Torvalds recounts AI debug session from hell; Z.ai flash model undercuts Nvidia

Torvalds details debug session from hell with AI that insisted the bug was unsolvable

Linux creator Linus Torvalds described a grueling debugging session in which his AI assistant repeatedly recommended giving up on a problem it called impossible and unsolvable. The eventual fix required changing a bogus round_up() function to round_down(), but the path there involved 24 patches adding debug information and 18 kernel boots to narrow down the issue. Torvalds noted the problem was not a syntax error or a typical bug, since the function was technically correct but not doing what he wanted it to do.

Torvalds credited his own stubbornness for the breakthrough, saying he suspects the AI was trained by people who may not be quite as stubborn as he is. He did acknowledge the AI’s role in faithfully adding debug code and analyzing results when pushed, and he allowed it to write the final commit message. The episode highlights both the limits and the utility of AI-assisted development, especially when the model’s confidence in failure is misplaced.

Linus Torvalds details his ‘debug session from hell’ pushing an AI to fix an error it repeatedly said was ‘impossible and unsolvable’. He then made it write the report →

Z.ai’s GLM-5.3-Flash matches top models at a fraction of the cost, runs without Nvidia

Z.ai released GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series, with 320 billion total parameters and 18 billion active. It is released under an MIT license, offers a one-million-token context window, and its weights are available on Hugging Face. Artificial Analysis measured the model at 57 points on its Intelligence Index at maximum reasoning effort, three points behind the larger GLM-5.3 and level with GPT-5.6 Terra and Muse Spark 1.2. Cost per task on the index is 0.09 dollars versus 0.68 dollars for GLM-5.3, making it roughly 7.5 times cheaper and placing it on the Pareto frontier of intelligence and cost.

On Z.ai’s API, GLM-5.3-Flash costs 0.15 dollars per million input tokens and 0.50 dollars per million output tokens, a little over ten percent of GLM-5.3’s price. On the GDPval-AA v2 agentic benchmark, the model scores an Elo of about 1770, matching GLM-5.3 and Grok 4.6, and trailing only Claude Opus 5. Roughly 90 percent of its output tokens went to reasoning. Before launch, Z.ai tested the model anonymously as ox-alpha on OpenCode and OpenRouter, where it became the most popular model of the week, with all traffic running on Chinese AI chips. SemiAnalysis reports it served 100 trillion tokens a day. Z.ai says its hardware efficiency and cost per token are on par with common Nvidia GPUs, and SemiAnalysis views this as another test of the CUDA moat. Z.ai built its own serving software on top of SGLang and broke processing into stages that scale independently, tripling throughput over its first attempt on the same hardware, with an agent based on GLM-5.3 helping with the optimization.

GLM-5.3-Flash matches top models at a fraction of the cost, and runs without Nvidia →

Google’s Gemini Omni 1.1 Flash cuts video generation costs and adds scene extension

Google released version 1.1 of its Gemini Omni Flash video model, introducing scene extension that analyzes up to ten seconds of existing video rather than only the last second for more visually consistent results. Scenes can be extended in 10-second increments up to a total of 40 seconds. Developers can upload up to three seconds of external footage as a style reference to carry over characters or motion patterns, and can set start and end frames to create camera movements between keyframes.

A 360p draft mode runs up to 60 percent faster at a third of the cost compared to 720p, with videos upscalable to 1080p or 4K. Per-second pricing is 0.03 dollars for 360p, 0.10 dollars for 720p, 0.15 dollars for 1080p, and 0.30 dollars for 4K. Omni 1.1 is available through Google AI Studio and the developer docs, making advanced video generation more accessible and affordable for developers.

Google’s Gemini Omni 1.1 Flash makes AI video generation cheaper and more flexible →

Wharton study finds AI shopping agents inconsistent and easily swayed by sources

Researchers at the Wharton School tested how consistently AI shopping agents recommend products when the search process changes, using the ACES simulator with six current models tasked to pick a fitness watch from a fixed product grid. Even without external sources, the models showed different baseline preferences. When agents saw just one external source before the product page, recommendations shifted dramatically in some cases. Wirecutter had the strongest pull, with the probability of picking the Fitbit Inspire 3 jumping by 90 percentage points for Claude Opus 4.8 and by 99 percentage points for Gemini 3.5 Flash compared to the control condition.

In further experiments, multiple sources didn’t balance out recommendations, and more sources led to more variability. Presentation order itself proved to be a driver of product selection, with Gemini 3.1 Flash Lite the most sensitive and Claude Haiku 4.5 the most stable. When a product was objectively superior on every dimension, several models still shifted selections toward pricier options based on short user memory statements like I love hiking. The researchers conclude that letting an AI agent buy on your behalf doesn’t guarantee consistent or optimal purchase decisions, and that optimizing for AI shopping will be harder than traditional SEO since sellers don’t know which model is shopping, what it read, or how it processes information.

AI shopping agents aren’t ready to buy on your behalf, study finds →

Anthropic’s new MHS protocol lets AI agents control physical devices

Anthropic introduced the Model Context Hardware (MHS) protocol, a new hardware standard designed to let AI agents control physical devices. The company demonstrated the standard with examples like Claude adjusting a laser, checking results via a separate camera, and repeating the process to calibrate the system, as well as focusing a microscope, analyzing results, and automatically moving it to continue an experiment. In a video, Claude reasoned how to get a robotic arm to pick up an aluminum can without specific training on the steps.

MHS includes a standardized tagging system that encodes hardware’s physical characteristics, adjustable parameters, measurement options, and enforced safety limits, which can be integrated into a reference file for devices the AI has no previous training experience with. Anthropic is working with partners including Amazon Web Services (Strands Robots), Hugging Face (LeRobot), Raspberry Pi, Automata, and Universal Robots during a preview period to build safety evaluations and best practices. The plan is for MHS to become an open source and agent agnostic standard, with early testing showing reduced device integration time and faster iteration in experimental settings.

Anthropic’s new hardware standard lets AI agents control the physical world →