Technology News

Windows 11 and KB5023778 update

A new cumulative update for Windows 11, KB5023778, has been available for a few hours. Optional, unrelated to security. It offers some new features and bug fixes.

Download and installation are not automatic. Obviously it is necessary to intervene in Windows Update. Note that this is a preview of some of the content that will be offered during the next Patch Tuesday (April 2023).

KB5023778 targets PCs running Windows 11 22H2. The release notes mention several bug fixes, improvements, and changes. One of the most striking facts concerns the start menu.

Microsoft introduces notifications about accounts. At first glance, the approach seems interesting, but beware. For some, the suggested screenshot reveals an “advertising” approach. We have a notification that resembles some kind of advertisement, promoting the benefits of the company’s cloud service.

Windows 11 and OneDrive service notification from Microsoft

It suggests backing up some folders and files to a connected cloud service. While not specifically mentioned, it’s an invitation to use Microsoft’s OneDrive service. It is possible to read

Back up your files. Folders like documents and pictures are stored in the cloud to protect them.

The notification contains several options

• backup start,
• account switch,
• screen lock
• log out

The approach is not surprising as testing has been going through various builds released in the Windows Insider program for quite some time. At the moment this functionality is only offered to a small number of users. It is possible that Microsoft would like to collect comments to see whether or not a larger scale deployment occurs.

Windows 11 – Manage notifications in Settings

Windows 11 has an option to disallow this type of notification. It’s called “Suggest ways to get the most out of Windows and finish setting up this device” and is at this address

Settings > System > Notifications > Additional settings

Windows 11 22H2 and KB5023778, multiple bug fixes

KB5023778 also improves the taskbar search box with custom color modes. Microsoft states that it has corrected it

• An issue affecting the Notepad combo box in Preferences. Not all available options are shown.
• An issue with PowerPoint that may crash when using accessibility features.
• A screen reader crash that fails to read items in drop-down lists in Microsoft Excel.

Finally, it fixes the bug with USB printers perceived by Windows 11 as multimedia devices, while announcing new features and improvements for Microsoft Defender for Endpoint.

Experts from all walks of life have been warning of the potential of artificial intelligence (AI) to create flawed images and videos for years. In its early days, deepfaking was fairly easy to distinguish, but with recent advances and because of the dangers of misinformation it poses, the “open bar” that was Midjourney and its deepfaking services had to be shut down (deepfakes) online.

We know the rapid advancement of AI, especially in ChatGPT-4. But the new version 5 of Midjourney’s deep faking software has dramatically improved the quality of images depicting people.

For example, misshapen hands, which used to be a feature of AI-generated images, are now less of an issue. Lighting and fabrics are also more realistic, and the system is capable of generating images of countless celebrities and public figures for TheVerge to read.

These improvements have resulted in a number of midjourney shots going viral in recent weeks. These include not only fake photos of Trump’s arrest and the fashionable Pope, but also French President Emmanuel Macron marching through the ongoing protests in Paris and Elon Musk holding the hand of Alexandria Ocasio-Cortez.

Montage of three fake photos implying the Pope, Elon Musk, Alexandria Ocasio-Cortez and Donald Trump.

Full enforcement of moderation policies

Midjourney’s content restrictions are more permissive than some competing services (like OpenAI’s DALL-E) but more restrictive than others (e.g. Stable Diffusion).

Additionally, Midjourney’s response to the escalating politically sensitive threat of fake images has been piecemeal thus far, without significant overhauls to its moderation policies.

There is actually a banned words list based on complaints from users from some countries, but almost nobody pays any attention to it. Even simpler: until recently, using other terms was enough to get the desired deepfaking results.

• Hear Martin Bouchard, President and Co-Founder of QScale, on the mic mario dumont, Available as a podcast on QUB radio :

too much abuse

Currently, non-paying users can no longer generate images, with Midjourney claiming a shortage of GPUs. Then the tech hinted that the misuse of the software was the main reason for stopping the free trials.

However, the service remains open to paying subscribers.

Ce logiciel nourri à l’intelligence artificial permet de créer des images hyper-réalistes, mais also de veritables oeuvres d’art.

Pour tester Midjourney, il faudra désormais sortir le portfolio. Ce logiciel d’intelligence artificielle, qui permet de créer des images hyper-réalistes, an annoncé ce jeudi 30 mars arrêter temporairement son essai gratuit, rapporte The Washington Post. En cause: “une demande incroyable et des abus dans les essais”, a écrit ce jeudi le fondateur et PDG David Holz sur le réseau social Discord.

