21–23 May 2024, Seattle Convention Center and online (hybrid). Build 2024 ran three days in Seattle, with roughly 4,000 people in the room, around 200,000 registered online, and something on the order of 60 separate announcements crammed into the programme. We were there, as we have been every year since 2012, and we left with a clear sense that this was a transitional Build: the year cloud-native .NET stopped being a preview hobby and became something you could put your name on in production, and the year Microsoft learned — publicly, painfully — what happens when you ship an AI feature that reads everything on your screen.
This is the long version of a short post we published the same week. We have gone back, checked the dates and numbers against the official record, and corrected a couple of things we got slightly wrong in the rush.
.NET Aspire reached general availability
The announcement that mattered most to us landed on Day 1, 21 May 2024: .NET Aspire became generally available. Microsoft's own framing was deliberately plain — "Aspire is a new stack that streamlines development of .NET cloud-native services and is now generally available."
One small correction to our original post, for the record: we called it "Aspire 1.0," and that is not how Microsoft branded it. The GA release shipped aligned with the existing .NET versioning (it was Aspire 8.0, not a fresh 1.0), and the official blog never used a "1.0" label. We are flagging that because getting the small things right is the whole job — if we are sloppy about a version number on our own blog, why would you trust us with your release pipeline?
What Aspire actually is, stripped of the marketing: an opinionated set of tools, project templates, and NuGet packages for building observable, distributed, cloud-native .NET applications. The centrepiece is the app host — a single C# project that describes your whole system (the API, the worker, the database, the cache, the message broker) as code, wires up service discovery and configuration between them, and gives you a local dashboard with traces, logs, and metrics out of the box. If you have ever spent an afternoon hand-rolling connection strings and docker-compose files to get a multi-service app running on a teammate's laptop, Aspire is the answer to exactly that pain.
We had been using the Aspire previews on customer projects since late 2023, so the GA was less a discovery than a vindication. The AppHost model we had been assembling by hand was suddenly the officially recommended way to do it. We rebuilt the threebit.io site onto Aspire the same week — if you look at the src/ThreeBIT.AppHost project in this repository, that is what you are looking at.
One honest caveat we will repeat, because Microsoft was careful about it too: the built-in Aspire dashboard is a near-real-time local view, not a replacement for a full APM system. For production observability you still wire Aspire's OpenTelemetry output into something like Azure Monitor or Application Insights. Aspire gives you the plumbing and the standards; it does not pretend to be your monitoring stack.
Copilot+ PCs: Windows grows a neural processing unit
The consumer-facing headline of the week was a new class of Windows hardware. Copilot+ PCs are machines built around a neural processing unit (NPU) fast enough to run AI models locally — the class bar is 40+ TOPS (trillions of operations per second), with the launch silicon delivering around 45 TOPS. The pitch is that a meaningful slice of AI work no longer has to round-trip to the cloud: it runs on the laptop, offline, at low power.
The most interesting piece for developers was Phi-Silica, a 3.3-billion-parameter small language model built specifically to run on those NPUs. It is the first language model Microsoft shipped to run locally inside Windows itself, designed for short, fast, device-resident tasks rather than acting as a cloud chatbot. The on-device direction is the one we care about: a model small and cheap enough to embed inside a product, capable enough to be genuinely useful, running without a per-token cloud bill or a network dependency, is a much better fit for a lot of our customers' regulated, sometimes-offline environments than a call out to a frontier model.
Underneath sat the Windows Copilot Runtime and the Windows Copilot Library — a set of on-device model APIs (Phi-Silica among more than 40 models) that developers can call from their own Windows apps. This is the part that turns "Copilot+ PC" from a marketing tier into a platform you can actually build on.
The broader Phi-3 family
Build 2024 was also a coming-of-age moment for Microsoft's small-model family. Phi-3-mini and Phi-3-medium reached general availability in Azure AI's model-as-a-service offering, Phi-3-small became available, and Phi-3-vision — a 4.2-billion-parameter multimodal model that reasons over images, charts, graphs and tables — arrived in preview.
The throughline with Phi-Silica is consistent and, we think, correct: not every problem needs the biggest possible model. A small model that is cheap to run, easy to deploy, and good enough at the specific task in front of it often beats a giant one you cannot afford to call at scale. That is a message a German software shop building line-of-business systems on a budget is very happy to hear.
Recall: the feature that got a reckoning
Now the part Microsoft would rather we did not dwell on, and the part we think honesty requires us to dwell on.
Alongside Copilot+ PCs, Microsoft announced Recall — a feature that takes periodic snapshots of your screen and builds a searchable, AI-indexed timeline of everything you have done on your PC. On stage it looked like magic: ask your computer in plain language about that document you saw three weeks ago, and it finds it.
Then security researchers got their hands on it, and the magic curdled fast. Kevin Beaumont examined the preview and described it bluntly as a "keylogger built into Windows." The original design stored its screenshots and extracted text in a local SQLite database that was effectively unencrypted — meaning any malware that got onto the machine, or anyone with local access, could read a complete visual history of the user's activity, passwords and all. Beaumont demonstrated how an attacker could automate the exfiltration of that database. For a feature pitched at every Copilot+ PC, on by default, this was about as bad as a privacy story gets.
