Siri Finally Lets Every App In
Apple just stopped deciding which jobs your app is allowed to do inside Siri.
- A Developer Preview in the iOS 27.3 betas lets any app expose its own custom actions to Siri AI through App Intents, replacing Apple’s fixed categories. New “skills” let developers chain several actions into one spoken request.
- Users can skip naming your app when it’s already on screen or part of the conversation. That makes Siri a real discovery surface. It also comes with a ceiling: server-side Siri AI now has daily usage limits, and Apple says expanded access will cost a fee it hasn’t priced yet.
- App Store submissions open in spring, with the first integrations reaching customers in a few months. Terms and usage policies are still unpublished. Siri AI stays off in the EU and for anyone under 13.
Apple built the on-ramp and bolted a toll booth to it.
Pick the one action your users repeat most and wire it up as an App Intent in the 27.3 beta, so you’re testing before the spring submission rush.
SpaceX Is Snapping Up 418 Homes
SpaceX found a cheaper fix for turbine noise than turning off the turbines.
- MZX, a SpaceX subsidiary, wants to buy 418 homes in Southaven’s Colonial Hills, right by the power station feeding its Colossus II AI complex. To qualify, homeowners have to withdraw from a class action over gas turbine noise.
- Plaintiffs describe a nonstop roar like a jet engine that wakes them at 2am. The city says its monitoring has never crossed the 70-decibel legal limit. Some specialists warn that long exposure near 50 decibels can still harm health. Buying out the people suing turns a legal fight into a real estate deal.
- SpaceX says it will pull the unpermitted turbines within a year, once a 1.2-gigawatt power station comes online. The mayor says the empty houses could go to incoming employees.
What does a gigawatt of compute cost a neighborhood?
If you build on hosted AI compute, find out where your provider’s power comes from, because local backlash is now a supply-chain risk.
“Ram Altman” Walks Back the Memory Grab
Last week your RAM was the problem. This week Sam Altman wants you drowning in it.
- A week ago, Altman called it “quite terrible” that so much of the world’s scarce DRAM sits in personal computers, a comment aimed at people running AI models locally. Now, after renaming himself “Ram Altman” on X, he told a user who asked why he wanted their memory that he wants them to have plenty, and the industry should build more.
- The reversal landed hours after midterm results handed Democrats control of Congress, with many candidates campaigning on pausing data center construction. X attached a community note to his post pointing back to the earlier quote.
- Moratorium proposals that stalled under the old Congress now have a realistic path. AI firms that leaned on executive orders to fast-track buildouts lose that shortcut.
Half of US adults now say data centers are bad for their home energy bills.
Price out a local-model fallback for your product, because data center capacity just became a political variable.
Anthropic Is Creating a System to Verify the Identity of AI Models
Anthropic is developing an open-source system called FlashProve that aims to cryptographically verify which AI model generated a particular output.
- FlashProve uses what researchers call “provable inference”: a cryptographic proof that a specific output was generated by a particular AI model.
- The goal is to let users confirm that an answer actually came from the model a company or developer claims to be using, rather than a different model substituted behind the scenes.
- It could also help verify third-party claims, such as whether a software library was scanned by a frontier cybersecurity model or a business plan was evaluated by a specific advanced model.
- For the system to work, AI labs would publish a unique fingerprint for each model. These fingerprints could then be matched against outputs and verified cryptographically.
- FlashProve is designed to be lightweight, allowing verification to take place locally on a user’s device.
The project is still in the research stage. Anthropic says there are currently no plans to deploy FlashProve internally or at other AI labs, and is sharing the work so researchers and developers can examine and improve it.
Apple Opens Siri AI to Third-Party Apps
Apple is expanding Siri AI’s capabilities beyond its own ecosystem, allowing third-party developers to integrate the assistant into their apps.
- As part of the iOS 27.3 update, Apple announced that developers can now integrate Siri AI’s new features in “all apps.”
- This lets the voice assistant draw on a user’s personal context, understand natural language prompts and use on-screen awareness to complete requests.
- The update also introduces “app skills,” a new feature that lets developers bundle several tasks into a single action and teach Siri how to coordinate multi-step requests.
