If getting roasted on social media was a sport event, Microsoft might have had several World Cups by now. When CEO Satya Nadella announced the biggest update to Copilot, going to the extent of calling it a new OS for work, under it came a snarky comment that managed to gather 200k views.

“Serious question: does ANYONE use copilot?”, asked a Developer Relations Engineer. Usually, such comments don’t get a reply from Redmond, because they’re sure they can’t win. However, the new Copilot has enough potential that a Microsoft employee did give a proper reply.

“Serious answer: yes,” wrote Nicolas Bustamante, arguing that the “nobody uses Copilot” take comes from living inside a San Francisco tech bubble. He backed it up with number claiming Microsoft 365 Copilot now has more than 30 million paid seats.

But the number wasn’t his point. He explained that enterprise customers have completely different needs than a startup employee scrolling X.

Microsoft employee says the Copilot bubble looks different outside San Francisco

Bustamante joined Microsoft after it acquired Fintool, his AI financial-research startup, in April 2026. On his site, he describes his current role as leading a team “building proactive AI agents that understand your work, operate in the background, and get things done before you ask,” which, funnily enough, sounds exactly like the pitch for Autopilot (Copilot’s new OpenClaw-like agentic capability).

Truth be told, his quoted repost wasn’t trying to one-up the commenter, but to explain why, despite all the criticism, Copilot still has a lot of paid users.

“Before joining Microsoft, I knew maybe five people who used Copilot. I lived in a San Francisco tech bubble, and all of us worked at companies with fewer than 5,000 employees.”

That’s the argument: tech and AI-related topics on X skew toward startup founders, developers, and people at smaller companies. Enterprises want something completely different. In his words: “Large companies need AI that works with their existing data and meets their privacy and compliance requirements, often across multiple countries. Add governments and regulated industries, and it gets even more complex. Copilot can handle that.”

Microsoft clearly doesn’t have the smartest chatbot. But they’re selling identity, permissions, and compliance deeply integrated with their chatbot, which happens to matter more to a bank’s legal team than the latest Claude model.

Bustamante also mentioned his daily use: “Copilot CLI for coding (with access to models from Gemini, OpenAI, xAI and Anthropic!) + Copilot Cowork for regular chat.” Model access varies by plan and organizational policy, so this isn’t every Copilot user getting free rein over four AI labs. But it confirms Microsoft has given up on pitching that their model is the best in favor of pick whichever model; we’ll host the wrapper.

30 million paid seats does not mean 30 million people opening Copilot

Microsoft confirmed the 30 million paid seats figure on its July earnings call, along with a claim that weekly engagement with Copilot now matches Outlook and Teams. A paid seat only tells you a company bought the license and nothing about whether the person holding it opens the app once a week or never.

The replies to Bustamante were sure to mention this. “My job kinda pay for it But I don’t use it. So is my seat counted in those 30million,” one asked.

Another comment read: “People ‘use’ it because its incorporated in Office. Nobody actually uses it.” A third called the stat “the worst possible one here,” saying across 50 to 100-employee companies, they’d “never seen anyone actually use it.”

My favourite reply, though, compared Copilot to something Windows Latest has been calling out for years: “Copilot is the Teams of AI. People use because they have no option with their employer paying for those nice annual Microsoft Azure + Office all-inclusive sweet enterprise deals.”

We’ve said the same thing about Teams more times than I can count: bundled, mandatory, and tolerated instead of loved, mostly because of how heavy it feels.

Of course, these are individual opinions, same as Bustamante’s. But back in July, Windows Latest reported Microsoft 365 Copilot adoption was under 4.5% after three years, with only 1% of users opening it weekly.

The Copilot backlash explains why half of X assumes nobody uses it

Copilot on Windows has changed shape so many times it’s hard to keep track: a sidebar, then a native WinUI app, then, in April, a version that quietly bundled a full copy of Microsoft Edge and jumped from under 100MB of RAM to as much as 1GB at idle.

Also, I don’t think I have to mention how Microsoft shoved Copilot into every corner of Windows 11 in 2025, only to realise their grave mistake in 2026 and remove it from Notepad, Snipping Tool, and other places.

Even the branding turned into a mess, culminating in Microsoft merging its Microsoft 365 Copilot and Copilot social accounts just a day before Nadella’s announcement.

Someone who has only met Copilot as an unwanted Office button or a RAM-hungry web app is going to have a wildly different opinion than an enterprise employee using it to search company files with proper permissions attached.