A la difference of logiciels traditionals de photomontage like Photoshop, qui nécessite quelques compétences, midday permet de created instantanément en indiquant seulement quelques mots-clés de description pour le résultat recherché. En un an seulement, l’entreprise basée à San Francisco a séduit déjà 13 millions de membres.

Un essai gratuit jusqu’à 25 images

Mid-journey autorisait jusqu’ici tous les internautes à create gratuitement 25 images, avant de proposer un subscription allant de dix à soixante dollars par mois. Mais l’entreprise américaine a published a new version of its logic a la mi-mars, qui a significant amélioré of the quality of the image – notamment celles representative of the personnalités publiques.

De quoi provoquer la creation de multitudes d’images devenues virales sur Internet ces dernières semaines, comme Emmanuel Macron sur un tas d’ordures ou an arrestation fictive de l’ancien président des Etats-Unis, Donald Trump. This is the series of photos that was part of the story and a comment on Twitter, so that his creator – the Britannian journalist Eliot Higgings – was a mid-day break.

A moderation peu restrictive

“Nous avons essayé de réactiver les essais gratuits avec de nouveaux gardes-fous contre les abus mais ils ne semblaient pas suffisants. Nous les désactivons donc à nouveau pour maintenir le service pour tous les autres”, a écrit David Holz.

Le logiciel est réputé pour imposer peu de restrictions. Dans ses lignes directrices, il precise que les “utilisateurs doivent avoir au moins 13 ans” et doivent “éviter de créer des contents choquants et dérangeants”.

En October dernier, le fondateur avait indiqué sur Discord que Midjourney avait établi a list de mots interdits “liés à des sujets dans different pays sur la base des plaintes des utilisateurs de ces pays”, sans thunder davantage de précisions. Le logiciel revendique une quarantine de moderators, selon le Washington Post.

Artificial intelligence is moving fast, and faster and faster. Now GPT-5 is coming, probably before the end of the year if we believe the rumors circulating on social networks.

We’ve talked a lot about ChatGPT since last fall, which allows you to have a conversation in a fairly natural language with an artificial intelligence that seems to know everything, and this application originally worked with a speech engine called GPT-3. The fourth generation of this engine, GPT-4, went live just two weeks ago.

GPT-5 could be a very significant upgrade. GPT-4 was already 10 to 1000 times more powerful than GPT-3. And GPT-5 would also be several dozen times more powerful than GPT-4.

When that happens, experts say we could witness the birth of the first real “artificial general intelligence,” called AGI in English. It sounds trite to put it that way, but it is the name of an AI that would be able to flawlessly perform any task that any human can perform.

For example, she would know how to fix a leaky sink, how to bake the best chocolate cake or how to balance your family budget. The only thing that even the general AI can’t do yet is our tax return, and that’s a shame…

Of course, the potential of such AI could also be used for malicious purposes. This is where things start to worry… But it seems it’s not enough to stop OpenAI, the company working on the development of the GPT-5 engine.

The next few months are therefore likely to be quite busy for AI experts.

Squid and cuttlefish are masters of camouflage, blending in with their surroundings to evade predators or surprise prey. Some aspects of how these cephalopods become reversibly transparent remain unclear, largely because researchers cannot grow cephalopod skin cells in the lab. Now, however, the researchers report that they have mimicked the tunable transparency of certain squid skin cells in cultureable mammalian cells. The work could not only shed light on the basic biology of squid, but also lead to better ways to image many cell types.

The researchers will present their findings at the American Chemical Society (ACS) spring meeting.

Alon Gorodetsky, Ph.D., and his research group have been working on squid-inspired materials for many years. In previous work, they developed “invisibility stickers” consisting of squid reflecting proteins produced by bacteria and stuck to sticky tape. “So we had this crazy idea to see if we could capture some aspects of the ability of squid skin tissue to change transparency in human cell cultures,” says Gorodetsky, the project’s principal investigator.

The University of California, Irvine team focused their efforts on cephalopod cells called leucophores, which have particle-like nanostructures made up of reflective proteins that scatter light. Typically, the reflectins clump together to form the nanoparticles so that light is not absorbed or transmitted directly; Instead, light scatters or bounces off them, making the leucophores appear bright white.

