The Shift from AI Models to the Full AI Stack
This week made one thing very clear to me: AI is no longer just about models.
For the past two years, I have watched the conversation get dominated by capability — which model is smarter, faster, cheaper. That still matters, but it is no longer the centre of gravity.
What I am seeing now is a shift across the entire stack: from chips, to models, to interfaces, to market dynamics. And, importantly, all of these layers are starting to move at the same time.
That creates a different kind of momentum — and a different set of risks. Let me walk you through the signals that stood out to me this week.
In this issue:
- Microsoft Releases New Multimodal Foundation Models
- Anthropic’s Most Powerful Model Is Being Held Back
- AI Is Starting to Design the Chips That Power AI
- AI-Native Devices Are Emerging as the Next Platform
- AI Adoption Is Rising — But Trust Is Falling
- AI Startup Valuations Are Heating Up Again
Weekly AI Signals Summary: Chip Design, Model Releases, and Device Launches
- Microsoft launched new multimodal foundation models.
- Anthropic confirmed a powerful new model but is not releasing it yet.
- A startup raised $60M to use AI for chip design.
- Companies are preparing AI-native devices like smart glasses and earbuds.
- A new poll shows rising AI adoption but declining trust.
- AI startup valuations continue to surge at early stages.
Model Releases and Safety Strategy
1. Microsoft releases new multimodal foundation models
Announcing 3 new world class MAI models, available in Foundry — Microsoft AI, 2 April 2026
Microsoft releases new AI models to expand beyond OpenAI — GeekWire
On 2 April, Microsoft put three new in-house models into public preview through Microsoft Foundry:
- MAI-Transcribe-1 — speech-to-text across 25 languages, at roughly 50% lower GPU cost than the leading alternatives, and ranking first by FLEURS score in 11 of those languages.
- MAI-Voice-1 — speech generation that can produce 60 seconds of expressive audio in under a second on a single GPU. It is already powering Copilot, Bing, PowerPoint, and Azure Speech.
- MAI-Image-2 — Microsoft’s strongest text-to-image model yet, debuting at #3 on the Arena.ai image leaderboard.
The three MAI models are not research demos. They are shipping into products I already use, which tells me Microsoft is not just experimenting with in-house models — it is quietly replacing third-party dependencies with them.
Takeaway: Major platforms are building their own multimodal model stacks.
Why this matters to you
In my view, choosing a model increasingly means choosing a platform. As vendors bake models directly into their own ecosystems, switching costs and architectural lock-in become harder to ignore — worth weighing before you build a workflow that assumes one vendor’s stack forever.
2. Anthropic’s most powerful model is being held back
Project Glasswing: Securing critical software for the AI era — Anthropic
Why Anthropic is refusing to release its most powerful AI model — Times of India
On 7 April, Anthropic publicly disclosed Mythos, calling it a “step change” in capability. It is not making Mythos generally available: the model is good enough at finding and exploiting software vulnerabilities that Anthropic judged broad release too risky, and instead launched Project Glasswing, giving vetted partners access specifically to find and fix flaws in their own systems.
I find that reasoning more convincing than most safety statements I read, because it names a concrete, testable capability — vulnerability discovery — rather than a vague “it’s very powerful.”
Anthropic’s decision marks a shift in how frontier models are handled:
- Capability alone is no longer sufficient for release
- Deployment is gated by risk assessment and controlled rollout
Takeaway: The most important model event this week was a non-release.
Why this matters to you
The Mythos non-release tells me the best models may not be immediately available to everyone, and access may be staged, restricted, or delayed rather than a straight line from announcement to API key. If you build on frontier models, plan for uneven access to capability, not just steady improvement.
Infrastructure and Industry Shift
3. AI is starting to design the chips that power AI
Cognichip raised a $60 million Series A, led by Seligman Ventures with Intel CEO Lip-Bu Tan joining the board, to build “ACI” — Artificial Chip Intelligence — a physics-informed foundation model that fuses logic and physics-based reasoning to design semiconductors. The company claims it can cut chip development cost by more than 75% and halve the timeline, and says it is already engaged with 30+ semiconductor companies. Worth noting: it cannot yet point to a finished chip built with the system, and it has not named any of those customers.
Chip design remains one of the slowest and most complex parts of the AI pipeline. Automating it could unlock significant acceleration across the entire stack — if the claims survive contact with an actual production chip.
Takeaway: AI is now being applied to its own bottlenecks.
