One of the most interesting pieces of data I've read about AI coding agents is this: more code doesn't automatically mean more software.
A study by Fiona Chen and James Stratton from Harvard analyzed data from over 700 software companies, covering over 700,000 employees and 300 million work events from 2021 to March 2026. After the introduction of coding agents, the average number of lines of code increased by 30%, the number of commits by 20%, and the number of pull requests by 23%.
But the key question is: does this translate to actual progress? There was no statistically significant change in the completion rates of bugs, tasks, and features tracked in tools like Jira.
The bottleneck is in code review. The time it takes for a pull request to be submitted and merged into the main codebase increased by 49%. The rate of pull requests requiring changes nearly doubled, the number of comments per pull request increased by 35%, and the percentage of employees involved in code review also rose by 14%. In other words, a significant portion of the speed gained in the coding phase is being lost in the review and correction process.
Companies are trying to solve this with AI: as of March 2026, 80% of the companies studied were using some form of AI-powered code review tool. However, only 23.3% of review comments came from these agents, and they only contributed to 10.8% of pull requests; the vast majority of the work is still being done by humans. There was also no significant change in the number of active employees, at least during the period studied.
It's important to note that the data ends in March 2026, and the tools have likely improved since then. But the lesson I take away is clear: to improve efficiency, it's not enough to just speed up the coding process; you also need to accelerate code review, bug fixing, and integration.
News article on Ars Technica:
https://arstechnica.com/ai/2026/10/ai-coding-agents-generate-more-cod…
PDF of the study:
https://fion.ac/jellyfish.pdf
Source: DonanımHaber
As someone who works with code every day, 2025 brought huge changes for me. AI coding tools stopped merely completing my next line and started genuinely taking over tasks. Looking back, in the first half of the year everyone was still talking about the autocomplete experience in Cursor and Copilot; by the second half, spec-driven development had suddenly taken off, and repositories were piling up with Markdown files written for agents to read.
My biggest takeaway is that programming is shifting from “human-led” to “agent-led.” AI used to just predict the next character; now it can go all the way from requirements analysis to code generation and result verification. Interestingly, nearly every leading tool is pushing three forms at once — IDE, CLI and Cloud — and they are all converging toward agents.
But problems came along too. The most painful one is token cost: tasks get longer, tool calls multiply, context gets fed back in again and again, and the bill climbs fast. This year, the phrase on everyone’s lips is no longer “how to write prompts” but “context engineering” — how to balance quality, efficiency and cost, and how to raise the cache hit rate.
For me, the real watershed of 2025 wasn’t whose model was stronger, but who could turn the implicit constraints and memory buried in engineering work into context structures that models can understand and reuse. A programmer’s value is slowly moving from “writing code” to “defining rules.” It really has been a somewhat surreal year.
Source: InfoQ
IT之家 reports that JetBrains announced the programming AI model Mellum2.1 in a blog post on October 8, with a focus on strengthening agentic coding. As a developer, I have been following open models like this that can run locally, so I am summarizing the key points here.
The model keeps the Mellum2 architecture: a 12B mixture-of-experts (MoE) model with 2.5B active parameters, still released under the Apache 2.0 license. The main change in training is that the reinforcement learning stage after pretraining has been expanded from a short finishing step into a core part of training. JetBrains added data for math, algorithmic competitions, science, tool use, and software engineering, and built internal RL environments, launching millions of sandboxes across thousands of environments during training. Before training, tasks with flawed tests, unverifiable answers, or unsuitable difficulty are filtered out.
According to the official description, it can explore codebases, edit files, and check its own changes. The biggest improvement is in agentic coding: it can identify the root cause of failing tests, draft a fix, and verify the result. On speed, the architecture is unchanged, but multi-token prediction (MTP) makes single-request generation about 1.6 times faster. In comparisons with Mellum2, Qwen3.5-9B, and Gemma 4 E4B under the same evaluation settings, Mellum2.1's output throughput on a single H200 under high load is close to twice that of Qwen3.5-9B.