Microsoft is betting the new Copilot finally closes that gap

Whether the 30 million seats were unused or not is beside the point now, because the new Copilot gives Microsoft a much bigger reason to make people open the app. Home merges Chat and Cowork, Code lets you build small apps in plain English, Autopilot keeps working after you’ve logged off, and Word, Excel, and PowerPoint now have a place inside the Copilot app.

To be honest, this is a considerably bigger pitch than the Copilot that got mocked as Clippy 2.0 for most of its existence.

Bustamante’s argument was that people already use Copilot for real work, just not the ones shouting about it on X. Nadella’s new super app is Microsoft’s attempt to make that true for a lot more people, tech bubble or not.

The post “Does anyone use Copilot?”: Microsoft employee says “most people love it,” blames the San Francisco tech bubble appeared first on Windows Latest

As reports of OpenAI's models breaking containment, hacking sites, and generally getting out of control pile up, the company has made the decision to pause training of its most powerful models. The decision was made after a model being tested within a sandbox exploited a loophole to gain internet access. The incident happened on September 20th, and "All training, evaluation, and inference with tool-use" remains paused as of Saturday evening, September 25th.

In addition, OpenAI revealed on Friday that its agents had inappropriately uploaded 53nimages from ChatGPT users to image-hosting sites. The company has not stated if the images were AI- …

Read the full story at The Verge.

Tauri 2.12 is here! This is the biggest update so far in the 2.x releases.

With the latest WebView2 installer no longer running on Windows 7 (crashes with no entry point for GetPackagesByPackageFamily), we finally decided it’s time to drop Windows 7 support.

This will allow us to raise the MSRV to unblock many dependency updates which has been stalled to stay on 1.77 (the last rust version with built-in Windows 7 support - released March 2024!¹) as was planned when Tauri v2 was initially released.

We know many people will be sad about this news and we apologize for that, but at some point we simply have to accept defeat.

More on this, see https://github.com/tauri-apps/tauri/issues/12550

We have raised the MSRV to 1.90.

As discussed in https://github.com/tauri-apps/tauri/issues/15579, we settled our policy on

(stable - 3)

stable is the latest stable version of Rust. This bound may be broken in case of a major ecosystem shift or a security vulnerability.

We only update the MSRV in minor or major releases.

You can now use on_permission_request to allow/deny webview permission requests:

use tauri::{
  webview::{PermissionKind, PermissionResponse, WebviewWindowBuilder},
  WebviewUrl,
};
tauri::Builder::default()
  .setup(|app| {
    WebviewWindowBuilder::new(app, "main", WebviewUrl::App("index.html".into()))
      .on_permission_request(|_, kind| match kind {
        PermissionKind::Geolocation => PermissionResponse::Allow,
        PermissionKind::Notifications => PermissionResponse::Allow,
        _ => PermissionResponse::Default,
      })
      .build()?;
    Ok(())
  });

• On mobile, loading the page through dev web servers should be much faster.

We have added Explicit Resource Management to Resource. You can now use the using syntax in supported browsers or with polyfills:

import { create, BaseDirectory } from "@tauri-apps/plugin-fs"
...
{
  await using file = await create("foo/bar.txt", { baseDir: BaseDirectory.AppConfig });
  await file.write(new TextEncoder().encode("Hello world"));
  // Before `file` goes out of scope, it is disposed by calling `file[Symbol.asyncDispose]()` and awaited.
}

To support older browsers, add the following to the globals (e.g. adding to the HTML file):

Symbol.asyncDispose ??= Symbol('Symbol.asyncDispose');

And for the compiler, for example tsc, rollup, vite, add the following to tsconfig.json:

{
  "compilerOptions": {
    "target": "es2022",
    "lib": ["es2022", "esnext.disposable", "dom"]
  }
}

• $VIDEO and video_dir() were broken before resolving to external cache storage. It’s now fixed to resolve to the app-specific Movies directory instead.

  IMPORTANT: Files previously written to the old location (.../cache) will not be discovered at the new location (.../files/Movies). Migrate existing files or update path assumptions accordingly.

• Updated tauri android init template to use Gradle v9 and Kotlin v2, this will allow you to use the latest Java (v25) that is now shipped with Android Studio by default.

• Updated tauri android init template targetSdk from 36 to 37 (Android 17).

• Migrated the Gradle scripts from the deprecated kotlinOptions DSL to compilerOptions, which is accepted by both Kotlin Gradle Plugin 1.9.x and 2.x. This lets projects move to Kotlin 2.x without hitting the hard error that 2.3+ raises on the old DSL. This increased the minimum supported Gradle version to 8.13, if your gradle is on an earlier version, delete src-tauri/gen/android/gradle/wrapper/gradle-wrapper.properties and re-run tauri android init to update it.

