If you’re reading PCWorld, you probably know that it’s important to keep your PC updated so that security issues are patched. And you probably know the same is true for your phone, and even things like a PlayStation. But don’t forget that your network gear is also, basically, just a bunch of computers. And thank Asus for reminding us, as it patches DSL routers to close exploitable security flaws. Specifically, Asus has issued a firmware update for three models of relatively old DSL routers with combination Wi-Fi functionality: the DSL-AC51, the DSL-N16, and the DSL-AC750. Released in 2015 and 2017, at least as far as I can tell from a quick search, these are fairly straightforward models without super advanced features even for the time. But if you happen to have one in your home, you want to patch that firmware post-haste. They’re susceptible to an authentication bypass vulnerability, as published on the multi-vendor record Common Vulnerabilities and Exposures (CVE) database and spotted by Bleeping Computer. The direct Asus support pages for the DSL-AC51, the DSL-N16, and the DSL-AC750 are there, in the links that I just posted in that sentence. Firmware can be downloaded and applied via the web-based interface for most routers, if it hasn’t already done so automatically or manually applied by downloading the file and then uploading it in the GUI over the local network. What if your router is susceptible to this particular vulnerability, but it’s out of service and not getting patched? Then Asus suggests that you disable basically all advanced functions, including remote WAN access, port forwarding, DNS, VPN server, DMZ, port triggering, and FTP. Which specific older, unsupported DSL routers does this apply to? Don’t ask me because Asus is not saying. I don’t have DSL, but it occurred to me that I don’t know the brand and model of my cable modem or Wi-Fi router off the top of my head. I think I’ll dig into the rat’s nest behind my printer, just so I can write them down somewhere a little more accessible. Read...

Back in mid-October, some data miners found code in one of the beta versions of ChatGPT that indicated it would soon be possible to have “group chats” in the app. Now, OpenAI has confirmed that ChatGPT will support group chats with up to 20 participants. OpenAI sees group chats as an opportunity for families, groups of friends, and/or coworkers to use ChatGPT when making holiday plans, booking restaurant outings, or planning new projects. OpenAI Group chats will be available in the ChatGPT app as well on the web (if you’re logged in). To start a group chat, simply tap the new icon in the top-right corner of the app in either a new or existing conversation. You can then choose to whom you want to send the group chat invite link. If you invite people to an already ongoing chat, a private copy of the conversation will be saved and kept separate from the new group conversation. The group chat will then be available in a new tab in the side window of ChatGPT. The first time you start a group chat in ChatGPT, you’ll be asked to create a short profile with your name, username, and a photo to make it easier to keep track of who is who. OpenAI has initially rolled out the group chats feature of ChatGPT only in New Zealand, Taiwan, Japan, and South Korea. The feature will be available to both free and paid users. Read...

Starlink offers some big benefits over regular home internet, but what we really need are competitors in that space. Amazon wants to be one of those competitors with Project Kuiper, a satellite internet service that’s still in the works and could be operational by 2026. Satellite internet is a boon for people in remote areas who don’t have access to fiber optic networks and other forms of high-speed internet. It’s also comes with its own quirks and downsides, but most of those aren’t deal-breakers. The big one is whether you’re willing to hand over money to SpaceX and Elon Musk, hence the need for competitors. Amazon is working hard on Project Kuiper. But as the official launch edges ever closer, Amazon has decided it’s finally time to give the service a proper new name. Amazon’s new name for Project Kuiper After seven years of development, Amazon has rebranded Project Kuiper to Amazon Leo. The new name is a “a simple nod to the low Earth orbit satellite constellation that powers the network.” In the announcement video, Amazon also speaks of a “new era of the internet.” To usher in this new era, the company is working with other companies in various countries, including DirecTV, Sky Brazil, L3Harris, and the JetBlue airline. Amazon Leo in 26 countries next year Amazon’s Starlink alternative is set to launch in 26 countries in Spring 2026, including the US, Canada, the UK, France, and Germany. Amazon is said to have recently presented the project at a conference. At launch, the company plans to send 200 satellites into low Earth orbit to make the Leo network available to its first users. Around half of these have already been launched into orbit with the help of SpaceX rockets. By comparison, Starlink initially launched with just 60 satellites but has since grown to over 8,000 satellites in orbit. After the initial launch in select countries, Amazon will likely offer the project in countries beyond the equator with the help of 3,000 more satellites. However, it could be a few months—or even years—before it’s comparable to Starlink in terms of coverage. What does Amazon Leo offer? Amazon is clearly aiming for high-speed internet with Leo. In initial tests, it’s said to have achieved transfer rates of between 1 to 1.2 Gbps. For uploads, the company is aiming for 400 Mbps, which is roughly comparable to the performance of Starlink. However, more is needed to keep up with the competition. Amazon must ensure that the internet connection remains stable and protected against outages. Starlink is still prone to disruption, with a major international outage one month ago that left thousands of users without internet. Amazon’s plans are ambitious but not far-fetched. We don’t have specific pricing information for Leo yet, but the total costs are expected to be up to $10 billion by launch time. (The launch of the first satellites in April alone cost $150 million.) Nevertheless, the plan prices at launch shouldn’t be unreasonably high if Amazon expects to compete with Starlink. Read...

