Enterprise AI Customers Are Drifting Away from OpenAI and Anthropic Over High Costs

According to a report by QZ (Quartz), enterprise customers of OpenAI and Anthropic are increasingly scaling back or rethinking their usage as ROI (return on investment) remains hard to see. The backdrop: costs keep piling up while the expected productivity gains are difficult to quantify.

The topic also came up on Hacker News, where commenters made points like "once AI tool subscriptions cover dozens of seats, you're looking at tens of millions of yen per year" and "they're useful for prototyping, but in many cases the cost doesn't stack up for production use." It seems the "muscle soreness" of AI investment is starting to influence corporate decision-making.

Note that the trend of trimming AI spending has been reported since last week; what's new this time is that the two major players, OpenAI and Anthropic, are being named explicitly.

What Is the "Dreams API" for Rebuilding Claude Agent Memory?

An article published on Qiita explains the challenges of memory management when running Claude agents over the long term, along with a solution approach built on the Dreams API.

In agents that operate for extended periods, problems emerge: the same knowledge piles up in memory files with different wording, and past assumptions linger on, contradicting current ones. The Dreams API reportedly tidies up and consolidates this scattered memory, providing a mechanism for the agent to autonomously rebuild "what it knows and what it should forget."

For developers running agents over the long term, bloated and contradictory memory is a real pain point, and this qualifies as a practical approach worth paying attention to.

Before AI-Driven Development Floods You with PRs — Visualizing with Labeler

Also on Qiita, a visualization technique using GitHub's Labeler was introduced for teams whose PR (pull request) lists, amid the spread of AI-driven development with Claude Code and GitHub Copilot, suffer from problems like "the title alone doesn't tell you what's inside" and "you can't tell how big the diff is until you open it."

Specifically, the approach automatically attaches labels to AI-created PRs so that their size and type can be grasped instantly from the list view. As the volume of PRs generated by AI tools keeps growing, this is practical know-how that directly feeds into prioritizing and streamlining reviews.

What to Know Before Wrestling with Unusable AI Code at Work

Another Qiita article points out a problem: while the spread of ChatGPT and GitHub Copilot has ushered in an era where anyone can build their own business systems and Excel/VBA macros without programming expertise, an increasing number of workplaces are being left scratching their heads over "unusable AI code."

Even when AI-generated code looks like it works, it can contain fatal flaws in production environments — security issues, lack of maintainability, missing error handling, and the like. The article emphasizes the concrete pitfalls to watch for when using AI coding tools, as well as the importance of a human review process.

Local STT Takes On Dragon Professional

A lively discussion played out in Reddit's LocalLLaMA community over the search for a local speech recognition (STT) tool on par with Dragon NaturallySpeaking (Dragon Professional).

Local STT tools such as Handy.computer and SottoScribe do exist today, but they reportedly fall short of Dragon Professional's advanced features, such as correcting text during dictation, selecting words, and editing text that has already been entered. Open-source projects like Speech to Windows Input (STWI) and Dictate were also introduced, but each remains limited to partial functionality.

The quality of STT leveraging local LLMs has improved dramatically, but the "incremental editing" capabilities required for professional use apparently have yet to reach a practical stage.

AI Is Eroding the Culture of Reading

An article published on Substack, "AI Erodes a Legacy of Reading," drew attention on Hacker News. It discusses a cultural concern: as AI makes summaries and interpretations of text readily available, people are losing the very habit of reading long-form writing.

While the speed of information consumption improves, some warn that the critical thinking and contextual understanding gained through deep reading may be undermined. It's an article that makes you think again about the price of having AI "read in your place."

---