei brought against the generative AI service Perplexity over the unauthorized use of their articles was held on May 15 at the Tokyo District Court. The two publishers are seeking a combined 4.4 billion yen in damages plus an injunction to stop the unauthorized use. Perplexity asked the court to dismiss the claims.

Legal battles over generative AI companies' use of media content are spreading around the world, and this case could set an important precedent in Japan.

Beating GPT-3.5 for $100 — Small-Model "Self-Learning" Draws Attention

On Reddit, an independent researcher's writeup of **a method for having small models learn from their own mistakes** has been attracting attention.

The approach has a Qwen 2.5 7B base model solve coding problems and solutions it generated itself, then uses correct/incorrect answer pairs as training data. No human-written data is used at all — the Python interpreter serves as the sole judge. As a result, HumanEval performance improved from 25 to 112 problems solved, and Qwen 2.5 14B achieved major gains with just 95 minutes of H100 runtime ($3.50).

Particularly interesting is the **discovery of a threshold**: when training data is scarce (36 pairs or fewer), sampling produces better results than fine-tuning. This threshold effect has not been documented before, and verification by the research community is expected.

Orchard: An Open-Source Agent Modeling Framework Debuts

**Orchard**, featured on Hugging Face Daily Papers, is an open-source framework that enables scalable agent training.

At its core, Orchard Env is a lightweight environment service providing reusable primitives for sandbox management. Three recipes are built on top of it. Orchard-SWE targets coding agents, scoring 67.5% on SWE-bench Verified — a state-of-the-art result for open-source models of its size. Orchard-GUI, a 4B vision-language model, achieved 74.1% on WebVoyager, making it the best-performing open-source model.

A distinctive point is that the lightweight environment layer allows data, training recipes, and evaluations to be reused across different domains.

---