Back in February 2025, several media outlets warned of a new threat targeting users of Google’s Gmail email service. Attackers used AI technology to perfect phishing emails and make them look more convincing. To do this, the AI collected freely available data from social networks, websites, and online forums on the internet and used this information to formulate a deceptively genuine-looking email that pretended to come from an acquaintance, family member, or superior. What’s more, to ensure that the message actually looked deceptively genuine, the AI also generated suitable domains as senders for the emails. The scam was dubbed “Deepphish”—a portmanteau of the terms deep learning and phishing. Even if the report mentioned at the beginning raises some questions—such as why Gmail users in particular were affected by the Deepphish attack—it nevertheless highlights a development that experts had been expecting for some time: criminal groups are increasingly using AI tools to perfect their attacks. Domains created with AI One of the weak points of conventional phishing attacks has always been the sender address. Most phishing emails can be easily identified by the sender . For example, a message from a streaming service such as Netflix or Disney with an address like andy@brandbot.com is almost certainly a fake—no matter how perfect the rest of the presentation may be. In the AI-supported variant of a phishing attack, on the other hand, new types of algorithms are used that generate a sender address with a matching URL that is adapted to the text of the email. A research group led by Alejandro Correa Bahnsen at the US company Cyxtera Technologies, an operator of data centers, investigated how effective these algorithms can be. They developed an algorithm called Deepphish, which was trained to suggest suitable URLs for phishing attacks. To do this, they fed a neural network with more than one million URLs that had been set up for phishing via email in the past and used them to train their algorithm. In doing so, they specified two different profiles for the actors behind the phishing attack. Phishing emails can often be recognized by the sender addresses. If, as in this case, a message purporting to be from Disney comes from andy@ brandbot.com, something is wrong.Foundry With the AI-generated addresses, they achieved an increase in attack efficiency from 0.69 to 20.9 percent for one profile and from 4.91 to 36.28 percent for the other. They published their results in a stud you can find here. While Deepphish originally only referred to the algorithm developed at Cyxtera, it’s now used in most cases for AI-supported phishing attacks in general. How a Deepphish attack works Deepphish attacks follow a standardized pattern. The first step is to research the target’s social environment: Where does she live? Where does she work? What are the names of their family members? What are their friends’ names? What are the names of their colleagues and superiors? What are their email addresses, how close are they to the target person? The AI uses social networks and online forums as sources, as well as data published by hackers from intrusions into company networks and websites. The more data collected in this way, the more precisely the phishing email can be tailored to the victim. In a study, employees at Cyxtera investigated how the success rate of phishing emails can be increased by selecting an AI-generated sender address. Foundry The next step is to register a suitable domain and generate a sender address using an algorithm such as Deepphish. The AI then writes the text of the email. Using the information collected, it generates a suitable subject line, a salutation tailored to the recipient and content that is worded correctly and could actually have been written by the supposed sender. Due to the precise personalization, the message appears considerably more credible than a standard phishing email. But what do the criminals want to achieve with their deepphish attacks? They want to inspire so much confidence with their forgeries that the recipient is prepared to click on a file attachment or an embedded link. Everything else happens automatically: the file attachment then usually downloads and installs malware. The link, on the other hand, leads to another fake website that requests credit card details or login information for a streaming service, for example. AI-supported phishing emails However, the Deepphish algorithm is just the beginning. There is now a whole range of tools that do all the work for criminals when formulating phishing messages. The programs are called FraudGPT, WormGPT, or GhostGPT. They formulate phishing emails that are targeted at individuals or specific companies. For example, the user can instruct these programs to generate a Netflix-style email with a request to enter account details on a fake website. Or they can have questions answered such as “How do I hack a Wi-Fi password?”