Boston Dynamics' Atlas Performs Goal Celebrations at the World Cup
Atlas, the humanoid robot developed by Boston Dynamics, a Hyundai Motor subsidiary, put on a public demonstration at a 2026 FIFA World Cup venue. During the round-of-16 match between Brazil and Norway, Atlas emerged walking through the players' tunnel and re-enacted several goal celebrations. It reportedly imitated the 'surfing' celebration of Brazil's Matheus Cunha and the 'camera' celebration of South Korea's Son Heung-min, and was also seen handing the match ball to the referee.
Hyundai has mass production and factory deployment of Atlas in its sights, and showcasing the robot on soccer's biggest stage offered a chance to tout its technological progress to a broad audience. Cases of humanoid robots being demonstrated at large public events are on the rise, and the accuracy of mimicking human movements in real time in a sports context points to steady improvement.
Microsoft 365 Raises Prices by Up to 42% to Pass On AI Costs — Some Are Calling It an 'AI Tax'
Microsoft has announced new pricing for Microsoft 365, with increases of up to 42% applied to some plans. The company cites 'continued innovation' as the reason, but the community widely sees the cost of bundling in Copilot features as the main driver. On Hacker News, critics are calling it an 'AI tax,' with concerns raised about a structure that increases the cost burden even for users who don't need AI features.
The bundling of AI features into enterprise subscriptions could spread to competitors as well. As the practice of passing AI investment costs on through pricing picks up steam, users will likely need to scrutinize cost-effectiveness more carefully.
GLM 5.2 FP8 Quantization Scores 79.8% on Terminal-Bench 2.1
Benchmark results for GLM 5.2's FP8 quantization (with FP8 KV cache) have been a hot topic in the local LLM community. Running Terminal-Bench 2.1 in a basic configuration with sglang on an H200 cleared 71 out of 89 tasks, for a pass rate of 79.8%. There was one timeout error, but the result reportedly came without any reruns.
FP8 quantization is spreading as a technique that cuts inference costs while keeping accuracy loss to a minimum. GLM 5.2 posting high marks on coding-agent benchmarks is yet more evidence of the practicality of open-weight models.
Sakana AI Releases 'Fugu', a Multi-Agent LLM Orchestrator
Sakana AI, a Japanese AI research company, has published 'Fugu' on GitHub, an orchestrator that coordinates multiple LLM agents through a single API. It is designed to handle complex tasks efficiently by combining multiple agents, and its distinguishing feature is the ease of building a multi-agent system with just one API.
The project has drawn moderate attention on Hacker News, reflecting strong interest in how agents are designed to collaborate. Multi-agent architecture is anticipated as the next evolution of LLM applications, and simple unified interfaces like Fugu could lower the barrier to development.
The Risk of Entrusting Research IP to AI Companies — Heated Debate on Hacker News
A Hacker News post arguing that AI research IP should not be entrusted to the AI agents of big tech companies has stirred up debate. The poster points out the risk that research data entered into tools like Claude Code or ChatGPT/Codex could be used by these companies for competitive advantage. Even without direct theft, they are concerned about indirect channels, such as incorporation into training data or leaks when human reviewers 'take inspiration' from it.
The post cites Uber's past use of user data as a weapon against competition and regulation, arguing that AI companies have similar incentives. Among AI developers, managing the risks of handling highly confidential research data with external AI tools is once again being recognized as a critical challenge.
Distilled LivePortrait Model Achieves 25fps in the Browser — Under 30ms on an RTX 5090
A distilled LivePortrait model has achieved real-time operation in the browser, as reported on Reddit. According to the developer, the ONNX version of LivePortrait initially took 30 seconds per frame running on Chrome WebGPU, but by distilling and shrinking the model they achieved under 30ms per frame on an RTX 5090. It now runs entirely in the browser and can generate video at 25fps.
While quality is limited and results depend on the portrait, it is a solid proof of concept built with just a few hours of training. Generating AI animation entirely in the browser without compromising privacy is a major practical advantage, and with further improvements, a wide range of applications can be expected.
Kakuyasu Uses Generative AI to Decode Its 30-Year-Old Legacy System
According to ITmedia, Kakuyasu, an instant alcohol delivery service, is using generative AI to tackle the 'black box' problem of a core system that has been running for 30 years. For a system that keeps working but that no one fully understands internally, the company adopted an approach combining AI-driven analysis with operational knowledge from the field. The key, they say, was establishing the techniques to 'control AI' in-house rather than blindly trusting it.
Modernizing legacy systems is a shared challenge for many companies, and leveraging generative AI as an analysis and augmentation tool is gaining attention as a practical solution. The stance of treating AI not as all-purpose magic but as a complementary aid combined with field knowledge is an example other organizations can learn from.
---