AI in practice Archive

Jan 28, 2026

AI startup Decart has unveiled Lucy 2.0, a real-time video transformation model. The system can modify live video at 30 frames per second in 1080p resolution with near-zero latency. Users can swap characters, place products, change clothing, and completely transform environments - all controlled through text commands and reference images while the video is still running.

According to Decart, Lucy 2.0 doesn't rely on depth maps or 3D models. Instead, the system's understanding of physics comes entirely from patterns learned during video training. A new technique called "Smart History Augmentation" prevents image quality from degrading over time, letting the model run stably for hours, the startup says.

The technology runs on AWS Trainium3 chips. A demo is available at lucy.decart.ai.

Comment Source: Decart

Maximilian Schreiner

Jan 28, 2026

AI in practice

Google Deepmind gives Gemini 3 Flash the ability to actively explore images through code

Maximilian Schreiner

Jan 28, 2026

Short News

OpenAI has launched Prism, a free AI workspace for scientific writing. The tool runs on GPT-5.2 and combines a LaTeX editor, reference manager, and AI assistant in a cloud-based environment. Researchers can create unlimited projects and invite collaborators.

The AI has access to the entire document and can help with writing, editing, and structuring. Users can search and incorporate academic literature from sources like arXiv. Whiteboard sketches or handwritten equations can be converted directly to LaTeX via image upload. Real-time collaboration with co-authors is also supported.

Prism is based on Crixet, a LaTeX platform that OpenAI acquired. The tool aims to eliminate the need to switch between different programs like editors, PDFs, and reference managers. Prism is available now for anyone with a ChatGPT account at prism.openai.com. Availability for Business and Enterprise plans will follow later.

Comment Source: OpenAI

Matthias Bastian

Jan 27, 2026

AI in practice

OpenAI CEO Altman admits he broke his own AI security rule after just two hours, says we're all about to YOLO

Matthias Bastian

Jan 27, 2026

Short News

AI research institute Allen AI has released SERA, a family of open-source coding agents designed for easy adaptation to private code bases. The top model, SERA-32B, solves up to 54.2 percent of problems in the SWE-Bench-Test Verified coding benchmark (64K context), outperforming comparable open-source models.

According to AI2, training takes just 40 GPU days and costs between $400 to match previous open-source results and $12,000 for performance on par with leading industry models. This makes training on proprietary code data realistic even for small teams. SERA uses a simplified training method called "Soft-verified Generation" that doesn't require perfectly correct code examples. Technical details can be found in the blog.

The models work with Claude Code and can be launched with just two lines of code, according to Allen AI. All models, code, and instructions are available on Hugging Face under the Apache 2.0 license.

Comment Source: AllenAI | Technical Report

Maximilian Schreiner

Jan 27, 2026

AI in practice

Mistral AI has unveiled Mistral Vibe 2.0, an upgrade to its terminal-based coding agent powered by the Devstral 2 model. The tool enables developers to control code using natural language, orchestrate multiple files simultaneously, and leverage full codebase context.

New in version 2.0 are custom subagents for specific tasks like testing or code reviews, clarifying questions when instructions are ambiguous instead of automatic decisions, and slash commands for preconfigured workflows.

Mistral Vibe is available through Le Chat Pro ($14.99/month) and Team plans ($24.99/seat). Devstral 2 moves to paid API access – free usage remains available for testing on the Experiment plan. For enterprises, Mistral additionally offers fine-tuning, reinforcement learning, and code modernization services.

Comment Source: Mistral

Matthias Bastian

Jan 27, 2026

AI in practice

Former Tesla AI chief Andrej Karpathy now codes "mostly in English" just three months after calling AI agents useless

Just last October, Andrej Karpathy dismissed AI agents: “They just don’t work.” Now he says 80 percent of his coding is agent-based and calls it the “biggest change to my basic coding workflow in ~2 decades.” A typically measured voice is joining the agent coding hype, but with some warnings attached.

Read full article

Comment

Matthias Bastian

Jan 26, 2026

Short News

OpenAI is charging around $60 per 1,000 impressions for its initial ChatGPT ads, far above typical online advertising rates in the low single digits and closer to what advertisers pay for premium TV spots like NFL games, according to The Information. The ads show up below ChatGPT responses in the free and lower-cost "Go" tiers.

OpenAI is also reportedly charging per impression rather than per click. Advertisers typically prefer click-based billing because it's easier to measure results. The decision to go with impressions likely reflects how AI chatbot users behave differently than traditional search users: they click on external links far less often. Perplexity uses the same approach, also charging per 1,000 impressions.

The move toward advertising—at premium prices and in a format that's less appealing to advertisers—suggests OpenAI needs to ramp up revenue quickly to justify its high valuation to investors. Sam Altman previously called ChatGPT advertising a last resort and a potential dystopia.

Comment Source: The Information

Matthias Bastian

Jan 26, 2026

AI in practice

MCP Apps, the Model Context Protocol's first official extension, turns AI responses into interactive interfaces

Matthias Bastian

Jan 26, 2026

Short News

Microsoft has unveiled its new AI inference chip, Maia 200. Built specifically for inference workloads, the chip delivers 30 percent better performance per dollar than current-generation chips in Microsoft's data centers, the company claims. It's manufactured using TSMC's 3-nanometer process, packs over 140 billion transistors, and features 216 GB of high-speed memory.

According to Microsoft, the Maia 200 is now the most powerful in-house chip among major cloud providers. The company claims it delivers three times the FP4 performance of Amazon's Trainium 3 while also outperforming Google's TPU v7 in FP8 calculations—though independent benchmarks have yet to verify these figures.

Microsoft says the chip already powers OpenAI's GPT 5.2 models and Microsoft 365 Copilot. Developers interested in trying it out can sign up for a preview of the Maia SDK. The Maia 200 is currently available in Microsoft's Iowa data center, with Arizona coming next. More technical details about the chip are available here.

Comment Source: Microsoft