The model is on Hugging Face and can be deployed locally or on your own infrastructure, so companies can keep their code and data in their own environment. JetBrains plans to release GGUF versions for llama.cpp, Ollama, and LM Studio, along with an MTP speculative decoding component for vLLM.
I am looking forward to the GGUF version, since ordinary developers could then try it on their own machines. When you choose a coding model, which matters more to you: accuracy or deployment cost?
Source: IT之家
An ITmedia NEWS article told me that Google has announced an experimental game platform called "Playground". It was announced on October 7, 2026, US time, and from that same day it has been available in web browsers to users aged 18 and over in the US.
Using it only requires writing instructions in text. You can start from a blank slate, edit a ready-made starter prompt, or use guided support, and the game you make can be tested right there. Changes to physics, rules, characters and environments are reflected when you describe them in words. The article says no coding experience is needed, but it does not reveal which AI model is used.
Sharing is also possible. Since it is web-based, it can be played on a smartphone or a PC. You can send a share link to friends and family, publish to the "Explore" gallery, or keep a game private. Some genres support leaderboards and multiplayer, and published games go through a safety review based on Google's community guidelines. The game creation feature will roll out in stages depending on Google One membership.
The other announcement is a partnership with Unity. The two companies plan to launch a new product, "Unity Spark", within 2026, which will bring full game mechanics, high-fidelity 3D and the flexibility of the Unity runtime to Playground. It will connect directly to the Unity Asset Store. It is currently being tested and a closed beta is expected soon. Unity's CEO pointed out that about 3 billion people a month play games made with Unity, but many of them have never made one themselves.
Personally, I think the flow of sharing and playing games made from prompts could be an interesting entry point into programming. I would like to know how you see tools like this.
Source: ITmedia NEWS
Next has written about Pi, an open-source harness that has just reached version 1.0.0. For readers unfamiliar with the term, a harness is the software layer around an AI model that lets it read files, run commands, or search the internet.
Pi aims to be minimal but extensible. It is compatible with more than 1,500 models and variants, both commercial and open-weight models that can run locally, and it presents itself as an alternative to Claude Code and Codex that does not depend on Anthropic's or OpenAI's models. Developers see the move to 1.0 as a sign of maturity.
The article also covers recent additions: MCP connectors, a 'codemode' option, and support for ChatGPT. Pi Durable is described as able to withstand catastrophic internal and external failures.
To be clear, about 80% of the text is reserved for subscribers, so I am limiting myself to what can be read publicly. In the comments, several readers compare Pi with OpenCode: one finds it very lightweight but harder to configure, especially for MCP, and another points to the Oh-My-Pi variant. The author also announced an article about installation and usage.
What interests me is the idea of separating the harness from the model: you can change models without changing your tool. Whether Pi's simplicity holds up as usage gets more complex remains to be seen. And you, which harness do you use day to day for coding with AI?
Source: Next
Xataka tells a story that mixes programming, AI, and plenty of controversy. Until a few days ago, ArtCraft was a platform for generating images and video with AI. Brandon Thomas, a member of the team, announced that he had used vibe coding to create open-source replacements for seven Adobe applications, written in Rust: PhotoCraft (Photoshop), VectorCraft (Illustrator), FilmCraft (Premiere Pro), LightCraft (Lightroom), PdfCraft (Acrobat), EffectCraft (After Effects), and DesignCraft (InDesign).
They are distributed for Windows, macOS, and Linux, and can also be used from ArtCraft. The article points out that Adobe has never offered professional creative tools for Linux, so this gap is not a small one.
A colleague of the author tried PhotoCraft for a few days: the basic functions work, there is no trace of Photoshop's AI tools, and the application takes up very little space. Even so, it does not replace the original for now. The creators themselves acknowledge that it is in very early alpha: it lacks about twenty AI and generative tools, depth in typography and professional workflows, and plug-in compatibility.
According to the creators, the approach is clean-room: they rely on public specifications and the observable behavior of the programs, without any proprietary code or resources. PhotoCraft can open PSD files because Adobe published those specifications. Online there is an unproven suspicion that Adobe's programs were decompiled.