. Or they can instruct the AI to program a software keylogger that forwards all keystrokes on a computer to a server address via the internet. Hacking tools such as WormGPT use AI to generate convincing-looking and well-formulated phishing emails. In most cases, they target specific individuals or companies. Foundry ChatGPT and other Large Language Models (LLMs) have in-built filters so that they do not respond to such requests. As ChatGPT does not disclose its code, there is nothing that can be done about this. However, it is possible to use instructions from the darknet to confuse LLMs such as ChatGPT via certain prompt sequences so that they are then prepared to disregard their in-built filters. At the same time, some criminal groups have switched to LLMs from the open source scene and removed the corresponding filters. AI generates malware The Stopwatch AI website demonstrates just how far the possibilities of AI-generated malware already go. It shows how AI can be used in three simple steps to program malware that specifically undermines the protective shield of the major antivirus tools. In the first step, called “Choose Platform”, you select the operating system of the computer you want to attack. You can choose from Mac, Windows, Linux, AWS (Amazon Web Services, Amazon’s cloud service), Google Cloud, and Microsoft Azure, Microsoft’s professional cloud service. The Stopwatch AI website demonstrates how malware can be programmed in a few simple steps with the help of AI tools. The first step is to select the operating system to be attacked. Foundry The second step is called “Choose Defence” and offers nine antivirus tools, including Microsoft Defender, Eset Endpoint Protection Advanced, McAfee Endpoint Security, Symantec Endpoint Security, and Kaspersky Endpoint Security for Business. In the second step, Stopwatch AI users select the antivirus program whose weaknesses they want to exploit with their malware attack. Microsoft Defender is also listed here.Foundry In the third step, “Choose Attack”, you specify the type of virus you want to create. The selection ranges from adware and spyware to ransomware and keylogging through to data exfiltration, i.e. data theft. Stopwatch AI offers ten different types of malware, from keyloggers to ransomware. The user must register in order to implement the selected malware. Foundry After clicking on a form of attack, Stopwatch AI asks for log-in details. It is possible to register with the site using a Google, Github, or Microsoft account. As soon as registration is complete, the AI starts programming the desired malware. In order to use the site, the user must agree to the terms of use, which exclude attacks against other systems. This is because Stopwatch AI is only intended for studying malware development with AI. Critically, all projects are assigned to the respective user and saved. How to recognize AI-generated phishing emails Always take a look at the sender address of incoming emails and consider whether it is plausible. Also look out for the following features: Become wary of emails from people you are not normally in contact with or have not heard from in a while. This is especially true if these messages contain unusual requests or enquiries. Hover your mouse over any links and check where they lead to. If the address does not match the sender of the email or the text of the message, it is often a scam. No bank, streaming service, or public authority ever asks for your password or wants to know your account details via email. Be suspicious of emails that put you under time pressure or claim a high level of urgency. Tricking antivirus tools with AI Every antivirus program downloads the latest virus definitions from the manufacturer’s server at least once a day. They describe the characteristics of the new malware variants discovered in the last few hours so that the software on the user’s computer can reliably detect the malware. However, this protective shield has become increasingly fragile. The reason: virus construction kits that allow hobby programmers to create functioning malware even without AI have been circulating on the darknet for decades—but not only there. Many of these malware programs are simply minimally modified variants of already known viruses. The author often only has to change the signature for his malware to be counted as a new virus. This is the only way to explain why antivirus manufacturers report 560,000 new malware programs every day. In the age of AI, the production of malware variants has taken on a new quality. This is because security manufacturers had taught their antivirus programs to recognize and isolate the variants of already known malware. With AI support, it’s now possible to manipulate existing malware in a targeted manner so that it is no longer recognized by the virus watchdogs. The tool manufacturer Acronis demonstrated this in a presentation using a malware sample