“We wanted to manipulate mammalian cells in such a way that they form stably rather than transiently reflective nanostructures for which we can better control light scattering,” says Gorodetsky. In fact, when the cells transmit light with low scattering, they appear more transparent. Alternatively, scattering much more light makes the cells opaque and more visible. “Then we thought we could predictably change the transparency of the cells relative to the environment or background at the cellular level or even at the culture level,” he says.

To change the way light interacts with cultured cells, Georgii Bogdanov, a graduate student in Gorodetsky’s lab who is presenting the results, introduced squid-derived genes that coded for reflecting into human cells, which then read the DNA used to produce the protein. “A key advance in our experiments was getting cells to stably produce reflections and form light-scattering nanostructures with relatively high refractive indices, which also allowed us to better image cells in three dimensions,” explains Bogdanov.

In experiments, the team added salt to cell culture media and watched reflective proteins clump together into nanostructures. By systematically increasing the salt concentration, Bogdanov obtained detailed 3D time-lapse images of the properties of the nanostructures. As the nanoparticles got larger, more light was reflected off the cells, adjusting their opacity.

Then the COVID-19 pandemic hit, and researchers wondered what they could do to advance their investigation without being physically in the lab. So Bogdanov spent his time at home developing computer models that could predict the expected light scattering and transparency of a cell even before an experiment was started. “It’s a nice loop between theory and experiment, where you introduce design parameters for reflective nanostructures, get specific predicted optical properties, and then make the cells more efficient – for any light scattering properties that might interest you,” says Gorodetsky.

Basically, Gorodetsky suggests that these results will help scientists better understand squid skin cells that have not been successfully cultured in the laboratory. For example, previous researchers have postulated that reflective nanoparticles disassemble and reassemble to alter the transparency of tunable squid leucophores. And now Gorodetsky’s team has shown that similar rearrangements occur in their stable mammalian cells modified by simple changes in salt concentration, a mechanism that appears to be analogous to what has been observed in tunable squid cells.

The researchers are now optimizing their technique to develop better cell imaging strategies based on the cells’ intrinsic optical properties. Gorodetsky envisions that reflectin proteins could act as genetically encoded markers that don’t turn white in human cells. “Reflectin as a molecular probe offers many opportunities to trace structures in cells with advanced microscopy techniques,” adds Bogdanov. For example, the scientists suggest that imaging approaches based on their work could also have implications for better understanding cell growth and development.

Researchers acknowledge funding from the Defense Advanced Research Projects Agency and the US Air Force Office of Scientific Research.

The development and spread of artificial intelligence programs worry several experts so much that a hundred of them, including Elon Musk, have requested a six-month pause in system training.

• Also read: Experts call for a six-month hiatus for artificial intelligence

• Also read: French: that ad that puts powder in your eyes

• Also read: Artificial intelligence: ChatGPT saves a dog’s life?

However, the spread of images created from scratch by artificial intelligence has already begun.

Can you tell a real photo from a fake? French magazine L’édition du soir, in collaboration with AFP, has put together a little guide to help you find your way around.

Here are five tips on how to recognize an image created by artificial intelligence:

Analyze the picture carefully

In some cases, it’s relatively easy to tell if an image was created by an artificial intelligence. Enlarging a photo with a zoom can make it possible to see small assembly errors.

Some programs that create images using artificial intelligence watermark their creations. A careful examination of the image can therefore make it possible to discover these signatures left by the programs.

Identify the source of the image

Knowing who posted a photo and on which site can certainly help identify a fake image. An official account can confirm the authenticity of the photo to us, while we are entitled to question more about the authenticity of the photo if the image was posted on an anonymous or very recent account.

It’s also a good idea to analyze a photo’s metadata to know if it’s true. Metadata is the digital identity of a file and can therefore help to determine whether the image being viewed was created by an artificial intelligence.

How do you find metadata? Just save the image to your computer, then go to “Properties”, then “Details” of it (PC) or go to “Read Info” (Mac).

However, it should be noted that most social networks delete metadata.

Reverse image search

When searching for the source of the photo, you can also use Google’s reverse image search to find out if the file already exists elsewhere and when it was first uploaded.

You may also discover the original file and determine if the photo you are wondering about has been modified and how. You may also be able to find a higher resolution version of the image, which will help you spot possible editing errors that suggest the photo is fake.

