Australian PM Albanese: "Act on the AI Race for Humanity's Sake" — Urging a Middle-Power Role and US-China Cooperation
Australian Prime Minister Anthony Albanese said that middle powers can play an important role in mitigating the risks posed by AI, calling for cooperation between the United States and China. Bloomberg reports.
At a UN meeting scheduled for the coming days, world leaders are reportedly set to discuss an international framework for AI. The remarks came shortly after Albanese met with Apple's Executive Chairman Tim Cook in California.
Around the AI development race unfolding as a US-China duopoly, the standoff between those urging deceleration and those pushing acceleration continues. Rather than siding with either camp, Albanese's position is that middle powers should act as a bridge for risk mitigation. How far the UN-stage framework-building will become concrete is something to watch going forward.
Goldman Sachs: "AI Investment and Government Borrowing Are Pushing Up the Cost of Capital"
On the economic impact of the AI boom, it has been reported that Goldman Sachs compiled an analysis concluding that AI investment and government borrowing are pushing up the global cost of capital.
The argument: massive capital spending on AI data centers and fiscal stimulus from governments around the world are proceeding at the same time, inflating global demand for funds. The framing is that it is not AI progress itself but its "overhead" that is feeding into funding costs such as interest rates — a reframing of the AI boom through the lens of financial markets.
Qwen 3.8 27B Takes on an Unsolved Math Problem on an RTX 5090 — A Public Live Experiment
In Reddit's LocalLLaMA community, a public experiment challenging a 27B local model with unsolved mathematical problems is drawing attention. The person behind it previously published an [experiment attacking the Riemann hypothesis](https://www.reddit.com/r/LocalLLaMA/comments/1wi9fau/qwen_38_27b_running_for_63_hours_on_a_rtx_3090_to/) — running Qwen 3.8 27B autonomously for 63 hours on an RTX 3090 — which drew a huge response. This time, as a sequel, the model is being set to work on a more easily verifiable unsolved problem, on an RTX 5090.
The target is C(25,15,5), one of the [covering design](https://en.wikipedia.org/wiki/Covering_design) problems — a combinatorics question asking how many "groups" of 15 elements each, chosen from a set of 25, are needed so that any combination of 5 elements is covered by at least one group. The best-known solution at this point is 42 groups, and the goal is to reach 41.
The highlight of the experiment is that the correctness of a solution can be machine-checked instantly. On the [live site](https://artificium-covering-experiment.gr.bio/), the agent's public messages, memory, live loop, and code are all viewable as-is, and an independent checker automatically verifies the solutions the agent proposes. It is built so that third parties can track, on the spot, whether the model is genuinely making progress.
The experimenter emphasizes that "even a small open-source model like Qwen 27B can push science and innovation forward on a consumer GPU," and expresses the view that autonomous research by agents should be conducted in a form open to both humans and AI. Compared with the previous attempt — too grand in scope to verify — the shift to a design that prioritizes verifiability is also noteworthy.
Further Japanese Experiments with Jev — The Ambiguity of「大丈夫です」and the openjev Development Story
Reports of experiments and development work with Jev in Japanese continue to appear. Jev is a model that has drawn attention for returning probability distributions over typed questions instead of generating text.
One is a record of building "OpenJev," an open-weight implementation, firsthand. Jev is a "System One model" that returns calibrated probability distributions for typed questions — yes/no, multiple choice, scores — but officially neither its weights nor the details of its training method have been disclosed. The author therefore attempted a reproduction using only public information as clues, and the article covers the process and results all the way to [the release on Hugging Face](https://qiita.com/xeje/items/48c9835a3e5c7b3d5489).
The other is an experiment having Jev judge ambiguity in Japanese. When asked — as a three-way choice of "accept," "decline," or "cannot tell" — whether「大丈夫です」(daijōbu desu) in reply to「袋をお付けしますか？」("Shall I include a bag?") means acceptance or refusal, Jev reportedly answered "decline" when given the entire conversation, but "accept" when given only the reply sentence. Taking this gap as an entry point, the author ran an [experiment](https://qiita.com/kagi_to_packet/items/8c552b1e08b64d1d77e0) comparing 12 fictional conversations across 4 conditions, for a total of 48 evaluations.
The result — the same phrase flips its interpretation depending on whether context is provided — is suggestive for thinking about Jev's input design, which differs from that of LLMs. Measurement-based verification is accumulating in the Japanese-speaking community too, and the outlines of practically usable scenarios should gradually come into view.
Claude Code v2.1.270–278 — Changes Across 9 Releases Summarized
On Qiita, an [article](https://qiita.com/NaokiIshimura/items/6648411c8caf63c9c40e) has been published that organizes the CHANGELOGs of all nine Claude Code releases from v2.1.270 to v2.1.278.
The headline change: the Auto mode classifier has been moved to the server side, and the default is now that organizations are not billed for its cost. It also reportedly touches on how projects without a CLAUDE.md are handled. For the details, refer directly to the original article's summary.
"My Boss Keeps Bringing in Regurgitated AI Output" — A Consultation Depicting Friction at the Dev Frontline
On Hacker News, an [advice-seeking post](https://news.ycombinator.com/item?id=49771657) asking "how do you deal with a boss who brings AI bs into the codebase and the design?" has been drawing attention.
The poster is a developer in their first year at the company. Despite holding the lowest grade, they reportedly complete the most tasks on the team, the fastest and with the fewest defects. But things changed once the company, as a stock-price measure, declared it would go "all-in on AI development." Teammates began ingesting AI output without verifying it, and over-engineered code and technical debt piled up rapidly. Half of a five-page technical design document is littered with AI bs, and colleagues cannot explain how the code they are responsible for works.
Even on the poster's own pull requests, the boss repeatedly brings change requests claiming "AI reviewed this." Items they triaged together on a call — "this is nonsense, this is valid" — come back an hour later as AI suggestions once again. "Either nobody understands the codebase, or they can't understand the AI's code, or they've stopped thinking," the poster writes, worn out.
It is one individual's experience, but the pattern — an "all-in on AI" mandate colliding with quality control — could apply to many workplaces. The post reads as a lesson: rushing adoption without designing a process for who verifies AI suggestions and how makes this kind of friction likely.