AI in practice Archive

Feb 27, 2026

Meta has signed a multi-year, multi-billion dollar contract with Google to rent its AI chips—Tensor Processing Units (TPUs)—for developing new AI models. That's according to The Information. Meta is also looking into buying TPUs outright for its own data centers starting next year.

The deal takes direct aim at Nvidia, which dominates the AI chip market and has been Meta's go-to GPU supplier for AI training. Just days earlier, Meta had announced plans to buy millions of GPUs from Nvidia and AMD. Internally, Google Cloud executives have set a goal of capturing up to ten percent of Nvidia's annual revenue—roughly $200 billion—through TPU sales. Google has also launched a joint venture with an investment firm to lease TPUs to other customers.

Here's where it gets complicated: Google itself is one of Nvidia's biggest customers, since cloud customers still expect access to GPU servers. So Google has to keep buying Nvidia's latest chips to stay competitive in the cloud market, while simultaneously trying to eat into Nvidia's market share with its own silicon. OpenAI reportedly managed to negotiate 30 percent lower prices from Nvidia simply because TPUs exist as an alternative.

Comment Source: The Information

Jonathan Kemper

Feb 27, 2026

Short News

A new integration links Figma's design platform directly with OpenAI's Codex. Teams can automatically generate editable Figma designs from code and convert designs into working code. It runs on the open MCP standard, supports Figma Design, Figma Make, and FigJam, and is set up in the Codex desktop app for macOS.

Until now, moving between Figma and code was mostly a one-way street. Dev Mode offered basic HTML/CSS snippets, plugins exported designs as React or HTML, and Figma Make generated React components from text input. These tools worked in isolation without understanding the full project. The new integration creates an end-to-end connection where the AI accesses code, Figma files, and the design system simultaneously.

Figma was one of the first partners with its own ChatGPT app and uses ChatGPT Enterprise internally. According to OpenAI, over one million people access Codex weekly, with usage up more than 400 percent since the start of the year.

Comment Source: OpenAI | Developer Blog

Jonathan Kemper

Feb 27, 2026

Short News

Claude Code now remembers what it learns across sessions - automatically tracking debugging patterns, project context, and preferred working methods without manual input. Previously, users had to log this information themselves or use /init to populate CLAUDE.md files. The new auto-memory function builds on that that: Claude creates a MEMORY.md file per project, stores its findings, and pulls them up automatically in later sessions. Work through a tricky debugging problem once, and you won't have to explain the fix again. Users can also explicitly ask Claude to save specific information. The feature is on by default and can be disabled via /memory, the settings file, or an environment variable.

Another recent update: locally running sessions can now be continued on the go via smartphone, tablet, or browser at claude.ai/code - without data migrating to the cloud.

Comment Source: Anthropic

Matthias Bastian

Feb 26, 2026

Short News

Anthropic's AI assistant Claude is picking up new features in its desktop app Cowork. Users can now set up scheduled tasks that Claude handles automatically at set times, things like a morning briefing, weekly spreadsheet updates, or Friday presentations for the team.

Anthropic also points to the plugins already available that give Cowork specialized knowledge in areas like design, technology, and law. A full overview of available plugins is here. Moreover, there's a new "Customize" section in Cowork's sidebar where users can manage all their plugins, skills, and connections from one place.

Cowork is available as a research preview for macOS and Windows, open to all paying Claude subscribers. As with any agent-based AI system, there are cybersecurity considerations. It's worth being careful about which parts of your computer you give the software access to.

Comment Source: Claude via X

Matthias Bastian

Feb 26, 2026

Short News

Anthropic has acquired AI startup Vercept to boost Claude's computer use capabilities. Vercept built AI that works directly on a user's machine, understands screen content, and executes tasks. Founders Kiana Ehsani, Luca Weihs, and Ross Girshick are joining Anthropic with their team. The acquisition price hasn't been disclosed.

Vercept solves perception and interaction problems central to AI-driven computer use, according to Anthropic. The technology lets an AI model read and operate human-designed interfaces from screenshots without needing a dedicated programming interface (API).

Vercept will shut down its desktop AI agent "Vy" in the coming weeks. What likely caught Anthropic's attention is the startup's "VyUI" interface recognition model, which reportedly outperformed comparable OpenAI technology in benchmarks.

Benchmark (UI element identification / grounding)	VyUI accuracy	OpenAI model
ScreenSpot v1	92%	18.3%
ScreenSpot v2	94.7%	87.9%
GroundUI Web	84.8%	82.3%

Claude already handles multi-step tasks in running applications. With the recently released Sonnet 4.6 model, Claude scores 72.5 percent on OSWorld—a benchmark that measures how well AI models complete real-world computer tasks—up from less than 15 percent at the end of 2024. The Vercept team could push that number even higher.