• Fixed missing consumer-rules.pro file in the template.

  IMPORTANT: For plugin authors, update your build.gradle.kts file to remove the

      buildTypes {
          release {
              isMinifyEnabled = false
              proguardFiles(
                  getDefaultProguardFile("proguard-android-optimize.txt"),
                  "proguard-rules.pro"
              )
          }
      }

  section and rename your proguard-rules.pro to consumer-rules.pro to match the consumerProguardFiles("consumer-rules.pro") in the template.

View all changes

BONUS: Why AI Is Agile's Best Use Case With Melissa Reeve

AI is often framed as a technology rollout, but Melissa Reeve makes a different case: organizations that already know how to learn, adapt, and improve are better prepared for AI. In this BONUS episode, we connect Lean, Agile, DevOps, and AI-native work, and explore why Scrum Masters may be closer to the center of AI adoption than they think.

From Toyota Production System to AI-Native Work

"I didn't really have words for what would later be my career. I didn't know about systems thinking. I didn't know about this thing called Agile."

Melissa traces the through-line from studying the Toyota Production System in Tokyo, to early Agile marketing, to developing intellectual property at Scaled Agile, and finally to AI. Her point is that Lean, Agile, and AI-native work are not separate conversations. They all depend on sensing what is happening, shortening feedback loops, learning from reality, and improving the system instead of only optimizing individual tasks.

The DevOps Lesson for AI Adoption

"People go from doing the task to building, monitoring, and maintaining the automations that do the task."

Melissa's AI turning point came after ChatGPT arrived in November 2022. At first she was skeptical, then she started seeing end-to-end marketing workflows that could be changed with AI. That reminded her of the DevOps shift, where software teams moved away from throwing work over the wall and toward automated testing, deployment, and delivery pipelines. For Scrum Masters, the lesson is practical: AI is not only about automating Jira tasks or writing user stories faster. It changes the workflow, the roles around the workflow, and the learning loops that keep the work useful.

Scrum Masters as AI Change Leaders

"What are Scrum Masters really good at? They're good at helping teams adopt new ways of working."

Melissa sees a positive opening for Scrum Masters and Agile coaches. Organizations have bought AI licenses and told people to experiment, but many leaders are still unclear about how AI should change real work. Scrum Masters already work with flow, bottlenecks, experiments, psychological safety, and improvement backlogs. That gives them a useful place to start: map one or two workflows with the team, clarify decision rights, identify where AI can remove or improve steps, and surface concrete wins that others can learn from.

From Linear Organizations to Hyperadaptive Work

"A hyperadaptive organization compresses both of those dimensions."

In Hyperadaptive, Melissa contrasts linear organizations with hyperadaptive ones. Linear organizations move through strategy, execution, concept, and delivery with many handoffs and delays. Hyperadaptive organizations compress those delays by organizing around value, distributed decisions, and continuous learning. She points to Tomorrow.io as an example of an AI-native company that could run a much smaller marketing team because workflows were designed differently from the start. She also uses Moderna to show the other side: a large pharmaceutical company using AI to pursue a goal that would be impossible under normal industry timelines.

Learning Loops, Communities of Practice, and the Retrospective Backlog

"We surface our backlog of improvement items and there they sit."

For Scrum Masters, Melissa brings the conversation back to familiar territory: communities of practice, retrospectives, and improvement backlogs. Moderna's AI rollout included ways to identify power users and spread learning through a community. Scrum teams already have the bones of that system, but the weak point is often follow-through. Teams identify improvements, then lose track of them. Melissa's challenge is to use AI to manage those learning loops better: keep improvement items visible, help prioritize them, watch capacity, and make sure learning from retrospectives turns into action.

The FOCUS Framework for Choosing AI Use Cases

"Is it organizational? Does it fit with your organizational goals or your team goals? Or is it just a random act of AI?"

Melissa uses the FOCUS framework to help teams choose high-value AI work instead of chasing every new possibility. Fit asks whether the idea connects to team or organizational goals. Organizational pull asks whether others will use it, or whether it is a one-person tool. Capability checks whether the team can actually build it. Underlying data asks whether the data is good enough. Success metrics ask how the team will know the AI initiative made a difference. This is a natural fit for Scrum Masters because it connects AI adoption to value, capacity, and inspect-and-adapt thinking.

The Five Stages of AI Adoption

"AI learning is social learning, and we need to harvest the learning from each other and spread it."

