AI in practice Archive

Dec 16, 2025

OpenAI has updated its Realtime API with three new model snapshots designed to improve transcription, speech synthesis, and function calling. According to developers, the gpt-4o-mini-transcribe variant significantly reduces hallucinations. For text-to-speech tasks, gpt-4o-mini-tts cuts the word error rate by 35 percent. The gpt-realtime-mini model, which targets voice assistants, follows instructions 22 percent more accurately and improves function calling by 13 percent.

? New audio model snapshots are now live in the Realtime API with improvements to reliability, lower error rates, and fewer hallucinations:

- gpt-4o-mini-transcribe-2025-12-15: 89% reduction in hallucinations compared to whisper-1

- gpt-4o-mini-tts-2025-12-15: 35% fewer word... pic.twitter.com/E8clreR1R0

- OpenAI Developers (@OpenAIDevs) December 15, 2025

OpenAI also explicitly mentioned improvements for Chinese, Japanese, Indonesian, Hindi, Bengali, and Italian.

Comment

Maximilian Schreiner

Dec 16, 2025

AI in practice

Nvidia is taking over software provider SchedMD to expand its presence in open-source technology. On Monday, the company confirmed it will continue to distribute SchedMD's "Slurm" software as an open-source product. The platform helps plan large-scale computing tasks in data centers, ensuring server capacity is used efficiently.

Nvidia views the technology as critical infrastructure for generative AI, noting that developers rely on it to train models. Financial terms of the deal were not disclosed. Founded in California in 2010, SchedMD employs around 40 people and serves clients like cloud provider CoreWeave and the Barcelona Supercomputing Center.

Comment Source: Nvidia

Jonathan Kemper

Dec 16, 2025

AI in practice

Google is linking its NotebookLM research tool directly to the Gemini chatbot. This integration lets users select specific notebooks as context for their Gemini queries, effectively expanding the chatbot's knowledge base beyond its initial training data and standard web results. While NotebookLM already includes a built-in chat function powered by a Gemini model, it remains quite limited—it doesn't even save chat histories. The new feature addresses this by allowing users to leverage multiple notebooks simultaneously within the main Gemini interface. It also supports integration with "Gems," the personalized versions of the chatbot. The rollout appears to be gradual, starting with browser users, though app support will likely follow soon.

NotebookLM started as an experimental tool in 2023. It has since established itself as a software with exemplary AI integration, particularly in the education sector. The tool makes it easy to set up RAG environments and thus make large document collections analyzable and searchable. Google regularly adds new functions to NotebookLM, most recently including one for deep research.

Comment Source: via X

Maximilian Schreiner

Dec 15, 2025

AI in practice

Robot vacuum pioneer iRobot has filed for bankruptcy and plans to hand control to its main Chinese supplier, Shenzhen PICEA Robotics. According to Bloomberg, shares of the Roomba maker will be wiped out under the bankruptcy plan. While the company will be delisted, it intends to continue operations as a going concern.

To set up the deal, Shenzhen PICEA acquired $191 million of iRobot's debt from the Carlyle Group. iRobot attributed the filing to a post-pandemic sales slump, supply chain issues, and stiffer competition from cheaper rivals. The move comes after a planned acquisition by Amazon fell apart in 2022 following opposition from EU regulators.

The company listed assets and liabilities between $100 million and $500 million. In a statement, iRobot confirmed it would continue paying employees and suppliers throughout the court proceedings.

Comment Source: Bloomberg

Matthias Bastian

Dec 14, 2025

AI in practice

Adobe has integrated Photoshop, Acrobat, and Express directly into ChatGPT's interface. Users can now edit images and documents for free using text commands. The Photoshop integration lets people customize photos with simple descriptions; changing backgrounds or adding effects, for example. Adobe Express handles design tasks like creating invitations from templates, while Acrobat makes it possible to edit PDFs like resumes right in the chat.

