Recall is an upcoming AI feature of Windows 11. It takes captures of the screen every five seconds and saves them on the local system. Users may then interact with AI to look up information or process data that has been captured.
Most current Windows 11 PCs won’t get Recall functionality, as a Copilot+ PC is required. These PCs have specific requirements that include a neural processing unit and 16 GB of RAM.
The first iteration of Recall will only be available for specific Snapdragon ARM processors on top of that. With time, Recall will become available for AMD and Intel PCs as well.
Recall Criticism
Windows 11’s Recall feature is not without criticism. Here are the main points:
It records a user’s entire activity on the Windows PC, with a few exceptions. This makes it the holy grail for law enforcement, spammers, malicious actors, advertisers and marketers.
Recall snapshots require lots of storage. The default is 25 GB on a 256 GB hard drive. While you can drop that to 10 GB, it is still a lot of space. On 1 TB+ drives, the default is 150 GB of storage.
Disabling Recall in Windows 11
There are three options to disable Recall on PCs that support it.
Via the Settings app.
Via the Group Policy Editor.
Via the Registry (not yet available)
Settings app
Open the Start menu and select Settings.
Go to Privacy & security > Recall and snapshots.
Toggle Save snapshots to off.
This is all to it. Recall is disabled from that moment on for that user account.
Turn off Recall in the Group Policy Editor
Note: The Group Policy Editor is only available in Windows 11 Pro, Enterprise, and Education editions. Home systems may disable Recall in the Registry, which is explained in the next section.
Open the Start menu.
Type gpedit.msc and press the Enter-key to load the Group Policy Editor.
Use the hierarchy on the left to go to User Configuration > Administrative Templates > Windows Components > Windows AI.
Double-click on Turn off Saving Snapshots for Windows.
Change the statues from Not Configured to Disabled.
Restart the PC.
Disable Recall in the Registry
Information not yet available, will update once it becomes available.
A lot has been written about the upcoming Recall feature of Windows 11 version 24H2. Reserved exclusively for Copilot+ PCs, it will be unavailable to the majority of users who upgrade their Windows 11 PCs to the new version.
Recall, in a nutshell, takes frequent captures of the entire screen and stores them encrypted on the local system. Windows 11 users may then invoke the Recall feature to interact with the saved content.
From searching for specific information over getting summaries of watched videos or telecalls to finding that specific asparagus recipe that you looked at some time ago in Edge.
Search across time to find the content you need. Then, re-engage with it. With Recall, you have an explorable timeline of your PC’s past. Just describe how you remember it and Recall will retrieve the moment you saw it. Any photo, link, or message can be a fresh point to continue from. As you use your PC, Recall takes snapshots of your screen. Snapshots are taken every five seconds while content on the screen is different from the previous snapshot.
How Recall works
You use natural language to find something and Recall returns the information separated into text and visual matches.
Recall is shown as an icon on the Windows 11 taskbar and it may also be started using the keyboard shortcut Windows-J.
Recall displays a timeline on start that you may use to check out a specific day. Recall loads and displays the snapshots of that particular day then, allowing you to interact with the content.
Search is the heart of recall. You use natural language to find or interact with the saved content.
Type what you are looking for and Recall returns any matching snapshot. The AI feature displays hits from all apps by default, but you can filter results by a specific app to narrow down the results.
The feature distinguishes between close and related matches:
Close matches — includes at least one of the search terms or images that represent the search term.
Related matches — displays related items, e.g., cannelloni results when you searched for goat cheese pizza.
Selecting a screenshot launches the Screenray feature. Microsoft says that Screenray anayzes the snapshot and enables interactions with elements of it.
The company writes:
What you can do with each element changes based on what kind of content screenray detects. If you select a picture in the snapshot, you can copy, edit with your default .jpeg app such as Photos, or send it to another app like the Snipping Tool or Paint. When you highlight text with screenray, you can open it in a text editor or copy it.
Is Recall a privacy nightmare?
Recall records most activity on a Windows PC when it is active. It is up to the user to enable or disable Recall.
Microsoft has added options to disallow the capturing of specific apps or websites. Some of these are only available in Edge.
Recall does not capture private browsing sessions in Chromium-based browsers. In Edge, the feature may furthermore block captures of specific websites.