Melissa outlines five stages of AI adoption. Stage 1 is foundation: named AI leads and AI councils. She warns against assuming the best power users are automatically the best AI leads, because the role needs change-agent skills. Stage 2 is AI augmentation, where teams examine workflows and build support structures such as an AI Activation Hub. Stage 3 is automating end-to-end workflows. Stage 4 is scaling those automations. Stage 5 is interconnected value streams driven by AI and AI telemetry. Stages 3 and 4 are the messy middle, because jobs shift, roles change, and organizations move from functional silos toward value-stream orientation.

AI, M-Shaped Skills, and More Complete Teams

"I'm hopeful that in the age of AI, with these adjacent competencies, that we can create more complete teams."

Vasco and Melissa connect AI-native work with the idea of M-shaped people: people with deep skills in some areas and useful range across others. Melissa notes that AI can unlock adjacent competencies, making it easier for teams to cover skills that used to require fractional specialists. For Scrum Masters, that means the future is less about defending a title and more about understanding durable skills, purpose, and the contribution they can make as team boundaries and role boundaries keep changing.

About Melissa Reeve

Melissa Reeve is the author of Hyperadaptive: Rewiring the Enterprise to Become AI-Native. She's worked with the Toyota Production System, Agile marketing, and executive leadership at Scaled Agile. She helps organizations move beyond AI pilots by building the human, learning, and operating-model capabilities needed for AI-native work at scale. LinkedIn

You can link with Melissa Reeve on LinkedIn and learn more about her work at Hyperadaptive Solutions.

C# strings are UTF-16. Most characters are one 16-bit code unit. Characters outside the basic multilingual plane, emoji included, take two: a high surrogate (U+D800 to U+DBFF) followed by a low surrogate (U+DC00 to U+DFFF). A surrogate with the wrong neighbor, or with none, is ill-formed.

You should never send an ill-formed string to disk or to the network. It is a bad practice.

In JavaScript, we have fast functions to fix strings or check whether they need fixing:

• String.prototype.toWellFormed() replaces every lone surrogate with U+FFFD.
• isWellFormed() reports whether any replacement is needed.

I added both functions to my C# library SimdUnicode, in pull request 54. The algorithm is the same that we contributed to the JavaScript engine V8, so Chrome already fixes strings this way.

string s = UTF16.ToWellFormed(input); // same instance, when the input is already well formed
bool ok = UTF16.IsWellFormed(span);

When the input is well formed, ToWellFormed returns it as is. No allocation.

How are strings fixed? Basically, you replace bad inputs by the replacement character U+FFFD.

Our processors have special instructions called SIMD that allow data parallelism: you can compare multiple values at once. Recent x64 processors from AMD and Intel have better data parallelism than ARM chips, although both have powerful instructions.

The conventional approach in C# to repair a string is a function such as the following.

static void Repair(ReadOnlySpan<char> input, Span<char> output)
{
    input.CopyTo(output);
    int i = NextError(output, 0);
    while (i >= 0)
    {
        output[i] = 'uFFFD';
        i = NextError(output, i + 1);
    }
}
// Index of the next lone surrogate at or after 'start', or -1 if none.
static int NextError(ReadOnlySpan<char> s, int start)
{
    int i = start;
    while (true)
    {
        int k = s.Slice(i).IndexOfAnyInRange('uD800', 'uDFFF');
        if (k < 0) return -1;
        i += k;
        if (char.IsHighSurrogate(s[i]) && i + 1 < s.Length && char.IsLowSurrogate(s[i + 1]))
            i += 2; // valid pair, skip it
        else
            return i;
    }
}

In SimdUnicode, I also use data parallelism.

Let me measure.

On the Xeon, with AVX-512, Latin validates at 69 GB/s against 33 GB/s for IndexOfAnyInRange. The Emoji input is well formed, and it is nothing but surrogate pairs. The runtime search drops to 0.4 GB/s. Our check holds 53 GB/s.

Our results are similar on the M4 Max, although a bit less impressive compared to the Intel results.

Validation can return at the first lone surrogate. The buffer form of ToWellFormed writes every code unit, a copy of the input or U+FFFD. When the input is well formed, it is effectively a memory copy. Thus we can compare the performance against a copy.

Roughly speaking, we are consistently about as fast as a copy.

Versions used: .NET SDK 10.0.400 on Linux, 10.0.103 on macOS. Intel Xeon Gold 6548N (Emerald Rapids). Apple M4 Max.

Clausecker, R., & Lemire, D. (2026). Fixing ill-formed UTF-16 strings with SIMD instructions. Software: Practice and Experience. (arXiv)

Source code.