To set it up, go to "Apps & Connectors" in ChatGPT's settings, select the Adobe app you want, and click "Connect." Then tap the plus sign in the chat, find the app under "More," and type your command. Alternatively, type "/AdobePhotoshop," "/AdobeExpress," or "/AdobeAcrobat" followed by what you want to do.

Adobe says commands work best when they're clear and specific, with complex tasks broken into individual steps. After each edit, sliders let users adjust the results.

Comment Source: Adobe

Jonathan Kemper

Dec 14, 2025

AI in practice

LongCat-Image proves 6B parameters can beat bigger models with better data hygiene

An image generated by the AI shows the words "LongCat-Image" in the form of colorful, fluffy plush letters standing on a carpet in a bright children's room, surrounded by small fantasy creatures, also made of plush.

Matthias Bastian

Dec 14, 2025

AI in practice

OpenAI wants to boost risk tolerance among its workforce. According to The Wall Street Journal, the company has scrapped a rule requiring new hires to stay for at least six months before their equity vests. The change aims to ease employee concerns about being laid off before receiving their first batch of shares. Previously, OpenAI had already shortened this waiting period from 12 months to six in April.

The move underscores the fierce competition for AI talent. Tech giants like Meta, Google, and Anthropic are courting top researchers with high compensation. OpenAI is set to spend around $6 billion on stock-based compensation this year, nearly half its projected revenue. These high personnel costs are putting additional pressure on margins in an increasingly competitive market.

Comment Source: WSJ

Matthias Bastian

Dec 13, 2025

AI in practice

Google is integrating Gemini into Google Translate for better text translations and launching a beta for real-time voice translation through headphones. Gemini now handles idioms, local expressions, and slang more naturally instead of translating them word for word. The improved text translation is rolling out in the US and India for English and nearly 20 languages, including Spanish, Hindi, Chinese, Japanese, and German. The app is available for Android, iOS, and on the web.

The live translation feature taps into Gemini's speech-to-speech capabilities to preserve the speaker's tone, intonation, and rhythm. The beta is currently available on Android in the US, Mexico, and India, supporting over 70 languages. iOS and more countries will follow in 2026.

Google is also bringing its language learning tools to nearly 20 new countries, including Germany, India, Sweden, and Taiwan.

Comment Source: Google

Matthias Bastian

Dec 13, 2025

AI in practice

OpenAI appears to be adopting the skills system Anthropic introduced in October, according to a discovery by user Elias Judin shared on X. Support for these skills has surfaced in both the Codex CLI tool and ChatGPT.

Judin found directories named "pdfs" and "spreadsheets" containing "skill.md" files. These files provide specific instructions for processing documents and data. It's basically like your prompt calling a more specific prompt to solve a complex subtask necessary for the main goal—like extracting text from a PDF. Since it's just a folder containing a Markdown file and maybe scripts, it's easy to adapt.

A look at the "skill.md" file for PDF handling reveals specific instructions for reading and creating documents. | Image: Elias Judin via GitHub

The file structure suggests OpenAI is organizing AI tools into app-like modules designed for specific tasks. Judin, who found the feature while using a "5.2 pro" model, documented the findings on GitHub. Anthropic debuted this modular system in October to help its Claude assistant handle specialized tasks.

Comment Source: Github | Judin via X

Matthias Bastian

Dec 13, 2025

AI in practice

OpenAI claims its team built the Sora Android app in just 28 days by leveraging its code-generation AI, Codex. According to a report from OpenAI employees Patrick Hum and RJ Marsan, a small team of four engineers utilized an early version of the GPT-5.1 Codex model to build the application, processing around five billion tokens along the way.

According to the authors, the AI handled the bulk of the actual writing—specifically tasks like translating existing iOS code into Android-compatible formats. This allowed the human developers to focus on high-level architecture, planning, and verifying the results. The team described Codex as acting like a new, experienced colleague that just needed clear instructions to get the job done. Despite the rapid timeline, OpenAI reports the app is 99.9 percent stable. You can read a detailed breakdown of their process on the OpenAI blog.

Comment Source: OpenAI