Anthropic and OpenAI Court Wall Street as Retracted Study Undercuts AI's School Push

01Anthropic and OpenAI Each Built Wall Street JVs This Week to Sell Their Enterprise AI

On May 4, Anthropic announced a joint venture with Blackstone, Hellman & Friedman, and Goldman Sachs to build "a new enterprise AI services company." Hours later, TechCrunch reported OpenAI was launching a near-identical structure with its own asset manager partners. Both deals pair model labs with private equity to handle deployment, workflow redesign, and distribution into Fortune 500 buyers.

The same day, Sierra closed a $950 million round, bringing the customer-experience AI startup past $1 billion in total funding. Sierra says it intends to become the "global standard" for AI-powered customer service.

Three announcements, one structural conclusion: the labs have decided their enterprise pipeline cannot be filled by an API and a sales team.

Anthropic's framing in its post is direct. The new entity will help large enterprises identify use cases, redesign processes around them, and operate the resulting systems. That is consulting work — the kind Accenture and Deloitte have priced for two decades. By bringing in Blackstone and Goldman as co-investors and balance-sheet partners, Anthropic gets distribution into thousands of portfolio companies and limited-partner relationships without staffing an in-house consultancy.

OpenAI's parallel deal makes the pattern explicit. According to TechCrunch, both companies "partnered with asset managers to more aggressively market their enterprise AI products." Neither lab is positioning the JV as a model upgrade. They are positioning it as a go-to-market function.

Sierra rounds out the picture from below. The startup pitches itself as owning the customer-experience layer rather than the underlying model, and that pitch has now attracted a billion dollars from investors. Capital flowing to Sierra is capital that does not flow to whoever supplies Sierra's underlying inference.

Margin moves down the stack. The model layer keeps the headline numbers and the brand recognition. The deployment layer, the workflow-redesign layer, and the customer-experience layer collect the integration revenue and the seats inside the buyer. Both labs signed asset-manager partners this week because that is where the revenue is consolidating.

The next test is pricing. If the JVs charge consulting rates while their own model APIs sit on commoditized per-token pricing, the labs will know which side of the trade earns more.

02"Notepad++ has never released a macOS version" — its creator had to say it out loud

Don Ho posted a clarification this week that he should not have needed to write. "To be clear: Notepad++ has never released a macOS version," he wrote, addressing the appearance of a paid app called "Notepad++ for Mac" on Apple's App Store. He has maintained the original Windows-only text editor since 2003.

The Mac version was not his. According to Ars Technica, an unaffiliated developer used an AI coding agent to assemble a lookalike. They shipped it under a name Ho spent more than two decades building. Apple's review process let it through.

Ho's statement is conspicuously short on legal threat. He does not allege trademark infringement, demand a takedown, or name the developer. He simply states the fact: any "Notepad++" on macOS is not his, has not been audited, and carries none of his project's security guarantees. Users who installed it bought a clone.

The dispute lands on a softer surface than typical software piracy. The cloner did not lift Ho's source code; they prompted a model to recreate the look, the keybindings, and the workflow. What they copied is the brand and the user experience, both harder to defend than a copyrighted codebase. Apple's guidelines bar deceptive use of third-party trademarks. Enforcement is reactive, and the trademark holder here is a solo open-source maintainer in France, not a corporation with a legal team.

The timing rhymes with developer mood. "Agentic Coding Is a Trap," a post arguing that AI-generated code shifts maintenance burden onto humans, hit 422 points on Hacker News the same week. Ho's case shows the inverse problem. Maintenance burden is one thing. Reputation burden is another, and it now sits on the original author.

What happens next is administrative. Ho can file an Apple trademark complaint or pursue counsel in France, where Notepad++ is registered. Until he does, the clone keeps charging users.

03The study used to prove ChatGPT helps students just got retracted. Same week, OpenAI, Google and Microsoft pushed AI into schools.

An influential study claiming ChatGPT improves student learning was retracted this week over data red flags, according to Ars Technica. The paper had been cited hundreds of times, making it one of the most-referenced data points behind the broader claim that generative AI lifts classroom outcomes.

The retraction landed the same week Senators Adam Schiff and Mike Rounds introduced the Literacy in Future Technologies Artificial Intelligence Act. Their bill would route grants through the National Science Foundation to put "AI literacy" curricula in American schools. OpenAI, Google and Microsoft are publicly backing the legislation, according to 404 Media.

The funding pathway is awkward. NSF science research budgets have been cut sharply under the Trump administration, and the agency would now be asked to administer grants for teaching tools whose effectiveness remains contested. The bill text does not specify what evidence base AI literacy curricula must meet before reaching students.

Supporters argue digital literacy and critical evaluation of AI outputs are legitimate goals regardless of whether the tools themselves improve test scores. Teaching students to recognize hallucinations, verify citations and understand model limits does not depend on the productivity claims the retracted paper made.

But the policy timeline is moving faster than the evidence. The retracted study was a load-bearing citation in education-technology pitches arguing AI tools raise learning outcomes. With it removed, the empirical foundation for the "AI improves learning" claim narrows. Three of the largest model vendors are still pushing to embed their products into K-12 instruction, with federal money attached.

If passed, the bill would create a federally-funded pipeline of students trained on the three companies' tools before researchers settle whether classroom AI works.