Week 2025-29
It's important to know where the next battles will be fought, especially when they concern our privacy. @vlkodotnet
Insight of the Week: The Battle for Context Begins
While last week I mentioned the battle for data, a second battle has simultaneously erupted. What matters more than the data itself is the context surrounding that data. Large language models work with probability, and context is precisely what helps one model produce higher-quality output than another. Even a simple question like "What will the weather be like?" has different answers depending on whether I just discussed vacation plans with my wife or I'm standing in front of my closet deciding what to wear.
Anyone who has worked with an AI coding agent knows that specifying the exact files related to your problem helps the AI model generate better responses with enriched context. Your operating system has the most context about everything, which Apple and Google certainly understand—they just face privacy concerns. Creating an operating system where users would voluntarily sacrifice their privacy just to get better AI responses probably won't be feasible in the short term.
The approach must be different. Browsers will be the Trojan horse through which AI companies break into your privacy. Browsers are where we search, work, read, listen, and watch content. Nothing is more tempting than an agent that promises to do work for you. So we'll give it access to emails, calendars, and documents. Eventually, because it will be incredibly practical, even to our credit cards. We humans will do anything get rid of boring work, especially if someone offers it for free.
Microsoft has had a Copilot button in Edge browser for some time, but it's only for manual activation (for now?). Perplexity is launching the Comet browser, which beyond traditional features like summarization and content questions, can also handle task automation.
OpenAI is preparing Project Operator, which will have its own built-in browser. You simply tell it what you want, and it handles ordering clothes, planning and booking vacations, comparing prices, etc.
It won't be long before every AI company is forced to offer some version of their own browser, or the competition will steamroll them. And privacy? Well, that will suffer considerably. AI will be a helper on every page, with every search, when reading every email. My personal prediction is that alongside cloud AI, a small and fast AI model integrated into the browser will assist, handling preprocessing or simple tasks.
AI Work Staff Reshuffling
Working at a startup means either getting well paid (which often doesn't happen) or exchanging your work for less money, hoping for future value from employee stock options. This assumes you believe the startup owner won't abandon you, since they hold considerably more shares than you do. This model worked for years, but now with AI companies, it's starting to crumble. Zuckerberg started this trend by acquiring ScaleAI just to get its founder, Alexander Wang.
That was the better case because even though ScaleAI certainly won't continue its previous work, at least employees gained something since it was an acquisition.
The Windsurf case showed the opposite side. Recently there was talk of OpenAI acquiring them, but the deal fell through after Windsurf founders announced they were joining Google DeepMind. They reportedly agreed on licensing Windsurf technology, but anyone who understands knows that employee stock options don't pay out through such licensing. Executives get signing bonuses, but most employees who relied on an exit or IPO face an uncertain future. The company will likely slowly fade away.
For large companies, acquiring talent this way is simpler—just throw lots of money at it. Those "lots of money" are still less than acquisition costs and antitrust problems. This means that instead of employee stock options, AI startup employees will start preferring salaries to reduce risk. This will result in fewer AI startups being created due to lack of funding. When I think about it, isn't that exactly what these massive corporations want? Less competition?
Business Insights
Nvidia became the first company in history to reach a $4 trillion valuation. If you own Nvidia stock, this certainly pleases you. I expect they could grow for another year or two since companies still need more AI hardware. But there's risk here. If future breakthroughs in small on-device AI models succeed, entire AI data centers could be left abandoned.
Nvidia also had some negative publicity. The Ukrainian military discovered that Nvidia Jetson Orin began powering Shahed MS001 drones. It helps them identify targets and map territory. This happened despite official export restrictions to Russia. Nvidia confirmed they are aware of this and added that if they discover who's exporting these to Russia, contracts will be immediately terminated.
Apple sued Jon Prosser, who became known for leaking unreleased iOS 26 information. He allegedly obtained an employee's access code, tracked their location, and when the employee wasn't home for an extended period, secretly accessed their phone with the developer version (which obviously couldn't be taken on vacation). According to Jon Prosser, it happened differently and he obtained information from unknown sources.
Incidentally, Australia introduced age verification for using search engines. This means that in Australia, before searching anything on the web, you'll need to go through some form of more or less anonymous identity verification.
YouTube content creators will certainly welcome the following news. YouTube will restrict monetization of content created with generative AI.
I'm not saying AI will cause the demise of SEO companies, but they'll have it much harder. The first casualty is World of Good Brands (formerly known as Leaf Group or Demand Media Inc.). Once a successful company that lived by serving internet content that brought traffic to products they offered themselves. Their downfall was mainly due to a series of unsuccessful projects, and the current wave of AI search was just the final nail in the coffin.
AI Insights
Vercel, which offers AI-generated products, published AI research results. So take it with a grain of salt. But it's interesting reading, revealing that 88% of users use OpenAI. 60% of users changed their AI model provider in the last 6 months. So everyone goes where there's a better/cheaper model. 71% use vector databases, and apparently we're past the phase of mindless AI enthusiasm, meaning today AI needs to present real problem solutions—chatbots are obsolete.
We have a new AI agentic IDE called Kiro. It's different because instead of prompting, you create specifications and hooks. Specifications allow you to define what you want the application to do, and hooks verify the functionality of the solution.
Kimi K2 is a free model with an agentic approach. You probably won't be testing it at home—you'll need either an H200 or a cluster with 16 GPU cards. But the test results look quite good.
Mistral introduced two new features. Le Chat got Deep Research mode, voice prompting capability, and image generation and editing.
The mentioned voice feature is powered by the new Voxtral model. Voxtral Mini should be an excellent choice for video-to-text transcription.
Finally, for those who don't want to try Cursor or other IDEs but would still like to test an AI agent, here are experiences from 2 weeks of using Claude Code.
.NET Insights
We finally got the sixth preview version of .NET 10. Not much new for web developers was added. JSON got some new deserialization rules, API validation got a separate package, and Blazor WebAssembly got preloading, improved form validation, and diagnostics.
Link Drop
JavaScript is 30 years old, which is still less than my age, so it still has things to learn. What do you think were its 10 milestones that allowed it to survive to this day? I got into it somewhere between milestones 3 and 4.
So that not everyone blames Amazon for only using open-source projects and giving little back to the community, I discovered the Active-active Replication Extension for PostgreSQL (pgactive). This is useful for systems where we can ensure data access sharding while still wanting complete synchronization. Pgactive enables active replication—the ability to influence what happens to data during conflicts.
AmazingHand is a project that lets you create your own robotic hand for under $150.
A visually beautiful project that currently has only one section about how screens work.
Closing Visual
Not everyone understands the difference between correlation and causality.