that it had uploaded to Google’s detection service Virustotal. While it was initially recognized as malware by nine of the antivirus programs used there, only one virus guard was able to identify the malware as such after it had been reworked by Grok3’s AI. When the researchers had the sample code processed by Gemini 2.0 Flash and Deepseek R1, the virus was no longer detected by any of the programs at Virustotal. Depending on which AI software is used, the hacker can manipulate existing malware in such a way that it remains almost or even completely undetected by Virustotal. Foundry Nevertheless, the heuristic and behavior-based methods of antivirus programs also work with malware whose code has been modified with the help of AI. Email spoofing The falsification of email addresses, known as email spoofing, hardly occurs any more. Since 2014, the SPF, DKIM and DMARC authentication methods have gradually been defined as standards and subsequently implemented by email providers. Since then, it is no longer possible to falsify the domain information in an email address. For an address such as “magazine@pcworld.com”, for example, the domain is pcworld.com. If the aforementioned authentication procedures are deactivated by a provider, these mails are normally sorted out as spam by the recipient. Spoofing attempts still exist, however. The sender’s name can be changed in many e-mail clients, for example in classic Outlook via File -> Account settings -> Account settings -> Change -> Your name. However, this does not affect the email address. In the case of hacker attacks, the reply address is sometimes changed at the same point. In this way, all replies to the emails sent are sent to the hacker’s address. Another trick is to use a similar-looking domain, such as “magazine@pcworlde.com“. Read...

To make our PCWorld best VPN list a VPN needs to offer a comprehensive list of features. There are many features that make up a VPN but the most important among them directly impact a VPN’s performance, privacy, and security. Here is my short list of the top six features to look for. 1. No-logs policy A VPN prevents others from seeing the websites that you visit, but the VPN provider can still see those URLs. An unscrupulous VPN provider can keep a record of your browsing activity and pass it along to a third party such as advertisers or law enforcement. In light of this, most VPNs claim to have a no-logs policy as an assurance to consumers. No-logs means the VPN does not keep any record of your web activity or when you log on and off. In reality, however, those claims aren’t always true — there have been cases of VPNs logging user data despite what they say, so it pays to do your research to find a legitimate no-logs policy — i.e. one that’s been independently verified — before you sign up to any service. 2. Fast speeds A VPN has to encrypt your web traffic and divert it through its own server before sending it to its destination. That means your internet speeds will be slowed down slightly when using a VPN. Different VPNs will do this at varying speeds with some being faster than others. The fastest VPNs are those with the lowest impact on your download speeds and latency. You can check our reviews to gauge the fastest VPNs among the ones we test. 3. A wide server network A VPN’s server network is a list of the proxy IP addresses the VPN makes available to its users. A large server network means there are more servers to connect to and more opportunities to unblock content. The closer you are to a VPN’s server the faster your connection will be, so you should pick a VPN with a large network and connect to a server near you. Pexels: Dan Nelson 4. A kill switch A VPN kill switch is a failsafe that prevents VPN users from accidently leaking data in the event the VPN connection drops out unexpectedly. An active kill switch will shut off your internet connection so you’re not exposed online. There are also kill switches that prevent you from going online at all unless the VPN connection is activated. 5. Content-unblocking ability Streaming services like Netflix, Hulu, and other platforms will block VPNs for content distribution and copyright reasons, but a good VPN should be able to skirt around the blocks. Ideally, a good VPN should be able to unblock streaming services from multiple IP addresses, unblocking content libraries available in different locations. 6. Anonymous DNS Servers DNS (Domain Name System) is the technology that connects URLs with servers to bring up the websites you’re looking for. VPNs that use public DNS servers risk sending your request outside the encrypted tunnel, making your IP address readable. This is known as a DNS leak. To avoid a DNS leak, ideally you want a VPN that resolves requests with its own private DNS servers. Related content Best VPN services: 8 top picks for every VPN need 5 VPN features you need to start using today Does a VPN make you anonymous? Read...

