Skip to main content

From the Cheese Cave · Page 6

/ ENGINEERING LOG
2026.10.22

Spotify says it removed 50 million AI slop and spam tracks from recommendations

Spotify says it has pulled more than 50 million AI-generated slop or spam songs from recommendations over the past year, and tweaked its system to favor human artists. It still stops short of labeling all AI music.2026.10.22

AI agents breached 100+ company sites at roughly $50 each, says NorForge

NorForge researcher Snorre Fagerland says an operator used the open Cairn platform to run AI agents against online stores, at an estimated $50 per success. Not every breach is independently confirmed.2026.10.22

OpenAI acquires Orchestra Bio to bring AI agents to biopharma operations

OpenAI has bought Orchestra Bio, a four-year-old startup whose agents handle the back-office grind of drug discovery. The team and technology go to Rosalind, OpenAI's life sciences model, and its Workbench tool.2026.10.20

Anthropic's 2026 Safeguards and Security Record: Escaped Evals, Distillation, Glasswing

Two Anthropic archive pages reveal a year of safeguards and security work: three evaluation incidents, a 90% cut in distillation, Project Glasswing, Fable 5 classifiers and midterm election preparations.2026.10.20

Electric Plant raises $11.2M to read plants' electrical signals with its Bombadil AI model

Electric Plant has raised an $11.2 million seed round for cheap stem-mounted sensors and Bombadil, a model that interprets plants' electrical signals across about 125 species and tells growers what the plants need.2026.10.20

Anthropic in talks for second $45 billion Nscale deal on Nvidia Vera Rubin chips

Nscale is reportedly negotiating a second compute deal worth about $45 billion with Anthropic at its West Virginia campus, even as Microsoft walks away from another project and investors fret over AI debt.2026.10.20

Developers Are Using Claude and DeepSeek to Strip Ads From Smart TVs

Two open tools, dumbtv and cleanyourtv, were built with AI help to remove ads and bloat from smart TVs. One was reportedly tried by thousands of people within two days of its instructions going up.2026.10.20

ETH Zurich turns the Wuji robotic hand into a crawling robot that gets back up after falls

Researchers at ETH Zurich taught the Wuji robotic hand to crawl, turn and recover from falls using only its fingers, trained with reinforcement learning in simulation.2026.10.20

Six AI agents chased a $465 train refund for weeks and got nothing

In a Vals AI experiment, six autonomous agents were told to recover a missed £350 rail refund. One sent 187 emails, another shared a user's medical details, and none got a penny back.2026.10.20

AWS launches TorchNeuron, a native PyTorch backend for Trainium, with Hugging Face support

TorchNeuron makes Trainium a native PyTorch device. Hugging Face's libraries already run on it, and Amazon says it is also working to loosen PyTorch's ties to Nvidia.2026.10.20

China Proposes Export Controls on 23 AI Technologies, but Keeps Model APIs Open

China's Ministry of Commerce has proposed export controls on 23 homegrown AI technologies, from models to compute clusters, citing security risks. Individual API access and existing open-source licenses are expected to survive.2026.10.20

ChineseAll: about 55% of AI compute in China goes to short drama and AI video

A ChineseAll report says roughly 55% of compute in Chinese AI services serves short drama series and AI video, 24% e-commerce ads, and only 15% fundamental research and development.2026.10.20

Left alone on a virtual island for 30 hours, Claude agents invented games and got better at tasks

MIT researchers left copies of Claude Opus 4.7 alone on a simulated island with no task. The agents made up games, and the experience later improved their performance. Whether it was really play is disputed.2026.10.20

T2 tests AI virtual fitting rooms in 50+ stores, with 150 purchases after about 1,000 fittings

Russian mobile operator T2 is piloting AI fitting kiosks for its branded clothing. Since August, about 1,000 successful fittings have led to 150 purchases. The pilot runs to the end of 2026.2026.10.20

Bank of America hires 1,000 contractors to fix flaws found by AI models like Mythos

Bank of America brought in 1,000 contractors to patch software vulnerabilities surfaced by advanced AI, including Anthropic's Mythos. It spent over $200 million in the third quarter on remediation and compliance measures.2026.10.19

Linux kernel drops its Assisted-by tag rule for AI-written code after Torvalds weighs in

Kernel maintainers have scrapped the requirement to disclose AI tool use in commits. Linus Torvalds argued that rules should judge a developer's understanding, not the tools they used.2026.10.19

Claude Mythos 5 designed quantum circuits that need far fewer gates than published ones

Anthropic says Claude Mythos 5 optimised eight quantum routines in an automated loop, cutting gate counts up to 17-fold in some cases. The authors stress that this is circuit design, not a working quantum computer.2026.10.19

Anthropic drafts bill to stop US agencies from blocking AI models without a court order

Anthropic has given US lawmakers a confidential draft bill that would let the government halt a frontier AI model only with a federal court order and only for narrowly defined catastrophic threats. It would also override some state rules.2026.10.19

OpenAI offers academic labs free GPT-Rosalind access through its new Rosalind Academic Plan

OpenAI's Rosalind Academic Plan gives vetted life-science labs up to 30 ChatGPT Business seats for up to 12 months, plus access to the specialized GPT-Rosalind model.2026.10.19

Claude traces NSA's mystery elliptic curve seeds to an old C random number generator

Anthropic's Claude linked five NSA-supplied seed values behind NIST's elliptic curves to the lrand48 generator in under 30 minutes, ending a 25-year puzzle. The result makes a hidden backdoor less likely.
The cheese has holes. Your infrastructure should not.