AI in practice Archive

Dec 23, 2025

Zhipu AI has introduced GLM-4.7, a new model specialized in autonomous programming that uses "Preserved Thinking" to retain reasoning across long conversations. This capability works alongside the "Interleaved Thinking" feature introduced in GLM-4.5, which allows the system to pause and reflect before executing tasks. The model shows a significant performance jump over its predecessor, GLM-4.6, scoring 73.8 percent on the SWE-bench Verified test. Beyond writing code, Zhipu says GLM-4.7 excels at "vibe coding" - generating aesthetically pleasing websites and presentations. In a blog post, the company showcased several sites reportedly created from a single prompt. Benchmark comparisons show a tight race between GLM-4.7 and commercial Western models from providers like OpenAI and Anthropic. | Image: Zhipu AI

A table with benchmark results comparing the GLM-4.7 AI model with competitors; the model shows leading values in categories such as Reasoning, Code Agent and General Agent. — Benchmark comparisons show a tight race between GLM-4.7 and commercial Western models from providers like OpenAI and Anthropic.

The model is available through the Z.ai platform and OpenRouter, or as a local download on Hugging Face. It also integrates directly into coding workflows like Claude Code. Z.ai is positioning the release as a cost-effective alternative, claiming it costs just one-seventh as much as comparable models.

Comment Source: Zhipu AI

Jonathan Kemper

Dec 23, 2025

Short News

The Qwen team at Alibaba Cloud has released two new AI models that create or clone voices using text commands. The Qwen3-TTS-VD-Flash model lets users generate voices based on detailed descriptions, allowing them to precisely define characteristics like emotion and speaking tempo. For example, a user could request a "Male, middle-aged, booming baritone - hyper-energetic infomercial voice with rapid-fire delivery and exaggerated pitch rises, dripping with salesmanship." According to the manufacturer, the model outperforms the API for OpenAI's GPT-4o mini-tts, which launched earlier this spring.

The second release, Qwen3-TTS-VC-Flash, can copy voices from just three seconds of audio and reproduce them in ten languages. Qwen claims the model achieves a lower error rate than competitors like Elevenlabs or MiniMax. The AI is also capable of processing complex texts, imitating animal sounds, and extracting voices from recordings. Both models are accessible via the Alibaba Cloud API. You can try demos for the design model and the clone model on Hugging Face.

Comment Source: Qwen

Matthias Bastian

Dec 23, 2025

AI in practice

Ex-Tesla AI chief Andrej Karpathy shares four tips for AI startups competing with OpenAI

Matthias Bastian

Dec 23, 2025

AI in practice

OpenAI admits prompt injection may never be fully solved, casting doubt on the agentic AI vision

OpenAI is using automated red teaming to fight prompt injections in ChatGPT Atlas. The company compares the problem to online fraud against humans, a framing that downplays a technical flaw that could slow the rise of the agentic web.

Read full article

Comment

Matthias Bastian

Dec 22, 2025

AI in practice

Google is ramping up its AI infrastructure with a major energy acquisition. Parent company Alphabet is buying clean energy developer Intersect for $4.75 billion in cash, plus assumed debt.

Alphabet is acquiring Intersect's energy and data center projects that are currently under development or construction. The company holds assets worth $15 billion. By 2028, projects with roughly 10.8 gigawatts of capacity should be online—more than twenty times the electricity output of the Hoover Dam, as Reuters reports. Intersect will continue to operate separately from Alphabet. Existing plants in Texas and California aren't part of the deal.

The deal reflects a broader trend: big tech companies are pouring money into energy assets as US power grids struggle to keep pace with soaring electricity demand from artificial intelligence. Google says it plans to double its AI capacity every six months, aiming for a thousandfold increase in output within four to five years. To hit those targets, Google is also investing in advanced reactor technology.

Comment Source: Alphabet

Matthias Bastian

Dec 22, 2025

Short News

OpenAI has reportedly made major strides in improving the profitability of its AI services. The company's compute margin—the share of revenue left after paying for server costs from paying users—jumped from around 35 percent in January 2024 to roughly 70 percent by October 2025, according to internal financial data obtained by The Information. For comparison, Anthropic is expected to reach 53 percent by year's end.

OpenAI achieved these gains by cutting rental costs for computing power, optimizing its models, and launching a pricier subscription tier. Still, the company has a long road ahead before reaching profitability. CEO Sam Altman continues to plan major investments in additional computing power while pursuing further circular business arrangements.

OpenAI is reportedly working on a funding round of up to 100 billion dollars.

Comment Source: The Information

Matthias Bastian

Dec 21, 2025

AI in practice

Kling 2.6 adds voice control and motion upgrades as AI video tools race toward realism

Matthias Bastian

Dec 21, 2025

Short News

Nvidia has released a new base model for gaming agents. NitroGen is an open vision action model trained on 40,000 hours of gameplay videos from more than 1,000 games. The researchers tapped into a previously overlooked resource: YouTube and Twitch videos with visible controller overlays. Using template matching and a fine-tuned SegFormer model, they extracted player inputs directly from these recordings.

NitroGen builds on Nvidia's GR00T N1.5 robotics model. According to the researchers, it's the first model to demonstrate that robotics foundation models can work as universal agents across virtual environments with different physics engines and visual styles. The model handles various genres—action RPGs, platformers, roguelikes, and more. When dropped into unfamiliar games, it achieves up to 52 percent better success rates than models trained from scratch.

The team, which includes researchers from Nvidia, Stanford, Caltech, and other universities, has made the dataset, model weights, paper, and code publicly available.

Comment Source: via Jim Fan

Matthias Bastian

Dec 21, 2025

Short News

Alibaba's AI unit Qwen has released a new image editing model that breaks down photos into separate, editable components. Qwen-Image-Layered splits images into multiple individual layers with transparent backgrounds (RGBA layers), letting users edit each layer independently without affecting the rest of the image.

The model handles straightforward edits like resizing, repositioning, and recoloring individual elements. Users can swap out backgrounds, replace people, modify text, or delete, move, and enlarge objects. Images can be split into either 3 or 8 layers, and the process is repeatable - each layer can be broken down into additional layers as needed. The Qwen team describes this approach as a bridge between standard images and structured, editable representations.

The Qwen team has published the code on GitHub, with models available on Hugging Face and ModelScope. More details are available in the blog post and technical report. For hands-on testing, demos are available on Hugging Face and ModelScope.

Comment Source: Blog

Matthias Bastian

Dec 20, 2025

Short News

OpenAI now lets users customize how ChatGPT communicates. The new "Personalization" settings include options for adjusting warmth, enthusiasm, and formatting preferences like headings, lists, and emojis. Each setting can be toggled to "More" or "Less." Users can also pick a base style - like "efficient" for shorter, more direct responses.

OpenAI says these settings only affect the chatbot's tone and style, not its actual capabilities. The company notes that the new options likely work as an extension of the custom instructions feature available in the same settings window.

Comment Source: OpenAI