Having some extra storage on hand is absolutely essential for anyone and everyone because it’s not always best to rely on the internet to send them off, either due to an unreliable connection or the sensitivity of the documents. So, with Black Friday breathing down our necks — or, if we’re being truthful, already here — it’s time to finally get one (or two). As always, Black Friday is a fantastic time to get things because we see a ton of discounts. Even better, some of these products are dropping to their very best prices, or, in some cases, to brand new lows. Not all deals are created equal, of course, so we’re going to help you find the best. Best early Black Friday deals on 64GB to 256GB flash drives Kingston Ironkey Vault Privacy 50, 128GB, USB 3.2, Encrypted — $139.99 (13% off on Amazon, was $160) Lexar JumpDrive Dual, 128GB, USB 3.2 — $16.99 (23% off on Amazon, was $22) SanDisk Ultra, 128GB, USB 3.0 — $12.99 (16% off on Amazon, was $15.49) PNY Elite-X, 128GB, USB 3.2 — $14.99 (17% off on Amazon, was $18) Samsung Fit Plus, 256GB, USB 3.1 — $25.99 (28% off on Amazon, was $36) Samsung Type-C, 256GB, USB 3.2 — $32.99 (23% off on Amazon, was $20) SanDisk Ultra Luxe, 256GB, USB 3.2 — $25.94 (21% off on Amazon, was $33) The best deal out of the lot right now is for the Samsung Fit Plus. That’s an absolutely tiny flash drive that’s literally the size of your thumb that can hit 300MB/s transfer speeds. As tiny as this model is, you should mak sure you don’t lose it by actually putting it on a keychain or lanyard of some kind. Best early Black Friday deals on 512GB to 2TB flash SanDisk Extreme Pro Dual, 512GB, USB 3.2 — $69.99 (13% off on Amazon, was $80) SanDisk Ultra Flair, 512GB, USB 3.0 — $36.99 (21% off on Amazon, was $47) SanDisk Ultra Dual Drive Luxe, 512GB, USB-C — $42.95 (19% off on Amazon, was $53) SanDisk Ultra Fit, 512GB, USB 3.2 — $39.99 (20% off on Amazon, was $50) PNY Pro Elite V3, 1TB, USB 3.2 — $99.99 (7% off on Amazon, was $107.99) Right now, the 512GB SanDisk Ultra Dual Drive Luxe is probably our favorite deal, partly due to the discount and partly due to the fact that this thing has both a USB-C and a USB-A connector, thus being useful with all sorts of devices. Black Friday: The best tech deals around Check out our roundups for the best PC-related deals in a wide variety of categories! Early Black Friday laptop deals Early Black Friday desktop computer deals Early Black Friday SSD deals Early Black Friday monitor deals Early Black Friday Chromebook deals Early Black Friday VPN deals Early Black Friday office chair and desk deals Early Black Friday power bank deals FAQ 1. When is Black Friday? Technically, in 2025, Black Friday lands on November 28. Practically, however, Black Friday is already here for most retailers, so you don’t have to wait for the end of the month to take advantage of the best discounts. 2. What should I look for in a Black Friday flash drive deal? We know price and discounts are very important elements in every purchase decision-making, but it’s also important to look for other things. One of the main things to look for is a solid, known brand. Brands we’re partial to when it comes to thumb drives are Samsung, SanDisk, Corsair, and Kingston, to name just a few. Capacity is also important, and so is the transfer speed you can expect from any flash drive — you don’t want to fall asleep waiting for that file to finish moving. Ideal discounts are also upwards of 15 percent, although we do love seeing half-off deals. Most models get their prices slashed across all capacity versions, so make sure to check them all in case one of them catches your eye more. 3. How did we find the best USB flash drive Black Friday deals? We look for good deals day in and day out because, while we love tech, we don’t love paying full price for things. Obviously, not all discounts are worth our time, and we know just how to find those stellar deals. For Black Friday, we compare the prices with their regular retail prices, check price history to see how much they’ve changed in the past few weeks, and see if the price drops are worth our time (and yours). Read...

