Ad
Skip to content
Subscribe Now
Matthias Bastian
Matthias Bastian View the LinkedIn Profile of Matthias Bastian
Mar 5, 2026
Short News
Copy the url to clipboard Share this article Go to comment section
Read full article about: Alibaba's chief AI developer quits, taking key team members with him

Alibaba's lead AI researcher Junyang Lin has unexpectedly resigned. Lin was the driving force behind Alibaba's Qwen model series and shaped the company's open-source strategy. According to Chinese technology portal 36Kr, several core team members followed him out the door, including Binyuan Hui (Qwen coder), Bowen Yu (post-training), and Kaixin Li (Qwen 3.5/VL). A number of younger researchers reportedly also quit on the same day.

via X

Alibaba CEO Eddie Wu accepted Lin's resignation and announced a new "Foundation Model Task Force," which he will lead alongside CTO Zeming Wu and Jingren Zhou, The Information reports. The company says it plans to double down on open source and ramp up AI investments. "In technology, standing still means falling behind," Wu wrote in a memo to the team. "In technology, standing still means falling behind. Advancing foundation models is a core strategic priority for our future."

Comment Source: The Information | 36kr
Matthias Bastian
Matthias Bastian View the LinkedIn Profile of Matthias Bastian
Mar 4, 2026
Short News
Copy the url to clipboard Share this article Go to comment section
Read full article about: OpenAI's Codex app lands on Windows after topping a million Mac downloads in its first week

OpenAI has released its Codex app for Windows. Codex is an AI-powered coding tool that helps developers build software by running multiple agents asynchronously across projects, delegating repeatable tasks through Automations, and connecting agents to tools and workflows via Skills. Developers can review, guide, and step into agent work without losing context.

For the Windows version, OpenAI built its own sandbox that operates at the OS level with restricted tokens, file system access rights, and dedicated sandbox user accounts. This lets AI agents run directly in Windows environments like PowerShell without forcing developers to switch to WSL or virtual machines. OpenAI has published the sandbox code as open source on GitHub.

The Windows app arrives a few weeks after the Mac version, which was downloaded over a million times in its first week, according to OpenAI. More than 500,000 developers had already signed up for the Windows waiting list. OpenAI says Codex now has over 1.6 million weekly active users total. The app is available across all ChatGPT plans.

Comment Source: OpenAI
Matthias Bastian
Matthias Bastian View the LinkedIn Profile of Matthias Bastian
Mar 4, 2026
Short News
Copy the url to clipboard Share this article Go to comment section
Read full article about: GPT-5.4 reportedly brings a million-token context window and an extreme reasoning mode

OpenAI's GPT-5.4 could be the leap forward that the just-released 5.3 Instant for ChatGPT wasn't. The model should drop "sooner than you think," according to OpenAI, but no official details have been shared yet.

via X

According to The Information, GPT-5.4 will feature a one-million-token context window, more than double the 400,000 tokens in the current GPT-5.2. That would put OpenAI on par with Google and Anthropic. The model is also expected to be more reliable and make fewer mistakes on longer tasks that can run for several hours, which matters especially for tools like OpenAI's Codex programming agent.

One notable addition is an "extreme" thinking mode that lets the model burn significantly more compute on tough questions. This mode is aimed at researchers rather than everyday users who want quick answers. According to The Information, the more frequent model release cadence is designed to keep expectations in check. The hype around the GPT-5 launch set the bar so high it was nearly impossible to clear, and OpenAI's user growth has recently fallen short of internal projections.

Comment Source: The Information
Maximilian Schreiner
Maximilian Schreiner View the LinkedIn Profile of Maximilian Schreiner
Mar 4, 2026
Short News
Copy the url to clipboard Share this article Go to comment section
Read full article about: Meta creates new applied AI engineering division

Meta is building a new applied AI engineering organization, according to an internal memo obtained by the Wall Street Journal. The new teams will be led by Maher Saba, currently a vice president in Meta's Reality Labs division, and will report to CTO Andrew Bosworth.

The structure is designed to be extremely flat, with up to 50 employees per manager. The new division will work alongside Meta's Superintelligence Lab to build the "data engine" that speeds up improvements to Meta's AI models.

According to Saba, the organization consists of two teams: one focused on interfaces and tools, and another on tasks, data collection, and evaluations.