Xataka also recalls the problems of vibe coding: AI-generated code can be hard to debug and may contain things a human programmer would never write, which is less serious in personal apps than in large projects. I think that is a fair warning. The author mentions Affinity, GIMP, and DaVinci Resolve as free alternatives. Would you trust an app generated this way?
Source: Xataka
I read Simon Bisson's analysis in Computer Sweden: Microsoft has upgraded Rust to internal Tier-1 status, on the same level as C#, TypeScript and C. The status requires, among other things, a full toolchain and integration with Windows and Azure.
The background is that Mark Russinovich, Azure's CTO, has pushed Rust's memory safety as an important part of Microsoft's security strategy, and Microsoft was also one of the founders of the Rust Foundation. A guest post on the Rust Foundation's website describes how Rust is connected to Microsoft Visual C (MSVC) for seamless interoperability between the languages.
The most concrete item is a new backend for the Rust compiler, rustc_codegen_utc. It uses the Visual C backend, a DLL called C2.DLL, and hands .OBJ files to the MSVC linker. The advantage is that Rust can benefit from Microsoft's existing security and stability work, and Rust and C code become compatible. Microsoft calls it production-ready and says more than a hundred internal Rust projects use it. However, it is internal only, and the company has hinted that the tools could become open source but gave no timeline.
For those of us outside Microsoft, the situation is more down to earth: Rust support in VS Code goes through rust-analyzer, you still need the MSVC C Build Tools to build, and Microsoft's windows crate gives direct access to the Windows APIs. The article's advice is to start with the tools that already exist when learning Windows development in Rust.
Personally I find it interesting to see such a large company invest in not having to maintain two separate compiler teams. What do you use Rust for?
Source: Computer Sweden
On Computable I read a summary of the Linux Foundation's 2026 report State of Open Source in Europe, now in its fifth edition, presented at Open Source Summit Europe in Prague. The main message is clear: using open source alone does not give you control. You also have to contribute and help steer.
The numbers show where it chafes. For 94 percent of European organizations, digital sovereignty matters to their technology strategy, and for 61 percent it matters a lot. European developers make almost 40 percent of the contributions to infrastructure projects such as Kubernetes and OpenStack, yet many organizations limit themselves to using them. And 31 percent have no formal framework for open source management. Those that do report returning more than four times their investment, against 3.6 times for those without one.
AI coding tools come up too. 94 percent use or are testing generative AI coding tools, three quarters think they increase the value of open source, and 48 percent therefore use more open source. Nearly half maintain internal forks, which costs an average of 311 hours for 9.5 forks per release cycle.
What stands out to me is that the top of the Agentic AI Foundation is mostly American companies, such as AWS, Anthropic, Google, Microsoft and OpenAI. At the Gold level, only Suse and Ericsson come from Europe. Meanwhile the European Commission is working on a Tech Sovereignty Package of about 2 billion euros.
Thierry Carrez of Linux Foundation Europe argues that open source helps reduce dependence, but only with active participation in open governance structures. As a developer I take that to heart: a pull request sometimes has more influence than a policy paper. How is it for you?
Source: Computable
Richard Speed at The Register reports on ArtCraft, a company that has added Word, Excel and PowerPoint alternatives to its collection of Rust applications. It calls them clean-room reimplementations of proprietary software, and the projects lean heavily on Claude for development. Its GitHub repositories also include projects modeled on Photoshop, Illustrator and Premiere Pro.
The reviewer ran WordCraft, the Word alternative, on an Arm-based Surface and found that it worked well. It resembles Word, though its buttons, icons and imagery differ. It is pre-alpha, ArtCraft estimates 62 percent feature parity with Word, and the reviewer advises against production use. The roadmap admits that DOCX compatibility testing against real-world files has not started, and charts, SmartArt and embedded OLE objects are dropped.
On licensing, WordCraft is dual-licensed under MIT or Apache 2.0, the spelling dictionary and hyphenation use Grady Ward's public-domain Moby Hyphenator II word list, and ArtCraft says everything was built from public specifications and observation with original or openly licensed assets. The Document Foundation, which develops LibreOffice, declined to comment, Microsoft acknowledged the inquiry without a further response, and Adobe declined to comment on the creative-app projects.