Examine the details

Although very advanced, artificial intelligences are not immune to visual inconsistencies. As with the 7 mistakes game, if you look closely at the photo, you may spot ill-defined details in the image.

Watch the background

Speaking of details, the background is often overlooked when an image is created by artificial intelligence.

Some flaws, such as poor perspective or a poorly observed size scale, can therefore slip into the background of the photo. The furthest away elements are the ones most prone to inconsistencies.

Between e-readers and good old paper books, the rivalry will not fade. On one side we have these shelves that can hold the equivalent of an entire library, on the other we have these pages that smell like paper but weigh a bit heavy once in your pocket.

Buy the Kindle Paperwhite for $104.99 from Amazon

This Amazon Kindle Paperwhite promotion can help you decide, as this e-reader, considered by many to be THE best today, costs between €149.99 and €104.99 on Amazon.

No more heavy paper books: The Amazon Kindle weighs only 205 grams!

This is the latest version of Amazon’s Kindle Paperwhite. This reader offers a beautiful 6.8 inch screen : neither too big nor too small! she is ideal for reading in the evening with the lights off as the warm light won’t strain your eyes. You can even adjust the brightness of the lights to your liking!

The screen of this new model is much brighter and more pleasant to read with its anti-reflective treatment, while the reader is generally smoother and faster. Important point, it has a USB-C port for faster charging and above all to avoid having to carry too many different chargers with you.

The 8 GB Kindle Paperwhite can store several thousand books. Realize that going on vacation is much more convenient than carrying several paper books in your bag. In short, just tell yourself that you can have the equivalent of a well-stocked library in this small tablet.

If this reader supports the vast majority of eBook formats on the market, it does not read EPUB format and that is the only major flaw of this device. And if you have books in EPUB format, you can still read them, but you will need them. switch before.

Europol, the European criminal police agency, has just published a report dedicated to this The emergence of linguistic models, like GPT, GPT-J or even LLaMA. According to the agency, the rise of generative AI is likely to disrupt “all industries, including the criminal industries.”

Also read: ChatGPT leaked confidential user data

Why Europol fears the rise of AI

Tools like ChatGPT have “huge potential” for criminals, Europol notes with concern. The report also notes that the criminals already started to exploit the possibilities of AI to commit crimes.

“ChatGPT is already capable of facilitating a significant number of criminal activities, […] such as terrorism and the sexual exploitation of children,” explains Europol, stressing that the gangsters are “generally quick to exploit new technologies”.

For example, scammers can rely on ChatGPT Write compelling phishing messages or ransom demands. Europol officials are also concerned that generative AI is being used to spread false information and propaganda on social media. With images designed by Midjourney or Stable Diffusion and messages written by ChatGPT, it is possible to manipulate internet users and “encourage them to trust criminal actors”. Some criminals are already using AI-generated videos to trick YouTube users. These dummy sequences direct internet users to corrupted download links full of viruses.

The agency also fears that the AI ​​will be used by pirates to orchestrate brute force attacks or generate fake documents such as invoices and contracts. With the help of ChatGPT, some hackers even started designing malware like ransomware or data stealing without having much technical knowledge.

“The possible exploitation of such AI systems by criminals offers bleak prospects,” warns Europol.

A source of information

In general, a criminal can rely on ChatGPT or any other intelligent chatbot Learn more about how to commit specific crimes. To avoid discrepancies, OpenAI has implemented filters and safeguards.

Unfortunately, it is possible to circumvent these limitations using a “prompt injection” attack. This type of attack, known to OpenAI, consists of talking to ChatGPT to convince it to ignore its programming and change its behavior. This is how some users pushed the model to teach them how to build a bomb or produce cocaine. This information was already available on the web, but the chatbot makes it easier to access.

“ChatGPT can significantly speed up the search process by providing key insights,” the agency believes.

The counterattack is organized

In this context, Europol encourages law enforcement agencies to look into the matter and prepare for “what’s to come”. At the same time, the agency, which is primarily dedicated to the exchange of information between national police forces, recommends “raising public awareness” and building security mechanisms with the help of industry.

Finally, Europol believes that police forces should do so explore the possibilities that AI offers. The report discusses building a language model consisting solely of data provided by law enforcement agencies. This conversational agent would assist the police in their activities, just as ChatGPT facilitates the task of criminals…