The new Copilot 2025 Fall Update features also up the ante in terms of capabilities and the accessibility of generative AI assistance from Microsoft to users, so businesses relying on Microsoft products, and those who seek to offer complimentary or competing products, would do well to review them.

Read the full article here.

Source: VentureBeat.com

The post Microsoft Copilot gets 12 big updates for fall appeared first on Useful Tech Stuff.

Every boom likes to pretend it is different. The AI surge insists it is inevitable progress wrapped in exponential curves, not a bubble. Yet the jargon, the easy capital, and the copycat product roadmaps look very familiar. When this bubble bursts, the fallout will not be a single pop. It will be a long hiss. Here is what that will look like, why some parts of the ecosystem will be just fine, and where real value will quietly consolidate.

Funding will snap from story to substance

Capital always chases narrative until it chases revenue. We will see a sharp repricing of private companies that raised at peak multiples on the strength of slideware benchmarks. Bridge rounds will come with tougher terms and liquidation preferences that sting. Survival will depend on proof of durable unit economics, repeat customers, low churn, and gross margins that do not collapse under GPU bills. Startups that rely on perpetual subsidy to keep usage afloat will not make it. Those with minimal customer issues and disciplined cost control will.

The compute bill will come due

Cloud credits are not a business model. As subsidies fade, many teams will learn that their cost of goods sold is effectively the price of someone else’s datacenter. That pushes companies to either raise prices or invent cheaper ways to deliver similar outcomes. Expect a scramble toward smaller, task specific models, on device inference where latency and privacy matter, and ruthless pruning of nonessential features. The winners will treat compute as a precious resource, not a blank canvas.

Horizontal tools will give way to vertical stacks

Generic chatbots and “AI for everything” will feel like the social media dashboards of a past era. Useful at first, then undifferentiated. Value will migrate into vertical stacks that own workflow, data, and integration; radiology reporting that fits the hospital information system, claims processing that speaks the insurer’s rules, developer tools that plug into build pipelines and ticketing. These stacks will look boring. They will also capture the profit that general chat apps cannot defend.

Data moats will trump model bloat

Bigger models get headlines. Better data gets contracts. When budgets tighten, enterprises will ask how models were trained, how outputs can be traced, and how liability is handled. Vendors who can secure, clean, and govern domain specific data will beat vendors who only promise larger checkpoints. Expect more partnerships where customers keep their data inside their own cloud boundary, with models that visit the data rather than the reverse. Trust will be as important as accuracy.

Compliance will graduate from cost center to moat

Regulation always arrives late, then all at once. Once it does, compliance maturity becomes a competitive edge. Auditable pipelines, red teaming practices, provenance of training data, and incident response will matter in sales cycles. Vendors who baked in governance will pass due diligence without drama. Others will bolt on policies and bleed time. The bubble’s end will reward the dull habits of good engineering.

The talent market will rebalance

Right now, salaries reflect scarcity and hype. Some researchers will move from frontier work to applied roles. Many product teams will shrink. On the flip side, managers will hire for outcomes rather than buzzword familiarity. Engineers who can ship reliable systems and measure value will have the advantage. So will technical writers, designers, and support teams who can make complex capabilities usable and safe.

Consumers will keep what feels magical and drop the rest

A correction does not erase demand. It clarifies it. People will pay for tools that save time in ways they can feel; better search inside their own files, assistants that actually do the multi step tasks they promise, creative tools that produce results rather than add novelty for its own sake. Subscriptions for “AI sprinkled here” will disappear. Subscriptions for “this turns three hours into thirty minutes” will stay.

Open source will harden

The open ecosystem will look messy during the bust. Projects will be abandoned. Repos will go quiet. Yet the remaining set will mature and stabilize around pragmatic goals; performance on commodity hardware, privacy by default, and simple deployment. Community efforts that provide clear licensing, reproducible builds, and strong documentation will thrive.

The big platforms will consolidate quietly

Expect more acquisitions, more bundling of model credits into existing enterprise contracts, and more integrated stacks that make it easier to choose the house solution. For many buyers the safest choice in a downturn is the vendor they already pay. That is not always the most innovative option, but it is the one that clears procurement.

What to build in the chill that follows

1. Pick pain, not puzzles. Start with a costly bottleneck and design backward from a measurable win. Benchmarks are not the product. Throughput, error rate, and cycle time are.
2. Design for cost gravity. Assume compute will stay expensive relative to price sensitivity. Cache, distill, prune, and run smaller where possible.
3. Own the last mile. The moat is often workflow integration, change management, and trust. Get excellent at onboarding, metrics, and support.
4. Invest in data health. Provenance, labeling quality, deduplication, and feedback loops pay compounding dividends. Treat them as product features.
5. Practice resilience. Feature flags for rapid rollback, and clear incident playbooks will save you should things go awry.

