JetBrains Open-Sources Mellum2, a Coding-Specialized Model

JetBrains has published a technical report on Mellum2, a coding-specialized model designed for AI development workflows, and released the model itself as open source.

Mellum2 is built on a 12B-parameter Mixture-of-Experts (MoE) architecture with only about 2.5B parameters active at inference time — a deliberately lightweight design. According to JetBrains, its coding performance reaches the same level as Qwen 3.5 9B reasoning models. That said, it reportedly falls short of even Qwen 3.5 4B on general-purpose tasks outside coding, underscoring a deliberate trade-off: this model is specialized for coding and nothing else.

JetBrains already offers AI completion features in its own IDEs, and open-sourcing Mellum2 marks a symbolic step in the broader trend of developer-tool companies giving their in-house models back to the community. The model is available on Hugging Face.

OpenAI Sued by Florida Attorney General Over AI Harms

Florida's Attorney General has sued OpenAI. The legal action is grounded in AI harms, and a US state directly prosecuting an AI company could set an important precedent.

The Wall Street Journal has covered the details of the lawsuit, though much remains unknown, including the specific harms alleged and the damages sought. Either way, how the judiciary defines the scope of AI companies' responsibility is a question that will directly shape the future direction of AI regulation.

AI-related litigation has so far centered on copyright infringement, but a state Attorney General filing suit over consumer protection and safety-related harms suggests that regulation is entering its next phase.

DuckDuckGo Makes Its "No AI" Search Engine Easier to Access as Traffic Booms

DuckDuckGo has released browser extensions for Chrome and Firefox that provide easier access to its "no AI" search engine, which serves search results free of AI-generated content.

The company's traffic has been performing strongly — TechCrunch reports the service is enjoying a "traffic boom." As Google and Bing accelerate the integration of AI Overviews into search results, there is demonstrably real demand from users deliberately choosing "no AI," and that demand is driving inflow to alternative search engines.

This may look like a counter-trend at first glance, but it can also be read as a sign of a split among users. We may see growing polarization between those who want to embrace AI actively and those who value the purity of information.

Bain Survey: AI Investment Is a "Circular Bet," with ROI Falling Short of Expectations

Bain & Company's latest survey has laid bare the reality of enterprise AI investment. Many companies had forecast cost savings from AI, but actual results have fallen far short of those predictions.

The survey describes this situation as a "circular bet": a company invests in an AI company, that AI company supports another enterprise's AI adoption, and that enterprise in turn invests in AI — whether real productivity gains are happening anywhere in this loop is far from clear.

Meanwhile, according to research highlighted by Jack Clark of Import AI, the US "AI economy" has reached roughly $250 billion in nominal GDP (as of 2025), with quality-adjusted real growth estimated at around 2,600% per year. The reality is that it is growing rapidly at a scale that is nearly invisible in macro GDP statistics.

A circular investment structure on one side, and breakneck growth that is hard to see in the macro data on the other. The gap between these two views may be the most honest picture of AI investment today.

Zig Project Bans AI-Generated Code Contributions Outright

The project lead of the Zig programming language has announced a policy of banning AI-generated code contributions outright. The blunt criticism of AI-generated code as "invariably garbage" drew significant attention on Hacker News.

This is less a matter of anti-AI sentiment than a reflection of circumstances unique to systems programming. In low-level languages like Zig, the finer details of memory safety and performance are critically important, and code that looks superficially correct often contains subtle bugs. The judgment here is that the review cost of AI-generated code exceeds its value.

There was also recent news of the QEMU project loosening its AI code restrictions — attitudes toward AI code vary widely depending on a project's nature and maturity. A one-size-fits-all conclusion has yet to emerge.

Strava Tightens API Limits in Response to AI Apps and Scrapers

Strava, the fitness tracking app, has tightened restrictions on API access. The reason: a surge in "zero-code AI apps and scrapers."

According to The Verge, the proliferation of AI tools has caused explosive growth in apps and scrapers using Strava's API, raising both the load on infrastructure and the risk of data misuse.

This shows how AI's ripple effects are reaching into unexpected areas. It is not the AI apps themselves but the side effect of AI tools making it possible for anyone to build apps on top of APIs — and that is now forcing platforms to respond.