- Users won’t always need to name the app in their requests. Siri can understand which app they mean when it is open in the foreground, running a Live Activity or already part of the conversation.
Apple is also extending Siri AI’s expanded capabilities to its own apps, including Calendar, Files, Finder, Journal, Mail, Maps, Messages, Music, Notes, Photos, Preview, Reminders, Safari, Shortcuts and Weather.
How to Use Netflix’s New AI Discover Feature
Netflix describes this first version of AI Discover as “built around themes,” letting subscribers start with a mood or an idea and narrow down their choices from there.
Step 1: If you have access to the feature, you should see a red gem with “Discover” in the top menu next to My Netflix.
Step 2: Browse the available themes, each of which features a clip from a show or movie.
Themes are grouped in a few ways:
- Choose based on how you want to feel, such as “Something Light” or “Scare Me.”
- Choose based on who you’re watching with, such as “Family Movie Night.”
- Choose based on how much time you have, such as “Quick Watches.”
- Other, more general themes include what’s trending or what you’ve been watching lately.
Step 3: Once you select a theme, you will see a collection of matching shows and movies. (You can choose “Something Different” for a new set of themes.)
Step 4: Every title has a “Tune Your Recommendations” button alongside “Play” and “My List.” Use it to tell Netflix to keep a recommendation or to remove the selected show or movie from your recommendations.
Note: AI Discover is gradually rolling out in the US to members who use Netflix in English on supported TVs, so you might not have access yet. It will open up to more people over the coming weeks.
Google will turn off Search for schools that don’t want AI
Chris Phillips, Google’s VP for education, told The New York Times that “Search and AI are deeply integrated,” meaning schools that opt out of AI “will not have access to search, and they’ll see a message from their school that indicates it’s not available.”
Apple requires dating apps and AI companions to be rated 18+
The company is adjusting its App Store policies to make it more difficult for minors to use online dating apps or AI companions that engage in “sexual or romantic content.” The new policy will require those apps to carry an 18+ rating, and starting November 16, developers must also add content descriptors reflecting these features.
Sam Altman responds to controversy over RAM remarks
OpenAI CEO Sam Altman has clarified his comments about consumer DRAM usage, stating he wants people to have “as much RAM as they want in their computers” and that global supply simply needs to expand.
His response follows a recent interview where he criticized the amount of idle RAM in personal computers, triggering heavy pushback from users online.
Hardly anyone is using AI agents for their DoorDash orders
DoorDash CEO Tony Xu said the company has received only “several hundred” orders from third-party AI agents, despite completing an average of 3 million deliveries per day.
OpenAI to acquire software startup Pierre Computer Company
The company says the acquisition will help Codex “connect the code a company ships with the decisions its people make.”
Apple plans $1,000 home robot for 2027
Apple is reportedly building its first home robot, code-named J595 and possibly called HomeView Ultra, aiming for a 2027 launch at around $1,000, though delays could push it to 2028.
The tabletop device puts a 9-inch screen on a swiveling arm that tracks you during FaceTime calls. New CEO John Ternus is pushing robotics, but AI software is holding up mostly finished hardware.
The project began as a fancy video-calling gadget around 2019. Seven years later, engineers still hunt for features beyond FaceTime and light switching that justify a $1,000 price.
Canva cuts 1,271 jobs as growth slows
Canva is laying off 1,271 employees, or 22% of its workforce, after growth slowed for the first time in a decade, marking a major shakeup for the design software giant.
CEO Melanie Perkins says Canva will merge its product, growth, and platform teams, give employees broader responsibilities, and introduce hybrid work requirements as the company looks to streamline operations.
Rising AI costs already pushed Canva to slash its growth forecast from 30% to 20%. Now, with ChatGPT, Gemini, and AI design startups gaining ground, Canva faces pressure from both sides.
CrowdStrike launches an AI Kill Switch
CrowdStrike has unveiled an AI Kill Switch for its Falcon Guardian security platform, allowing businesses to immediately cut off AI agents that behave unexpectedly or pose security risks.
The switch disconnects agents from their AI models without shutting down the models themselves. Human administrators control activation, and the system logs every use for security monitoring and accountability.
