AI Is More Persuasive Than Human Experts — Confirmed in a Large-Scale Experiment

A joint research team from the University of Oxford, Stanford University, the London School of Economics, and the UK AI Safety Institute has published the results of a large-scale experiment on the persuasiveness of AI systems. The findings show that AI can reliably change people's minds on policy issues and charitable donation amounts more effectively than human experts.

The experiment involved 6,923 participants and 18,978 conversations, with AI outperforming human experts in text-based persuasion. Human persuaders matched the AI's results only when constraints were placed on the AI side. The conclusion — "AI systems were reliably more persuasive than expert humans" — once again highlights the risks this poses for future policymaking and public opinion manipulation.

The study was also covered in the latest issue of Import AI (issue 462), where it is discussed from the perspectives of "superpersuasion" and "paths to ASI (artificial superintelligence)."

The Future of Science on JUPITER, Europe's First Exascale Supercomputer

JUPITER — Europe's first exascale supercomputer, operated at Forschungszentrum Jülich in Germany — showcased real scientific results at ISC 2025 in Hamburg. Four projects running on the system, built with NVIDIA Grace Hopper Superchips and Quantum-X800 InfiniBand, drew particular attention:

Cell-scale mapping of the brain**: Detailed mapping of the human brain at the cellular level

km-resolution Earth climate simulation**: Computing the entire globe on a 1 km grid

AI for next-generation wireless networks**: AI systems underpinning wireless infrastructure such as 6G

Simulation of a 50-qubit quantum computer**: Emulation of universal quantum computation

Thomas Lippert, Director of the Jülich Supercomputing Centre, stated: "With JUPITER, Europe is not just joining the exascale era — it is leading the world in every field of science and AI." Meanwhile, Los Alamos National Laboratory (LANL) also announced new supercomputers named "Mission," "Vision," and "Veritas" based on the NVIDIA Vera CPU, positioning them to accelerate agentic scientific AI.

Sakana AI's "Fugu" Declares Frontier-Class Performance

Fugu, a model released by Japan's Sakana AI, has been reported to deliver performance surpassing generally available frontier models. On hard benchmarks spanning engineering, science, and reasoning, it is said to be on par with Fable 5 and Mythos Preview.

What makes this particularly notable is that Fugu delivers frontier-level capability "without taking on export-control risk." Given that Fable 5 and Mythos 5 were made unavailable at the direction of the US government, the existence of a frontier-class model unconstrained by geopolitics carries significant weight. Within the Japanese developer community, instructions for using Fugu from CLI tools such as opencode have already started making the rounds.

GLM Coding Plan Raises Prices 6x in Six Months

Z.ai's GLM Coding Plan, which exploded in popularity during the second half of 2025 thanks to its low "from $3 a month" pricing, has implemented a massive price hike just six months later. The Lite plan went from $3/month to $18/month (about 6x), while the Pro plan rose from $15/month to $72/month (about 4.8x). On top of that, weekly limits have been newly introduced, and users on the old plans are being forced to migrate to the new ones.

For developers using GLM models from tools like Claude Code, this dramatic shift in cost structure is a major blow. GLM-5.2 remains a strong option, but its positioning as a "cheap alternative" is fading.

"Half of AI-Generated Code Is Vulnerable" — Working Is Not the Same as Secure

An article gaining traction on Qiita puts the security risks of AI-generated code to a practical test. Have AI write code, and you get a working function in tens of seconds — one that passes its tests, too. But "working" and "secure" are two different questions.

The article proposes an approach that evaluates AI-generated code separately on "does it work" and "is it secure," and as a result, roughly half of the code was found to contain some kind of vulnerability. As AI coding tools continue to spread, this perspective carries important practical implications.

Good News for Local LLM Builders — EU DDR5 Prices Are Plummeting

On Reddit's LocalLLaMA community, data tracking DDR5 prices across four European countries (Germany, the Netherlands, Spain, and Belgium) over 25 days has been drawing attention. The results are clear: prices of major DDR5 memory kits have dropped 13–28%.

For example, the G.Skill DDR5 Aegis 2x16GB 6000 fell from €579 to €419 (-28%), and the Kingston FURY Beast RGB 2x16GB 6000 dropped from €499 to €369 (-26%). By country, the price gap between Germany and the Netherlands is wide, with competition heating up on the German side. Falling costs for the large memory capacities needed for local LLM inference could help broaden the reach of open-source AI.

---