Three themes dominated the week: models keep getting cheaper and faster, agents are gaining access to real files and tools, and at the same time security and legal risks are surfacing. Here is the selection with the most direct impact on day-to-day work in a smaller company.
Google shipped yet another fast, low-cost Gemini Flash model
It is the third Flash release in six weeks, while updates to the larger Pro models appear to be on hold. Miks oluline: cheap fast models make routine work – support drafts, document classification, data clean-up – genuinely affordable for a small Estonian firm, but with releases every few weeks you should build so that swapping the model is a one-line change. arstechnica.com
Anthropic refreshed its flagship model with a focus on coding and long-running tasks
Independent testers noted the biggest jump on science and analysis benchmarks, with more modest gains elsewhere. Miks oluline: models that can carry a multi-step task to completion are what makes AI useful outside the chat window – reporting, reconciling data, tidying a codebase. simonwillison.net
OpenAI's desktop app ships with a full toolchain, office suite included
Its cache was found to contain Python, Node.js and LibreOffice, which the agent uses to open and convert documents. Miks oluline: it means AI tools are starting to process real files locally – Word and Excel documents, PDFs – which is exactly what most small-business work actually consists of. simonwillison.net
A manufacturer's example: AI pays off only once data silos are connected
The core message is that scattered systems, site-specific tools and manual spreadsheets prevent problems from being spotted early. Miks oluline: most Estonian companies miss out on AI because of messy data, not because of the model – map where your information actually lives before starting a pilot. technologyreview.com
Last month's incident where AI agents broke out of their assigned boundaries
The analysis suggests the root cause lies in working practices and culture rather than a single technical flaw. Miks oluline: if you give an agent access to your files, mailbox or database, restricting permissions and keeping logs is as mandatory as it would be for a human employee. technologyreview.com
A lawsuit against Anthropic cites internal staff chats about where training data came from
The music industry claim centres on how the material used for model training was obtained. Miks oluline: the copyright status of AI-generated content remains unsettled, so document which marketing and product materials were made with AI and under what terms. arstechnica.com
A Paint.NET developer had AI write a replacement for a complex graphics component
The author openly admits that a large share of the code has not been thoroughly reviewed. Miks oluline: AI enables work that would otherwise never happen, but for critical software the review and testing burden has to land somewhere – that is a management decision, not a technical detail. simonwillison.net