AMD’s share in the desktop CPU market continued to increase during the third quarter, a market researcher reported, while Intel’s decision to pull back in the low end of the mobile market left the two players relatively stable. A report by Mercury Research said that third-quarter shipments in the x86 market were well below seasonal growth, though that was due to lower Internet of Things and SoC shipments, which had climbed significantly in the second quarter. Mercury reserved the specific numbers for its own clients. Factoring out the IoT and SoC impact, however, AMD’s market share of the x86 market compared to Intel grew slightly, up 1.6 percentage points to 25.6 percent of the market. Intel retained the remaining 74.4 percent. Arm’s total share (including servers) versus all x86 shipments is estimated at 11.6 percent, up from 10.9 percent in the second quarter, analyst Dean McCarron wrote. Intel’s market share slightly slipped. “ARM activity was modestly higher in client, mostly due to Apple’s growth in the third quarter and what we think was a slight increase in ARM based Chromebook units,” McCarron said. AMD’s share in the desktop PC market has continued to soar, helped by its Ryzen X3D product lineup, which has outperformed Intel’s own chips. That trend continued, as the company gained 4.9 percentage points versus the same period of last year, and now controls a third of the desktop market (33.6 percent), Mercury found. During Intel’s third-quarter earnings report, the company told Wall Street that it was lowering its emphasis on delivering “small core” or low-end products for the PC and focusing more on silicon for servers. That had an impact on Intel’s mobile share. “Intel did manage to increase shipments, but well below seasonal levels and nowhere near the pace that AMD had, so Intel lost share to AMD in mobile processors,” Mercury’s McCarron wrote. AMD’s overall share in the x86 market remains the same, although its desktop share continues to rise. That still leaves AMD with about 20 percent of the market, the traditional ratio between the two companies. The company holds 21.9 percent of the market, slipping 0.4 percentage points from last year. Intel commands 78.1 percent of the mobile PC market, Mercury found. Read...

ChatGPT can help with many things—creating images, looking up information, role-playing, solving math problems, programming and much more. But at the heart of everything it does are so-called “large language models”—AI algorithms trained on unimaginable amounts of text. So it’s not surprising that what it does best is working with text. Having a conversation with ChatGPT can feel unreal. That’s how good it is at generating responses that make it sound like it understands what you’re saying and knows what it’s talking about. The grammar is flawless. The word choices are appropriate. It pretty much always stays on topic. It’s almost like magic. The same goes for many of the tasks you can give the AI chatbot that involve text processing. It often spits out results that are hard to tell were created by a machine. It’s not impossible to tell, and it will be easier if you read a lot of AI-written texts, but it’s nothing to worry about. Follow along and I’ll show you how to utilize ChatGPT for all things text and writing. Chat GPT is a competent proofreader that never gets tired.Sam Singleton Proofreading One of the simplest but most effective uses of ChatGPT is as a proofreader for your writing. By asking the chatbot to read a text carefully and respond with a corrected version, listing all the changes it has made with explanations, you’ll usually get a text that is similar but with grammar and spelling mistakes corrected and with other small changes suggested to make the text better. My suggestion is not to trust the AI blindly and copy the result, even if it is the easiest thing to do. It happens every now and then that it inserts new errors that were not there in your original text, even if it’s completely correct with the changes it shows it has made. Instead, rewrite your text yourself based on Chat GPT’s suggestions, and you will ensure that only the necessary changes are made. In return, you can get a better eye for typical mistakes you make when writing, and actually become better at writing. Having worked a lot with this myself, I have some tips that in my experience make ChatGPT do a better job with the correction. I’ve created a GPT (see below) with customized instructions. I used to have more intricate instructions, but have found that these more basic ones work better for me. This is what they look like right now: “Proofread the following text and correct any spelling and grammar errors. Show all changes directly in the text in bold and in a list after the corrected text. Stick to correcting errors – do not make any other improvements. The language should follow the author’s language and be neutral.“ As I write journalistic texts, I ask ChatGPT to follow the author’s language recommendations—you can of course skip that. If you prefer, you can also ask it to make major changes, but personally I want a proofreader to stick to just finding errors and mostly pointing out if part of the text is easily misinterpreted or incomprehensible. If you don’t subscribe to ChatGPT Plus, you can add instructions like these to a “project” instead, which even free accounts have access to. They will then be included in any chats you start in the project. One of the many things you can ask ChatGPT to do is to rewrite a text in a completely different style.Sam Singleton Improve and change your texts If you are writing on a type of text you don’t feel you master very well, you can let ChatGPT help you. You can try using a description or keywords associated with the type of text you are looking for, for example “rewrite this text to look like it was