Microsoft did what, in fairness, a company should do when it gets this badly wrong: it backed down. On 7 June 2024 it published an update committing to make Recall opt-in, and on 13 June 2024 it pulled Recall from the broad Copilot+ PC launch entirely, sending it back to the Windows Insider Program first. The redesigned feature required explicit opt-in during setup (off by default), gated access behind Windows Hello biometric authentication and proof of presence, and encrypted the snapshots with "just in time" decryption so they are only readable while the authenticated user is actually looking at them.
We are recounting this because it is the most instructive story of the whole conference, and because it is exactly the kind of decision we make for customers all the time. The lesson is not "AI bad." The lesson is that a feature which is delightful in a demo can be catastrophic the moment you ask the boring questions — where is this data stored, who can read it, what happens when the machine is compromised, is it on by default. Those questions are not a tax on innovation. They are the job. Recall shipped without good answers to them, the internet supplied the answers instead, and Microsoft had to rebuild it in public.
Agents, Team Copilot, and the rest of the AI slate
The rest of the AI announcements pointed clearly at where 2025 and 2026 would go, even if most of them were still previews:
- Copilot agents in Copilot Studio entered an Early Access Program — copilots that can independently orchestrate tasks tailored to a specific role or business process, rather than just answering questions. This is the seed of the full agent platform Microsoft would build out over the next two years.
- Team Copilot extended Microsoft 365 Copilot from a personal assistant into a shared team member — meeting facilitator, collaborator, project manager — announced for preview later in 2024.
- GitHub Copilot Extensions previewed, letting third-party services (Azure, Docker, Sentry and others) plug directly into Copilot Chat, with GitHub Copilot for Azure arriving in private preview to help build, troubleshoot and deploy Azure apps from inside Copilot.
- Azure AI Studio reached general availability as the hub for building and operating generative-AI apps — the platform that would later be renamed and grow into Azure AI Foundry.
- Real-Time Intelligence in Microsoft Fabric previewed as an end-to-end SaaS solution for analysing high-volume, time-sensitive data.
The keynote framing
Satya Nadella set the tone in the opening keynote, returning to the argument he keeps coming back to — that the current wave is driven by the same kind of compounding curve that drove the last one:
"Just like Moore's Law helped drive the information revolution, the scaling laws of [deep neural networks] are really… driving this intelligence revolution." — Satya Nadella, Build 2024 keynote.
You can quibble with how durable the scaling-laws argument turns out to be — plenty of people have, in the two years since. But as a framing for why Microsoft was reorganising its entire developer stack around AI, it was coherent, and the product slate backed it up.
Why it mattered for us
Two years on, with the benefit of hindsight, here is the honest accounting of Build 2024 from where we sit in Ibbenbüren.
Aspire's GA was the cleanest "we picked the right stack" moment in years. We had bet on the previews; the GA turned that bet into the recommended path. Every Aspire-based project we have shipped since traces back to that week. That one is unambiguously a win.
The small-model thread aged extremely well. Phi-Silica and the Phi-3 family pointed at a future of cheap, embeddable, on-device intelligence that is a far better fit for our customers' constraints than a permanent dependency on the largest, most expensive model available. We have leaned into small models repeatedly since, and Build 2024 is where that conviction firmed up.
Recall was the cautionary tale we have quoted to clients ever since. When a customer asks for an AI feature that quietly hoovers up data because it would be convenient, "remember Recall" is shorthand for the whole conversation about defaults, encryption, and consent. The feature is fine now. It is fine now because it got torn apart first.
Strip away the spectacle and Build 2024 was, for a Microsoft-stack shop like ThreeBIT, a year of foundations: a supported cloud-native framework, a credible on-device AI story, and a very public reminder that shipping fast is not the same as shipping responsibly. We took all three lessons home.
Sources & further reading
- The Official Microsoft Blog — What's next: Microsoft Build continues the evolution and expansion of AI tools for developers (21 May 2024): https://blogs.microsoft.com/blog/2024/05/21/whats-next-microsoft-build-continues-the-evolution-and-expansion-of-ai-tools-for-developers/
- Microsoft — Build 2024 Book of News: https://news.microsoft.com/build-2024-book-of-news/
- .NET Blog — General Availability of Aspire: Simplifying .NET Cloud-Native Development: https://devblogs.microsoft.com/dotnet/dotnet-aspire-general-availability/
- VentureBeat — Microsoft introduces Phi-Silica, a 3.3B parameter model made for Copilot+ PC NPUs: https://venturebeat.com/ai/microsoft-introduces-phi-silica-a-3-3b-parameter-model-made-for-copilot-pc-npus
- DoublePulsar (Kevin Beaumont) — Microsoft Recall on Copilot+ PC: testing the security and privacy implications: https://doublepulsar.com/microsoft-recall-on-copilot-pc-testing-the-security-and-privacy-implications-ddb296093b6c
- Windows Experience Blog — Update on the Recall preview feature for Copilot+ PCs (7 June 2024): https://blogs.windows.com/windowsexperience/2024/06/07/update-on-the-recall-preview-feature-for-copilot-pcs/
- The Record — Microsoft reverses course, makes Recall feature opt-in only after security backlash: https://therecord.media/microsoft-reverses-course-recall-opt-in
Where figures vary between sources (attendance, exact announcement counts, TOPS thresholds), we have used Microsoft's own published numbers and noted ranges rather than picking a single precise figure.
Image credits
All photos are used under their respective Creative Commons licences; we are grateful to the photographers.