- Text The Verge
Microsoft Launches Recall and AI-Powered Windows Search for Copilot Plus PCs
-
Recall officially launches as an opt-in memory-enhancing feature for Copilot Plus PCs
-
Recall continuously takes categorized screenshots of your activity to help you “pick up where you left off” using natural language search.
-
Microsoft spent 10 months redesigning Recall’s security, adding encryption, sensitive content filtering, and opt-in controls.
-
Despite improvements, privacy concerns remain: Recall’s filtering can be inconsistent, and access protection is still relatively weak (e.g., PIN via Windows Hello).
-
-
AI-enhanced Windows Search enables natural language queries across your PC
-
Users can now search by describing content, like “brown dog photo,” instead of relying on exact filenames or metadata.
-
This functionality extends across File Explorer, the Windows search bar, and settings.
-
-
Click to Do debuts as a Windows-native alternative to Google’s Circle to Search
-
Triggered via Windows key + left click, this feature lets users act on-screen content — e.g., summarizing text or removing image elements.
-
Initially exclusive to Qualcomm-powered Copilot Plus PCs, with Intel and AMD support coming soon.
-
-
Availability and regional limitations
-
All three features — Recall, enhanced Search, and Click to Do — launch today on Copilot Plus PCs.
-
Not yet available in the EU, Iceland, Liechtenstein, or Norway; Microsoft plans to roll out there later in 2025.
-
-- Rebecca Bellan, TechCrunch, 4/23/25
- Text NY Times
Apple and Meta Are First to Be Hit by E.U. Digital Competition Law
-
Apple and Meta fined under the Digital Markets Act (DMA) for anti-competitive practices
-
Apple was fined €500 million ($570 million) for restricting how app developers communicate with users about alternative sales and offers outside the App Store.
-
Meta was fined €200 million ($230 million) for implementing a “consent or pay” model, which required users to either allow ad tracking or pay a fee for ad-free access to Facebook and Instagram.
-
-
DMA aims to curb the gatekeeping power of dominant tech platforms
-
The law, adopted by the EU in 2022, is intended to prevent companies like Apple, Meta, and others from abusing their control over digital platforms.
-
It prohibits practices that force business users or consumers into unfair terms, such as closed payment systems or forced data-sharing.
-
-
The fines escalate EU–U.S. tensions, prompting backlash from the Trump administration
-
U.S. officials labeled the ruling “economic extortion” and warned it may trigger trade retaliation.
-
A February White House memo explicitly said the U.S. would respond if EU laws unfairly target American tech firms.
-
-
Both Apple and Meta intend to appeal the fines and criticize EU policy as discriminatory
-
Meta claims the DMA acts like a “multibillion-dollar tariff” that forces it to offer reduced-quality services.
-
Apple argues the ruling forces it to hand over proprietary technology and undermines platform security.
-
Both companies frame the penalties as part of a broader anti-U.S. tech bias, especially as Chinese and European competitors are not held to the same standards.
-
- Text Anthropic
- Preface by the editor of this blog
-- Thinking models display their chain of thoughts, right? Not necessarily.
-- A couple of months ago, the editor posed a complex word problem that involved simple math to OpenAI's Deep Research agent that used OpenAI's o3 thinking model.
-- Among other things, the problem required the model to find the types of main courses served at upscale vegetarian restaurants and their prices.
-- At some point, the DeepResearch agent's chain of thought declared that it was looking at the online menu of a highly esteemed restaurant.
-- When it finished the problem, the agent displayed a few of that restaurant's main courses and prices.
-- I asked where it found the prices. It replied that it got the price’s from the online menu. However this restaurant never displays the prices of any courses on its online menus. The agent then conceded that it had, indeed, made up the prices. So its previous assertion was not a hallucination; it was a "fib".
-- While it was working on the problem, the model must have "thought" that it could not find the prices and it must have "thought" about its decision to make up prices for the main courses of this restaurant. But these "thoughts" were not included in the chain of thoughts that it displayed.
- True interpretability isn’t about the surface-level outputs
As Dario Amodei, the CEO of Anthropic, points out, true interpretability isn’t about the surface-level outputs. It’s about understanding the actual internal circuits — the “neurons” firing inside the model — that produce behavior. And right now, our ability to interpret those internal mechanisms is falling behind the rapid advances in model capabilities. - Why This Matters:
-- This is not another “feature release” story. It’s not Google launching a new app or Adobe upgrading image generation.
-- This is a fundamental shift in how we understand — and safeguard — powerful AI systems. If we don’t massively scale up interpretability research, we risk building smarter and smarter systems without understanding how they actually work.
- What Amodei is Proposing:
1. Massive Collective Effort: Anthropic and others have made real progress — starting with small models, working up to larger ones, mapping millions of circuits — but it’s not enough.
2. A Race Against Time: Models are improving faster than interpretability is. Without a coordinated, global effort, we won’t be able to keep up.
3. An Optimistic Call to Action: Amodei believes that if enough resources, talent, and attention are put into interpretability, we can solve this problem — but we have to start now.
No comments:
Post a Comment
Your comments will be greatly appreciated ... Or just click the "Like" button above the comments section if you enjoyed this blog note.