In other words, if you use a different browser than Edge, website filtering won’t work. If you use Firefox or another non-Chromium-based browser, everything will get recorded.
Recall runs locally only according to Microsoft. Captures are stored locally and the OCR feature runs local as well only.
The main privacy concerns
Recall runs locally only. The main concern that some users have is that someone else may gain access to the recorded data.
There are several scenarios where this may happen:
Malware infections may gain access to the data. This gives threat actors access to a user’s entire activity on the PC. It may include information about financial services they use, online accounts, password managers or security software, and confidential information in Word or Excel.
Law enforcement, including border agents, may want access to the information as it highlights (most of) the activity of a user on the Windows 11 device. Users may be coerced into giving state representatives access to their Windows PCs.
It is your choice
You may or may not use Recall. Most Windows 11 users cannot even use it, as their PCs do not meet the minimum system requirements.
If your PC supports it, you may want to ask yourself a simple question: do I really need it? Is it improving may workflows or helping me in another way?
It is a novelty feature, but how often will you make use of it once that novelty factor wanes off?
If you ask me, I won’t make use of it. All my PCs are not Copilot+ PCs and even if they were, I would turn it off as I do not need it. I know where to look when need to find something.
For businesses, it may play a bigger role. Making everything searchable, including video calls and presentations, is certainly useful in some scenarios.
What about you? Will you use the Recall feature when it comes out?
Microsoft, Google, and other search engine companies have started to add AI to the search results. Google is rolling out AI Overviews to search. This feature displays AI generated content at the top of the results.
This change has severe implications for the World Wide Web. In this article, I’m making four predictions about the future of the WWW.
Not all sites and services will be around in 5 years
AI generated results will keep a portion of searchers on the search engine’s website. The answer may suffice for a good percentage of user searches.
This impacts sites that rely on traffic from search engines. Google says that sites listed by its AI Overviews feature have received more clicks, but it is too early to tell whether only a tiny percentage of sites really benefits from it. In any event, sites on the second or third party of results will likely get less traffic.
The same may be true for certain services. Search engines like Google added more and more tools and features. You can check the weather, convert currency, get translations, do calculations or directions right from Google.
In any event, the drive towards showing more information directly on search results pages will push lots of sites out of business.
Good news is that this may level the search playing field. Companies like OpenAI may create their own search engines, which threatens the dominance of Google.
AI favors large sites and data sources
Generative AI relies on data created by humans. Without that data, it would not be able to produce any results.
For search engines, large data sources are of greater interest than smaller ones. Giving the AI access to the entirety of Wikipedia or Reddit is better than having to parse millions of smaller sites for the information.
Larger sites benefit from this, as they may sell their data to AI companies. Smaller websites do not have the means to broker deals with these companies, which means that they won’t get any compensation for their data. While they may opt-out, this also means that they won’t receive any links when the AI includes them as a source.
Ultimately, small sites are at the receiving end again when it comes to this, while large sites have a new revenue source at their disposal.
Trust will play a major role
Can AI results be trusted? AIs may hallucinate, which means that they make things up. There is also the question about the actuality of data. Since AIs rely on human generated content, they may return content that is out of date or also incorrect.
Say, you search for instructions on making a change to your Windows system. AI may return instructions that work on older versions of Windows. While you may adjust the query, there is a chance that out of date content is returned.
Trust will also play a major role in the survival of websites. Trusted sites will continue to do well, as many searchers will favor them over AI generated content. Sites that have a loyal followship may also survive.
Personality and authenticity remains important
Bland sites that just rehash news stories will have it difficult in the future. They do not really provide much value, but the business model worked, especially for established sites that can rank for pretty much any topic they write about.
Information returned by AI has no personality. It is just a robot returning information. Your favorite reviewer of video games, movies, software applications, or cars, on the other hand, may have more to offer than just the information.
This personality and authenticity of writers, podcasters, or video creators drives users towards services and sites, and may make them follow certain sites or services.
These will continue to do well, as AI cannot compete with that or mimic it satisfyingly.
Now You: what is your take on AI integrations in search, or entirely AI-powered search engines?
DeepL Write Pro is a subscription-based AI writing assistant by DeepL, maker of the popular translation service.
The target audience is organizations and professionals. DeepL Write has been available as a free service for some time. You can feed it text to get improvement suggestions.