Why this matters to you
I think this creates a genuinely recursive loop: better AI leads to better chips, which leads to better AI. Progress is no longer limited to scaling compute alone — it is increasingly driven by improving the infrastructure underneath it. I would still treat the cost and timeline numbers as vendor claims until a named customer ships something built on them.
Interface Shift
4. AI-native devices are emerging as the next platform
Nothing’s AI devices plan reportedly contains smart glasses and earbuds — TechCrunch, 1 April 2026
Nothing is preparing a new generation of AI-first hardware: AI-focused earbuds later this year, followed by smart glasses in the first half of 2027. Notably, CEO Carl Pei had previously resisted smart glasses; he has since told staff he wants a multi-device strategy extending Nothing’s OS beyond phones into wearables. The glasses will reportedly pair a camera, microphones, and speakers with a phone and the cloud to handle AI queries.
These devices are designed for continuous, ambient interaction rather than discrete app usage.
Takeaway: AI is moving from screens into the physical world.
Why this matters to you
I read this as the next interface shift: desktop, then mobile, then ambient AI. The most important AI experiences may soon happen without a screen at all — though I would want to actually wear the glasses before believing the “ambient” pitch over the “yet another gadget to charge” reality.
Adoption and Market Reality
5. AI adoption is rising — but trust is falling
As more Americans adopt AI tools, fewer say they can trust the results — TechCrunch, 30 March 2026
A new poll of nearly 1,400 Americans shows a growing disconnect: only 27% now say they have never used an AI tool, down from 33% a year earlier, and 64% report using AI in work or personal life in the past month. Trust has not kept pace — 76% say they trust AI outputs rarely or only sometimes, against just 21% who trust it most or almost all of the time.
Takeaway: Adoption is outpacing confidence.
Why this matters to you
In my experience this shifts the product challenge from capability to reliability and trust. Verification, explainability, and consistency are becoming essential features, not nice-to-haves you bolt on once the model works.
6. AI startup valuations are heating up again
AI-focused companies are now closing seed rounds at a median pre-money valuation roughly 42% higher than a comparable non-AI startup — a $10 million seed at a $40–45 million post-money valuation is, as one investor put it, “pretty typical” if you are an AI company. Seed deal count is actually down, but the deals that do close are priced higher, with large VC firms moving in earlier to secure a stake.
Investors are pricing companies based on future potential rather than current traction.
Takeaway: Capital is accelerating ahead of outcomes.
Why this matters to you
Higher seed valuations create a high-pressure environment: faster funding, higher expectations, and less room for slow iteration. If you are raising, that premium is real money on the table — but it is also a bar you now have to justify.
Structural Shift Across the AI Stack: Hardware, Models, Interfaces, and Market
This week’s signals point to a structural shift:
AI is evolving across the full stack — with new constraints
| Layer | What is changing |
|---|---|
| Hardware | AI designing chips |
| Models | In-house models + controlled releases |
| Interfaces | Wearables and ambient devices |
| Products | Embedded AI experiences |
| Market | Rising valuations + falling trust |
The Full-Stack Shift: Why Every AI Layer Is Moving at Once
The most important shift this week is not a single announcement. It is the realisation that AI is no longer a single layer.
AI is a stack — and every layer is evolving at once. The full-stack AI shift represents a structural change in which hardware, models, interfaces, and market dynamics move together instead of one layer leading in isolation.
That creates powerful momentum. But it also creates coupling: hardware affects models, models affect interfaces, interfaces affect trust, and trust affects adoption.
Understanding AI now means understanding how these layers interact, not just how any one model performs. Increasingly, the teams that win will be the ones who can navigate the entire stack — I include myself in still figuring that out.
Did you find this useful? I would love to hear your thoughts. Let me know if you have comments or suggestions!
References
- Announcing 3 new world class MAI models, available in Foundry — Microsoft AI
- Microsoft releases new AI models to expand beyond OpenAI — GeekWire
- Project Glasswing: Securing critical software for the AI era — Anthropic
- Why Anthropic is refusing to release its most powerful AI model — Times of India
- Cognichip wants AI to design the chips that power AI, and just raised $60M to try — TechCrunch
- Nothing’s AI devices plan reportedly contains smart glasses and earbuds — TechCrunch
- As more Americans adopt AI tools, fewer say they can trust the results — TechCrunch
- It’s not your imagination: AI seed startups are commanding higher valuations — TechCrunch
Enjoyed this? Get more like it.
Weekly notes on AI tools, Python, and what I'm actually building — plus two free gifts: the 15-page Fantastic AI: The 2026 Toolkit and a Git Commands & Contribution Workflow Cheatsheet.