Conclusion

Every technology wave overshoots. Then it right-sizes and becomes infrastructure. The internet did. Mobile did. Cloud did. AI will too. The end of the bubble will not be the end of value. It will be the end of delusion. What remains will be tools that make people better at their work, services that reduce complexity, and systems that act with reliability and care. That is not as loud as the boom phase. It is, however, more durable.

The best time to build sober companies is after the party. The music fades, the lights come on, and you can finally see who is still standing. Build for that room.

The post When the AI Bubble Bursts appeared first on Useful Tech Stuff.

Microsoft just rolled out another wave of Copilot upgrades for Windows 11, voice activation (“Hey Copilot”), deeper on-screen understanding via Copilot Vision, and new “agentic” abilities that can take actions on your behalf. The company’s own headline says it’s “making every Windows 11 PC an AI PC.”

First, what is an “AI PC”?

In Microsoft’s strictest sense, the Copilot+ PC label requires dedicated on-device AI horsepower, an NPU capable of ≥40 TOPS, alongside baseline RAM and storage, so features like Recall, Cocreator, and Auto Super Resolution can run locally, fast, and (in theory) more privately.

What changed with the latest Copilot update?

Three things stand out:

1. Ubiquity. Copilot is now more present and easier to invoke (voice, taskbar integration preview, and broader “Vision” availability), with Microsoft emphasizing AI woven through everyday Windows experiences, it is not just a sidebar chatbot.
2. Actions, not just answers. “Copilot Actions” can perform tasks. For example, making reservations or manipulating files, within permission boundaries you approve.
3. Not only for Copilot+ PCs. Much of this update runs on any Windows 11 machine, which is why Microsoft says “every PC” is becoming an AI PC. Independent coverage underscores that most of the new Copilot capabilities don’t require an NPU.

So, are we there yet?

Well, that very much depends on one’s definition of AI PC.

• If “AI PC” means AI features everywhere, discoverable in the shell and usable by mainstream users: Microsoft’s new update is a real step. Copilot is more proactive, more contextual, and increasingly able to do things, not just chat.
• If “AI PC” means on-device intelligence by default; low latency, private, battery friendly workflows running on your laptop’s NPU, then the Copilot+ definition still matters. Many headline features remain tied to NPU class hardware, and Microsoft’s own requirements haven’t changed.

The catch: fragmentation and fit

• Feature fragmentation. Some experiences (e.g., Recall. You can watch my video on that here) have been controversial and are still gated, optional, or hardware dependent. Organizations and even universities have urged caution; Microsoft now exposes Recall as an optional Windows feature you can enable or remove. Expect IT to curate which “AI” shows up on endpoints.
• Rollout reality. Several integrations (like taskbar level Copilot chat) are starting in Insider previews and may take time to reach everyone.
• Windows versioning. Big AI leaps have aligned with the Windows 11 24H2 platform and later; enterprise rollout pacing and known issues tracking still apply.

What it means for you

• On older PCs: You’ll see more Copilot entry points and useful assistance without buying new hardware, which is good for everyday guidance, light automation, and quick explanations.
• On Copilot+ PCs: You get the fuller vision; local AI effects, creation tools, and (where enabled) time machine style retrieval, thanks to that ≥40 TOPS NPU. This is where the AI-PC promise feels native and not bolted on.

Conclusion

Microsoft’s latest update makes Windows feel much closer to the AI-first experience it has been pitching: Copilot is more capable, more present, and beginning to act on your behalf. But the truest version of an “AI PC” still hinges on on-device NPUs and the Copilot+ class of hardware. In short: AI for everyone now; fully native AI when you buy the right silicon.

Sources: Reuters, Microsoft, Microsoft Learn, Microsoft Support, Windows Blog, Windows Central

The post Is Microsoft Finally Close to an “AI PC”? appeared first on Useful Tech Stuff.

By now we’ve all seen the glossy demos: humanoid robots smiling, making eye contact, mirroring gestures, chatting with a warmth that feels… almost real. The tech is impressive and it can also be off-putting. As robots become more life-like, many people don’t lean in; they pull back. That hesitation isn’t irrational. It’s a healthy reaction to a bundle of social, ethical, and safety questions that hyper-realistic machines force into the open.