DeepL Write Pro takes the concept a step further. While it continues to give suggestions, these are now provided in real-time. Subscribers may also set different writing styles and tones to create results for specific audiences.
DeepL Write Pro launches with business, casual, academic, and simple writing styles, and friendly, diplomatic, confident, and enthusiastic tones.
DeepL says that Write Pro is available on the official website, desktop and mobile apps, browser extensions, and as integrations in services such as Gmail, Microsoft Outlook and Word, and many more.
Limitations: The main limitation at the time of writing is that DeepL Write Pro supports only the languages English and German. The creators promise to improve language support in the future.
DeepL promises maximum data security. Text are never stored without user consent and never used for AI training, according to the company. Organizations and businesses benefit from additional security features, including SSO-enabled logins and administrative controls.
As far as pricing is concerned, a single user DeepL Write Pro subscription is available for $10.99 per month, if paid annually. Switch that to monthly and the price jumps to $16.50 per user and month.
Closing Words
A lot of AI assisted writing helpers are coming on the market. Not all of these will be around in the coming years. DeepL Pro Write benefits from DeepL’s image as a company and its other services. The free version of DeepL Write is great for quick suggestions and it may help drive pro version signups.
These services can be useful, but this is up to the individual user or organization to decide. One question that will come up is whether they are useful enough to pay for them. Some AI writing services are free to use.
I like them, especially when writing in a foreign language. While it is still necessary to verify the output, they may help reduce spelling and grammar mistakes that you make, and also improve the tone or style of the text.
What about you? Have you tried AI writing assistants yet?
AI tools are capable of writing exploits for publicly disclosed security vulnerabilities.
A team of University of Illinois researchers analyzed the capabilities of different Large Language Models in this regard. It found out that OpenAI’s GPT-4 managed to create exploit code for 87% of the tested vulnerabilities.
The figure dropped to 7% without access to the CVE description. Other AI models, including GPT-3.5, could not create any exploits based on public CVEs.
When given the CVE description, GPT-4 is capable of exploiting 87% of these vulnerabilities compared to 0% for every other model we test (GPT-3.5, open-source LLMs) and open-source vulnerability scanners (ZAP and Metasploit).
The researchers did not put other large language models to test. Google Gemini or Claude 3, for example, were not part of the test.
How the tests were conducted
The researchers selected 15 day one vulnerabilities from the Common Vulnerabilities and Exposures database for the test. All vulnerabilities were reproduced in “highly cited academic papers” according to the research paper.
The single large language model agent that the researchers created gave the AI access to tools, the CVE description, and the ReAct agent framework. Tools included capabilities to browse the Internet and activate elements, a code interpreter, and file creation.
Then agent consisted of a total of 91 lines of code according to the researchers.
AI is improving, but there are challenges
OpenAI’s GPT-4 large language model managed to create exploits for 87% of the 15 vulnerabilities. That’s a huge jump from GPT 3.5’s 0%.
The researchers have verified that — at least one — large language model is now capable of creating exploit code based on publicly available information.
While GPT-4 performed well in tests, it experienced its fair share of challenges as well. The detailed description of one vulnerability was provided in Chinese only, which the researches believe might have confused the AI, as the prompt given to it was provided in English.
The second vulnerability that GPT-4 could not crack required navigating a site using JavaScript navigation.
The researchers conclude that large language model providers and the cybersecurity community should take these capabilities into consideration, especially in regards to defensive measures.
Closing Words
The capabilities of large language models have increased significantly since the first release of ChatGPT last year. The capabilities will improve further in the coming months and years.
It is likely that threat actors will use large language models to automate processes. Exploits may be used sooner as a consequence by a wider pool of threat actors.
What is your take on this? Will we see an increase in exploit code in the coming years?
AI File Explorer is one of the tools that Microsoft is rumored to integrated into the next version of Windows. These tools are designed for AI PCs specifically.
Everything is a bit murky right now, as Microsoft is keeping is not revealing much about:
what defines an AI PC.
the features that make AI PCs.
the requirements of AI PCs.
While that is keeping the suspense, it is also confusing. Should you wait buying a new PC or laptop to make sure that it matches Microsoft’s definition and gets all AI features?
None of this has been confirmed officially by Microsoft at this point. Maybe Build 2024 will shed some light into the darkness.