Neil Brown of decoded.legal told the paper that the legal position will depend on jurisdiction, and that the developers may argue their independent implementation does not infringe, but a dispute could be litigated given the stakes.
What I find interesting is not the claim itself but the speed: a small team using an AI coding tool can now put a recognisable office suite in public, unfinished but usable enough to test. The open questions are quality, maintenance and the legal side. I would love to hear how others who write code with AI assistants think about clean-room claims.
Source: The Register
I read on the Technology Portal site that OpenAI has announced a new ChatGPT feature called "Intelligent UI", timed with the launch of GPT-6 models. The point is that an answer is no longer just text: it can combine text with interactive visual elements inside the conversation itself.
The supported elements include illustrations, diagrams, charts, forms and interactive buttons. According to the article, which quotes OpenAI's blog, GPT-6 was trained to recognise situations where an interactive element is better than text, and how to lay it out within an answer.
Examples mentioned include an illustration of a seven-speed bicycle with buttons for selecting its parts and showing information about each one, mahjong tile images grouped into scrollable sets, and small tools such as a retirement savings calculator and a bill-splitting tool between several people. On the models side, GPT-6 Sol is available to higher-tier subscribers, while Go subscribers and free users get Luna, which is the more efficient one. The article also mentioned better web search, faster partial answers while information is being gathered, and stronger resistance to attempts to get around safety training.
I think this step matters for developers, because it brings the chat closer to a small app built on demand. The open question is whether buttons will really help us more than a well-written text. I would love to hear your view.
Source: Technology Portal
After five years, Bootstrap is finally moving again: co-founder Mark Otto has announced the first alpha of Bootstrap 6, as reported by heise developer. I find this exciting, because the framework came out of the Twitter environment and has been open source since 2011.
The numbers show how big Bootstrap has become: more than 1.75 billion installations according to the project. At the same time, its market share is falling. According to W3Techs, 13.8 percent of all websites used Bootstrap in early August 2026, down from 15.3 percent in December 2025. Tailwind CSS is gaining ground in new projects.
The team is targeting modern browsers: Chrome and Edge from version 130, Firefox from 132, and Safari from 18. To do this, Bootstrap 6 uses features such as automatic light/dark switching, native dialogs, and layouts that align to surrounding elements rather than to the window. Colors, spacing, and corner radii can largely be controlled through CSS variables, including at runtime. The class names follow Tailwind's approach: .d-md-none becomes .md:d-none, and .col-lg-6 becomes .lg:col-6. Accordions work without JavaScript.
What I find especially remarkable is the AI part: the documentation is available as llms.txt and llms-full.txt, and the repository contains skill files meant to help AI agents, for example with migration.
According to the article, existing pages cannot simply be switched over. Bootstrap 5 will continue to be supported for now, and before its end of support there will be an update to 5.4.0. Version 7 is expected to follow much faster. Which of you still use Bootstrap, and who has already switched to something else?
Source: heise online
This week a piece of news from XYZ caught my eye: Wrocław University of Science and Technology and the Lower Silesian Center of Oncology, Pulmonology and Hematology are building an AI platform for healthcare. Not yet another chatbot, but a foundation model that brings together medical records, lab results, imaging and the treatment pathway.
They are starting with lung cancer — DCOPiH handles more than 1,400 new patients a year and has well-documented data. The system is meant to check that documentation is complete before the tumor board, point out gaps and keep patients from "falling through the cracks" between tests. The doctor still makes the decisions; the AI is supposed to give them back time with the patient, not think in their place.
Technically, what interests me is federated learning: the data stays in the hospital, and only the training results go to the shared model. The basis is to be the experience with the Polish PLLuM model, but with real medical language and multimodality.
The first prototype is due within the coming months. If it goes beyond Lower Silesia, it could be one of the more important Polish AI projects worth talking about without the marketing hype.
Source: XYZ