Meta restructured its AI operations last summer, creating the Superintelligence Lab under former Scale AI CEO Alexandr Wang. CEO Mark Zuckerberg announced in January that the company would release new models and products in the coming months.

Comment Source: WSJ
Maximilian Schreiner
Maximilian Schreiner View the LinkedIn Profile of Maximilian Schreiner
Mar 4, 2026
Short News
Copy the url to clipboard Share this article Go to comment section
Read full article about: Anthropic nears $20 billion revenue run rate despite Pentagon feud

Anthropic is on track to generate nearly $20 billion in annual revenue based on current performance, according to Bloomberg. The company recently surpassed $19 billion in run-rate revenue, more than doubling its $9 billion run rate from the end of 2025. Just a few weeks ago, the figure stood at roughly $14 billion. The growth is driven by strong adoption of Anthropic's AI models and products, particularly its coding tool Claude Code. The company is currently valued at $380 billion.

At the same time, Anthropic is locked in a conflict with the US Department of War. Secretary of War Pete Hegseth declared Anthropic a supply-chain risk, a designation typically reserved for companies from countries the US views as adversaries. The move came after Anthropic pushed for restrictions on the Pentagon's use of its AI for surveillance and autonomous weaponry. Former White House adviser Dean Ball called the designation "attempted corporate murder." Anthropic said the move is "legally unsound" and is prepared to challenge it in court. The company's main app recently topped Apple's download charts, reflecting a surge of support during the Pentagon standoff.

Comment Source: Bloomberg
Maximilian Schreiner
Maximilian Schreiner View the LinkedIn Profile of Maximilian Schreiner
Mar 4, 2026
Short News
Copy the url to clipboard Share this article Go to comment section
Read full article about: OpenAI is building a GitHub competitor that could challenge its biggest investor

OpenAI is building its own alternative to GitHub, Microsoft's widely used platform for code management and collaboration, according to The Information. The move was triggered by a string of GitHub outages in recent months that directly affected OpenAI's own developers. The project is still in its early stages and likely won't be ready for several months.

Internally, the company is reportedly discussing whether to offer the product to outside customers - a move that would put OpenAI in direct competition with Microsoft, one of its biggest investors. GitHub has attributed some of its outages to problems with Microsoft's Azure cloud, which the service is currently being migrated to. Google and Meta both run their own internal code platforms but don't offer them externally. OpenAI, Microsoft, and GitHub all declined to comment.

Comment Source: The Information
Matthias Bastian
Matthias Bastian View the LinkedIn Profile of Matthias Bastian
Mar 3, 2026
AI in practice
Copy the url to clipboard Share this article Go to comment section

Google's fastest and cheapest model Gemini 3.1 Flash-Lite got smarter but also tripled the price

Google Deepmind has released a preview of Gemini 3.1 Flash-Lite, the fastest and cheapest model in the Gemini 3 series. It’s significantly more capable than its predecessor, but output costs have more than tripled.

Read full article
Comment
Matthias Bastian
Matthias Bastian View the LinkedIn Profile of Matthias Bastian
Mar 3, 2026
Short News
Copy the url to clipboard Share this article Go to comment section
Read full article about: OpenAI releases GPT-5.3 Instant for smoother everyday conversations and better search

OpenAI has released GPT-5.3 Instant, an update to the standard ChatGPT model. The new version aims to make everyday conversations feel more natural and useful. According to OpenAI, the model delivers more accurate answers, better web search results, and fewer unnecessary warnings and refusals. Hallucination rates drop by up to 26.8 percent for web searches and 19.7 percent for internal knowledge, depending on the scenario. The writing style also feels less robotic and preachy, OpenAI claims.

The system card shows some trade-offs on the safety front. GPT-5.3 Instant beats the older GPT-5.1 Instant on average when it comes to catching unauthorized content, but it actually performs worse than its direct predecessor, GPT-5.2 Instant. The model also takes a small hit on health-related queries (HealthBench) compared to the previous version.

GPT-5.3 Instant is rolling out now to all ChatGPT users and is available to developers via the API as "gpt-5.3-chat-latest." The outgoing GPT-5.2 Instant will stick around for paying users for another three months before OpenAI pulls the plug on June 3, 2026.

Comment Source: System Card | Blog