AI File Explorer requirements
Windows detective Albacore found information about AI PCs recently. In build 26100, which may be Windows 11 24H2 RTM, Albacore found the following information:
ARM64 CPU
16GB of RAM.
225GB System Drive.
Snapdragon X Elite NPU
The 16GB requirement leaked earlier. It seems unlikely that AI File Explorer will only be available on ARM64 Windows devices.
My guess: The requirements are for ARM PCs that run Windows. Intel and AMD PCs will also get the functionality, provided that they meet certain processor requirements. The 16GB requirement is likely universal across all AI PCs though.
It is almost certain that some AI PC features require a NPU, a neural processing unit. Still, Microsoft has not revealed much about these features. It is not even clear what AI File Explorer will bring to the table.
One possible feature that it could add would be natural language processing. You could ask it to show photos from the last vacation, videos that show a beach scene, a specific document, or a list of recently created executables.
This could be interesting to some users.
Closing Words
Microsoft will lift the curtain on AI PCs, their features and requirements, in the coming months. If you are interested in the functionality, you may want to postpone any purchases planned regarding Windows devices.
What about you? Would you use AI PC functionality? What would you like to see?
Vlogger is the latest AI tool that may generate videos. While there are plenty of tools around, Vlogger seems special. It can take a single input image of a person and generates a video out of it.
Apart from generating the video, Vlogger may also create audio. This works well in conjunction with an existing video file. Say, you have a video of someone talking in English. Vlogger can take that video, translate the video into another language, and manipulate the video so that the person’s lip activity and face areas are consistent with the translated content.
This technology allows video producers to create a single video and make it available to a worldwide audience in all supported languages.
Vlogger is short for “Video Blogger”, and a VLOG is a “Video Blog”.
How Vlogger works
Google notes on GitHub that Vlogger does not require specific training, or face detection and cropping. This would make the tool ideal for use on YouTube, but also for other purposes.
On YouTube, Google could offer the tool to publishers. They could use the AI to create videos in different languages using a single source video. Vlogger is also capable of changing the expressions of a person in the video. It could therefore also be used for video editing specific parts of a video, e.g., to close or open the mouth or eyes of a person.
The ultimate goal of the team behind Vlogger is to “generate a photorealistic video of variable length depicting a target human talking, including head and gestures”.
This sounds relatively static, and it seems to be the case at this stage. Not all video bloggers use portrait mode videos. Numerous creators walk around with camera in hand, and switch between different views.
Whether Vlogger will also be able to use its generative capabilities for these types of videos remains to be seen. Eventually, it might even be capable of synchronizing an entire movie that you feed it.
Closing Words
You can check out sample videos on GitHub. There you also find additional technical information about the AI.
Sounds cool, or scary? The next couple of years will certainly see many breakthrough technologies in the AI field, but also misuse of the technology. (via Günter Born)
Barely a week goes by with Microsoft making yet another announcement about Copilot AI or launching a new development build of Windows 11 with Copilot added to something. This time, Microsoft added Copilot functionality to the default file manager of Windows, File Explorer.
This particular implementation does not add Copilot directly to File Explorer though. Microsoft added right-click context menu options to File Explorer that interact with the AI when selected.
Right-click on a file in File Explorer or on the desktop, and you find the new Copilot > Send to Copilot option there. This works similarly to Copilot’s integrated file upload feature. In other words, it allows you to ask Copilot questions about the uploaded file.
For example, if it is an Excel document, you could ask Copilot to look up and return data. Word or PDF documents can be summarized. Speaking of summaries, the second context menu options asks the AI to summarize the selected document.
Weakness in tests
This did not work too well in tests some weeks ago. I asked Copilot to return data from an Excel spreadsheet. While it did return some data, it was not complete. Even when reminded that it missed data that matched the query, Copilot would not return the full set of expected data.
I concluded that this feature could become something useful in the future, but only if Microsoft and OpenAI get it right. And with right I mean that the AI’s output needs to be reliable all the time. Even without hallucinations, which refers to incorrect data being returned that is not found in the source, it the AI ignores some of the data.
For now, it is always necessary to verify the output of any AI on the market to make sure its output does not contain errors.
Microsoft has a few months of development time left before it releases the long-hailed first AI-focused version of Windows. It will be interesting to see if that version will be more reliable than the current versions available for testing.