Comment Source: Anthropic | Vercept

Matthias Bastian

Feb 26, 2026

AI in practice

Suno investor admits she ditched Spotify for AI music, accidentally undermining the company's fair use defense

Suno investor C.C. Gong told X she barely uses Spotify anymore, accidentally undermining the company’s fair use defense and handing the music industry a powerful argument in its lawsuit against the AI music startup.

Read full article

Comment

Matthias Bastian

Feb 26, 2026

AI in practice

Google's Nano Banana 2 brings Pro-level image generation to Flash speeds at up to 40% lower API cost

Matthias Bastian

Feb 26, 2026

Short News

Andrej Karpathy, former AI developer at Tesla and OpenAI, says programming with AI agents has changed fundamentally over the past two months. According to Karpathy, AI agents barely worked before December 2025, but since then they've become reliable, thanks to higher model quality and the ability to stay on task for longer stretches.

As an example, he describes how an AI agent independently built a video analysis dashboard over a weekend: he typed the task in plain English, the agent worked for 30 minutes, solved problems on its own, and delivered a finished result. Three months ago, that would have been an entire weekend project, Karpathy says.

As a result, programming is becoming unrecognizable. You’re not typing computer code into an editor like the way things were since computers were invented, that era is over. You're spinning up AI agents, giving them tasks *in English* and managing and reviewing their work in parallel.
Karpathy via X

Karpathy also points out that these systems aren't perfect and still need human "high-level direction, judgement, taste, oversight, iteration, and hints and ideas." What makes his take especially notable is how recently he held the opposite view. As late as October 2025, he called the hype around AI agents exaggerated, saying the products were far from ready for real-world use. He fundamentally changed that opinion after the release of Opus 4.5 and Codex 5.2 in the winter and is now doubling down on it.

Comment Source: Karpathy via X

Matthias Bastian

Feb 26, 2026

AI in practice

Alibaba has expanded its Qwen 3.5 model series. The lineup includes four models: Qwen3.5-Flash, Qwen3.5-35B-A3B, Qwen3.5-122B-A10B, and Qwen3.5-27B. According to Alibaba, the models deliver stronger performance while using less compute. All four take text, images, and video as input and generate text as output. The series started with the release of Qwen3.5-397B-A17B in mid-February.

The smaller Qwen3.5-35B-A3B model outperforms its much larger predecessor, Qwen3-235B-A22B; a clear sign that better architecture, data quality, and reinforcement learning matter more than raw model size. The larger 122B and 27B variants aim to close the remaining gap to top-tier models, particularly in complex agent scenarios.

Benchmarks show Alibaba's Qwen 3.5 models matching or outperforming top Western models like OpenAI's GPT-5 mini, gpt-oss-120b, and Anthropic's Claude Sonnet 4.5. The largest model, Qwen3.5-122B-A10B, leads in several tests: it tops all competitors in agent-based tool use (BFCL V4, 72.2) and agent-based web search (BrowseComp, 63.8). In the HMMT math benchmark, it scores 91.4 - just behind GPT-5 mini (92.0). It also takes the lead in visual reasoning (MMMU-Pro, 76.9) and document recognition (OmniDocBench, 89.8). Claude Sonnet 4.5, on the other hand, clearly outperforms all Qwen models in agent-based terminal coding (49.4) and embodied reasoning (64.7). GPT-5 mini leads in multilingual knowledge (MMMLU, 90.0) and math. Notably, the small Qwen3.5-35B-A3B with just 3 billion active parameters keeps up with much larger models across many tests. — Alibaba's Qwen 3.5 models match or outperform leading Western models like OpenAI's GPT-5 mini, gpt-oss-120b, and Anthropic's Claude Sonnet 4.5 across multiple benchmarks. | Image: Alibaba

All models are available on Hugging Face, ModelScope, and through Qwen Chat. They ship under the Apache License 2.0, a permissive open-source license that allows commercial use, modification, and redistribution. Qwen3.5-Flash is the hosted production version with a context length of one million tokens and built-in tools. The API costs $0.10 per million input tokens and $0.40 per million output tokens.

Comment Source: via X

Matthias Bastian

Feb 25, 2026

Short News

Perplexity has launched "Perplexity Computer," a new chat interface that pulls together multiple agentic AI models into a single system. Similar to Claude Cowork, but browser-based and with access to models from different providers, it handles entire workflows on its own.

Users describe the outcome they want, and the system spins up sub-agents for web research, document creation, data processing, or API calls. According to Perplexity, AI models are becoming increasingly specialized, so a complete workflow needs access to all of them, a convenient argument for a company built on top of other providers' models, though that doesn't make it wrong.

Perplexity Computer currently runs Opus 4.6 as its core model, supplemented by Gemini, Grok, ChatGPT 5.2, Nano Banana for images, and Veo 3.1 for video. Each task gets its own secure environment with browser, file system, and tool connections. Perplexity Computer is available as part of the Max plan at 200 dollars per month.

Comment Source: Perplexity