Here are today's key topics, drawn from the news collected on September 17. The debate over accelerating versus slowing AI development continues, with former U.S. Vice President Al Gore and Michael Burry, the investor known for 'The Big Short,' staking out directly opposing positions. On the regulatory front, Spain's data protection authority has published its first data breach report related to AI agents. And for developers, word that a prior study of the Jev architecture we covered last time existed a year ago became a talking point — a day of movement on both the debate and the practical front.
Al Gore and Michael Burry Take Opposite Sides on the AI Slowdown Debate
Former U.S. Vice President Al Gore, in an interview with TechCrunch, struck a surprisingly measured tone on the backlash over AI data center emissions. By his account, what worries him far more than data center emissions keeping him up at night are the warnings that AI industry leaders themselves are issuing about where the technology is headed. For a figure known above all for climate change advocacy, it amounts to declining to join the criticism of data centers.
Investor Michael Burry, meanwhile, is attacking from the opposite direction. He slammed calls by OpenAI and Anthropic to slow AI development as "self-serving," arguing that major AI companies are using the slowdown to tilt the competition in their favor.
On Hacker News, an essay arguing that "a cartel of CEOs trying to monopolize AI is scarier than agents taking over the internet" has also drawn attention, and skepticism of the slowdown argument is shifting away from the substance of the risks themselves toward a question of power: who controls AI. In Japan, an article has been published that organizes the past week's debate not merely as "speed up or slow down" but across external evaluation, human oversight, concentration of power in a handful of companies, and training policy — a sign that the terms of the dispute keep expanding.
Spain Publishes First Data Breach Report Related to AI Agents
Reuters reports that Spain's data protection authority has published its first data breach report related to AI agents. Although the information made public so far is limited, this appears to be the first case in which a personal data breach originating from an AI agent has become the subject of an official investigation and disclosure.
As AI agents move into real business operations, information leakage in the course of agents' interactions with external services is becoming a tangible risk. In the EU, where data breach reporting is mandatory under GDPR, the key question going forward is whether incidents routed through agents will be explicitly positioned as matters for regulatory response. Whether this spreads to other EU countries is worth watching.
The Buzzed-About Jev Architecture: A Prior Study Steps Forward Claiming It Was "Fully Published" a Year Ago
An unexpected follow-up has emerged for the architecture we previously covered as "Jev, which throws away both code and reasoning" — the one that abandons autoregression and predicts probabilities over a JSON schema. A researcher has come forward in Reddit's LocalLLaMA community saying he had fully published the same kind of idea in March 2025 — more than a year ago — complete with an arXiv paper, Hugging Face models, training datasets, and a PyPI package.
According to the researcher, there are differences: his model uses PPO on sequence embeddings to output per-turn transition probability trajectories (probabilities from 0.0 to 1.0), whereas Jev uses parallel sampling (trained with RLCD) to output confidence distributions and schema selection. Still, he says, the underlying architectural thinking is very similar. He also vented his frustration that a frontier lab proposed the same idea as a "literal breakthrough" without a technical paper, public weights, or a public dataset. The post has struck a chord as a classic open-source story: months of labor eclipsed by a big-company announcement of the same work repurposed for broader use.
Who gets to judge novelty as an architecture, and whether frontier lab announcements cite prior work, will remain points to watch — issues that bear on the trust between the research community and industry.
FastRecall — Ultra-Cheap Memory You Can Carry Across Models
On Hacker News's Show HN, "FastRecall" was introduced by a developer who worked on memory features at OpenAI. It is a cross-model memory API that solves the problem of context management turning ad hoc — or context being lost entirely — every time you switch to a different AI model.
Its distinguishing features lie in pricing and design. Storing context is cheap, recall is completely free, and plans start at $2 a month. When you keep using the same model, it respects the provider-side cache to cut inference costs. For cases where full context isn't needed, it also ships with FlashCompact, a model-agnostic context compression. Notably, it explicitly assumes workflows where you hop between models via aggregators like OpenRouter.
With the model layer increasingly commoditized, this is a product built on the forecast that a layer for detaching "context and memory" from models and carrying it with you will become necessary. Whether this layer can actually stand on its own as an independent service is an interesting question.
Does "Start with a Monolith" Still Hold in the AI Era?
An article rethinking the software design rule of thumb "start with a monolith" for the AI era was published on Scalar Engineering (Medium) and has stirred debate on Hacker News. The claim: conventionally, factoring in a team's operational capacity made the monolith the sound choice for an initial architecture, but AI-driven acceleration of development is changing that premise.
If the spread of coding agents drives down the cost of splitting services and writing boilerplate, the very trade-offs of "splitting from the start" change. Whatever the conclusion, the perspective of revisiting design decisions on the assumption of AI-era development speed is an issue with direct consequences for day-to-day practice.
JetBrains Air's Cloud Tasks Run Spring Boot Builds and Tests Entirely in the Browser
On Japan's Qiita, a hands-on account of solving coding agents' operational pains with the cloud is drawing attention. Hand long builds and tests to an agent and your laptop's fans spin endlessly, other work slows to a crawl, and — since you don't want to kill the run halfway through — you can't even put the machine to sleep. In response to this pain familiar to nearly everyone who has started using agents, the author moved the execution environment to the cloud with JetBrains Air's cloud tasks.
The post describes actually completing everything from the Spring Boot build to the tests entirely in the browser. "Redesigning the execution environment" — deciding where agents do their work — looks set to become a major focus of development environment design, as strained local resources increasingly become the bottleneck.