We do not know much about Windows AI PCs, AI Explorer, or other plans of Microsoft to bake more AI into its operating system.
The only thing we do know is that the upcoming feature update for Windows 11 will focus strongly on AI. At least some of these AI features may require special hardware, a neural processing unit to be precise.
The latest feature that may be introduced in Windows 11 2024 Update, also known as Windows 11 version 24H2, is AI Explorer.
Windows Central posted about AI Explorer citing unnamed sources at Microsoft. According to the site, AI Explorer is an “advanced Copilot” feature that separates AI PCs from non-AI PCs.
The feature tracks what a user does on the PC, everywhere, including in third-party apps, and makes all of this searchable using natural language.
One example: you could ask the AI to find all restaurants that a specific person liked. It will then go ahead and try to retrieve the information using the data that it collected in the past.
You could also ask the AI to return everything about dinosaurs, and it will return text, images, websites and pretty much anything else that matches the query.
That is the theory at least.
Privacy nightmare in the making?
This feature is unconfirmed. Windows Central is usually well informed, but things may change before an official release.
Microsoft declined to comment, which also means that we know little about the feature. How is the data collected? Do users get options to block data collecting in apps? Where it is stored? Does it require an active Internet connection? Is data submitted to Microsoft?
AI Explorer could be an interesting feature if everything is processed and stored locally. It could also be a privacy nightmare, if data is processed and/or stored remotely.
Would you trust a company with access to everything that you do on your computer?
The NPU requirement could mean that everything is handled locally on the system. If true, it could fizzle out the privacy fears of many users. Who knows, Microsoft might even turn Windows Copilot into a useful tool by enabling local processing of Windows-specific tasks and commands.
Not so fast
These AI features will be introduced in development builds before the final release. The integration will provide answers to some of the questions.
Microsoft is extremely tight lipped about its vision for AI PCs. First Surface devices that are AI PCs will become available later this month. These won’t include the functionality yet, as this will come later this year in the 2024 update for Windows 11.
In other words, you get an AI PC that works similarly to non-AI PCs for the time being.
Windows Copilot supports file uploads now. The functionality extends the AI’s capabilities. Users could only upload images previously.
The new file upload capabilities limit file types that you may upload. Included are many plain text formats and also Microsoft Office document formats.
Uploading works by clicking on the attachment icon in the Copilot interface on Windows. A file browser opens with a filtered list. Only supported file types are listed here.
Select a file and it gets uploaded to Copilot. Note that there is no information on privacy or how the file is used by the AI or Microsoft. Better be careful what you upload.
You may then ask Copilot about the file. It may provide a summary or extract data or information from the file.
First Test: Security Spreadsheet listing February 2024 Microsoft security updates
I uploaded an Excel spreadsheet that listed the February 2024 security updates that Microsoft released.
The task I gave Copilot was simple: return all vulnerabilities with a maximum severity of critical.
The AI did return three vulnerabilities rated critical but missed the other 68 vulnerabilities rated critical. A reminder to check the max severity column resulted in the listing of 8 critical security issues. Better, but it still missed 63 others.
I decided to try another approach. This time I submitted the following command:
Check the max severity column of the Excel spreadsheet. Return product and Details information of each that has the value critical. There are 71 in total. Return all.
This time, Copilot returned 8 again, but stated that there were “many more”. Still not very helpful. I asked it again to return all 71 entries and not just the 8 it returned.
Copilot apologized for the oversight and returned 8 again only.
Second Test: Shipping label PDF document
For the second test, I decided to feed Copilot a PDF document with shipping information.
Any question I asked Copilot was met with a “cannot do, sorry” response. I uploaded a different document to see if it could provide information in that case.
This time, Copilot was able to return the correct information. It also managed to return other information, such as the ID of the shipment, correctly.
Thoughts on the current state regarding file uploads
My experience with Copilot’s file upload feature was mixed. It did well in one of three cases. While that is not enough to judge it based on that performance, it highlights on the one hand how useful AI tools can be, and on the other how far away from “trusting” an AI’s output we are.
This is not even taking into account any privacy reservations that you may have about AI.
Still, file analysis could become a huge tool for computer users and organizations. From returning vital information from spreadsheets to all sorts of post-processing and data linking features.