CrowdStrike’s CEO compares the feature to a race car’s brake, which agents can’t override. As businesses give AI agents more access, having an emergency stop button looks increasingly practical.
Netflix launches AI Discover recommendations
Netflix is rolling out an experimental AI Discover tab for English-language users on supported U.S. TVs, letting viewers find shows and movies based on their mood, available time, or viewing preferences.
Users can browse themes like “something light,” “scare me,” and “quick watches,” refine results, and use a new “Tune Your Recommendations” button to save preferred suggestions or exclude unwanted titles.
Netflix calls AI Discover its “test kitchen” and will use feedback to shape future features. If it works, choosing a movie might finally take less time than watching one.
Apple opens Siri AI to all app developers
Apple released a Siri AI Developer Preview in its new 27.3 betas, letting any app plug its own actions into the rebuilt assistant that launched in beta with iOS 27 in September.
Until now, Siri AI could only use third-party apps through Apple’s preset templates for common tasks. Developers can now build custom actions and “skills” that teach Siri to chain several steps together.
The walled garden cracks open, slowly. App Store submissions don’t open until spring, and Apple hasn’t given a date, so no third-party Siri AI integration will reach the App Store this holiday season.
GPT-6 Astra edits model weights directly, boosting math 6x without training
MIT CSAIL just showed something genuinely weird: GPT-6 Astra edited another AI model’s brain directly, no training required.
Think of it like this. Every AI model is made of billions of numbers called weights. Those numbers control how the model thinks. Normally, changing them means expensive retraining. Astra just… read them, found the bug, and patched it like editing a config file.
The target was Qwen3.5-0.8B, a small model stuck in reasoning loops. Astra touched only 0.016% of its weights and the results were:
- MATH-500 score jumped from 7.9% to 48.1%
- GSM8K math score went from 13.8% to 67.1%
- Instruction-following score tripled, from 21% to 62.3%
They call this Weight-Space Programming: using a big AI to fix a smaller AI by editing its weights directly. No GPU training runs, no datasets, no cost.
The fixed model is already on HuggingFace. MIT is now building tools so you can try this yourself.
Hydrogen-1.0 open-weight science model beats GPT-6 on chemistry benchmarks
Most science AI is locked behind closed APIs. You can’t inspect it, adapt it, or trust it fully. Hydrogen-1.0 flips that. Deep Principle just dropped a fully open, 753B parameter model built specifically for scientific research, and you can grab the weights right now on Hugging Face under MIT license.
It was trained on top of GLM-5.3, then fine-tuned on real lab records, full experiment cycles, and past discovery paths. That training approach is what makes it special: it learned not just facts, but how scientists actually think and work.
Here’s what it can do:
- Plan experiments from a short description of intent
- Run and debug molecular simulations end-to-end
- Catch lab procedure errors before they happen in a virtual rehearsal
- Find better material candidates with fewer attempts
Specs worth knowing: 1M token context, 128K output, scores 38.3% on FrontierScience-Research versus GPT-6-Astra’s 33.3%. You can self-host it with SGLang on 8 GPUs, or hit it via API at GLM-5.3 pricing.
Open-source tool unlocks Nvidia GeForce GPUs, boosting AI performance up to 87%
Your consumer GPU has been running at half speed. On purpose. NVIDIA artificially limits Tensor Core performance (the chips that handle AI math) on GeForce cards to push you toward pricier professional GPUs. A new open-source tool called gpu-unlock removes that limit.
Here is what actually happens: when your GPU does the AI math most training and inference relies on, GeForce cards issue those instructions at half the rate of professional cards. gpu-unlock rewrites two hardware registers before NVIDIA’s driver loads, then hands control back. No VBIOS flash, no patched driver.
Real-world gains reported:
- RTX 3090: up to +65% compute throughput in AI workloads
- RTX 4090: FP16 rate jumps from 165 to 331 TFLOPS
- Image generation and model fine-tuning: 25-55% faster
It is Linux-only, still experimental, and requires a cold boot each time. The warning is real: running tensor cores at full speed draws significantly more power. Check the repo at github.com/Zagitta/gpu-unlock.