The uncanny isn’t the only valley

We tend to blame discomfort on the “uncanny valley,” the eerie dip in affinity when a robot is nearly, but not quite human. But lifelike robots create other valleys too:

• The trust valley: When a robot sounds empathetic, we naturally infer understanding and intent. If the underlying system is brittle or narrow, the gap between performance and competence breeds mistrust. People become cautious not because the robot looks odd, but because it feels misleading.
• The consent valley: Social cues evolved to signal agency; eyes, nods, tone. When a robot deploys those cues, people may unconsciously grant it permissions they would never give a kiosk: proximity, personal disclosures, influence over choices. That’s a consent problem in disguise.
• The accountability valley: If a human-like agent errs, who is responsible, the manufacturer, the model provider, the owner, the operator? Lifelike design can mask the chain of accountability, dulling our instinct to ask “Who’s on the hook?”

Why realism raises the stakes

1. Manipulation gets easier. A comforting face and warm voice can nudge behavior in ways text boxes cannot. In retail, education, or care settings, a realistic robot can move people from “What’s the policy?” to “I’ll just do what my friendly helper suggests,” even when suggestions aren’t in their best interest.
2. Privacy expectations shift. We tolerate microphones in smart speakers; we bristle when a “person” stands nearby and listens. A face and gaze imply relational norms, yet the robot may be streaming data to a cloud service and third parties. Life-like presence increases the feeling of intimacy while often lowering actual privacy.
3. Role confusion in vulnerable contexts. In elder care, pediatrics, or mental health scenarios, a human-like helper can be beneficial, but also risky. Users might disclose sensitive information or rely on the robot for emotional support beyond its design. If the system fails or is withdrawn, that can feel like abandonment.
4. Labor and dignity optics. A humanoid serving coffee can spark a defensive reaction not just about jobs but about the human meaning of service work. The more a robot mirrors us, the more it invites comparisons and resistance to how we value people.

“Helpful” isn’t the same as “harmless”

The industry’s default defense is utility: if robots tutor kids better, help nurses lift patients, or fix potholes, why fuss about appearances? Because in social systems, form is a feature. A realistic face is not a skin; it’s an interface contract. It sets expectations about comprehension, empathy, and duty of care. When those expectations are unmet, people feel misled, and they retreat.

Pragmatic guardrails (designers and policymakers, this is your move)

• Make the machine-ness legible. Require clear, persistent identity cues; a tonal watermark, subtle visual markers, or on-device badges that signal: “I’m synthetic.” Not a one-time disclosure in a manual, but an ambient reminder during every interaction.
• Separate warmth from deceit. You can be friendly without impersonating humanity. Favor stylized, plainly robotic aesthetics for general public deployments. Reserve highly human-like embodiments for use cases where they’re clinically or functionally justified and only with ethics review.
• Expose capabilities and limits up front. Robots should announce what they can and cannot do, what data they collect, where it goes, and how to opt out. Use plain language, not legalese. If the model can hallucinate, the robot should say so before making recommendations.
• Design for handoff and override. Prominent physical controls (mute, pause, SOS to a human) should be visible and tactile, not hidden in an app. A life-like agent that can’t be easily interrupted feels domineering, not helpful.
• Contextual consent, not blanket consent. The robot in a hospital hallway should have different data defaults than the same model in a private home. Minimize collection by default; escalate only with explicit, time-boxed opt-ins.
• Audit the “social proof” layer. Test not only accuracy and safety but persuasion effects: Does the robot’s demeanor produce undue compliance? Are certain groups more susceptible? Treat this like accessibility testing – mandatory, not optional.
• Label the supply chain. A transparency tag (on-device and in settings) should name the hardware maker, model provider, and operator. If something goes wrong, users shouldn’t need a FOIA request to find the responsible party.

A better north star: appropriateness, not imitation

We don’t need robots that pass as people. We need robots that are appropriate to their context, legible, interruptible, respectful, and honest about their nature. In plenty of spaces (industrial floors, logistics, public transit), a clearly non-human form reduces confusion and friction. In others (therapeutic settings with evidence of improved outcomes), human-like elements can help, but only with safeguards and sunset plans.

The public’s caution is wisdom

If life-like robots make people hesitate, that’s not a barrier to adoption; it’s a compass. It points us toward designs that earn trust rather than borrow it from human mimicry. The goal isn’t to abolish faces and voices, it’s to ensure those features serve users, not engagement metrics. When we treat realism as a regulated capability, deployed sparingly, audited rigorously, and chosen for necessity rather than novelty, people won’t need to be deterred. They’ll be discerning. And that’s the kind of user a healthy robotic future depends on.

What are your thoughts on this subject? Let me know in the comments.

The post When Robots Look Too Human appeared first on Useful Tech Stuff.