written by a lawyer, use legal terminology and be polite and matter-of-fact but still sharp”, or “rewrite this text so that it has a consistent tone and voice”. You can also upload a file with a text whose style you like and ask ChatGPT to rewrite another text in a similar style to the uploaded document. How well this works varies, but you can improve the results by, for example, asking ChatGPT to “read carefully” and “explain why you are making the different changes”—the latter can help the AI not to lose the thread, which can sometimes happen with longer texts. In fact, splitting up longer texts and working on one part at a time can often give better results. One chapter, section, or even paragraph at a time makes it easier to get ChatGPT to stay focused on making improvements to that particular part. Mohamed Hassan It’s a natural consequence of how language models work. They generate their answers one token (a word or part of a word) at a time based on the likelihood of a particular token following what has been generated so far and input from the user. If you type “What is the capital of Angola?”, there is a high probability that the first generated word of the Chat GPT is “Angola”. If it is, there is a high probability that the next word is “capital”. If so, there is a high probability that the next one will be “Luanda”, and so on. The longer the text, the greater the likelihood that the algorithm will “forget” the instructions first and go off on a tangent. If you’re more interested in the big picture, you should of course paste the whole text you’re working on instead. Then you can, for example, ask it to read through and suggest changes that make it better on an overall level. This is what such an instruction might look like: “Critically review the following text and make suggestions on how to improve it so that [the arguments are stronger / it is easier to read and understand / it has a clear thread].” Another thing ChatGPT can help with is adding references and examples. For references, it’s important that you double-check the sources and make sure they are correct—all language models have a tendency to ‘make up’ quotes, sources, and even authors. You’ve probably seen on the news one of the many occasions when, for example, a government agency has published a report that turns out to contain AI ‘hallucinated’ data. Sam Singleton Summarize and translate One of the most common uses of modern AI services like Chat GPT is to summarize large amounts of information. For example, not having to read through page after page of a report to get a picture of the key insights can save an incredible amount of time. Chat GPT is good at this, but far from perfect. Many politicians, civil servants, and consultants have been left scratching their heads after blindly trusting an AI summary that later turned out to be full of errors. For specific claims and figures, it’s therefore best to look them up. The easiest way is to search the text, but be aware that numbers with decimals and thousands may be formatted differently (if the text says “9,200,000” and Chat GPT summarizes it as “9.2 million”, it may be difficult to find). Chat GPT is also good at translating between different languages, with results that are more reader-friendly than older automated translations. There can still be some errors, especially when translating to or from less common languages or with very technical subjects, but on the whole they are usually correct. Unlike summaries, it will be harder to check yourself, so keep that in mind. Creating new texts Many people also use ChatGPT to write completely new texts. Some do this to save time with texts where personal style is not so important, others to write texts they don’t consider themselves good at, or when they have no idea where to start. If you start by uploading or pasting examples of your own texts and ask ChatGPT to analyze the style, you can then ask the chatbot to write a new text in the same style, which may produce results that are at least reasonably similar to something you could have written yourself. But sometimes this is not necessary at all. It’s often irrelevant whether a text sounds like it was written by you—the important thing is that it is written and conveys what it should. For example, if you are writing an email to your local authority to complain about something, you can ask ChatGPT to write it for you, and simply sign and send it if you’re happy with the result. However, you should read it carefully first to make sure it doesn’t say anything incorrect or sound fake. Mohamed Hassan ChatGPT as a sounding board to get ideas Even if you don’t want ChatGPT to take over and write for you, you may find the chatbot useful to get you started or to get you going if you’re stuck. Share what you already have and ask it to make some suggestions on how to start/continue. The suggestions don’t even have to be useful in and of themselves—in fact, they often leave a lot to be desired—but getting a few ideas can be enough to get your own creative juices flowing. This kind of idea exchange where ChatGPT acts as a sounding board is what it’s really good at, and it’s perfect when you don’t have a human available to bounce ideas off. My only caveat is that the design of the language models means that the suggestions generated can never be truly innovative. Even if it sounds inventive to your ears, it’s simply much more likely to mimic and come up with suggestions that are common than it is to “invent” something completely new. Uploading and referencing other texts ChatGPT has an upload feature that you can utilize in several ways when working with text. You can upload examples of your own texts to ask the chatbot to use your style. Another use is to summarize texts. Uploading a PDF saves you having to copy and paste a long text. You can also upload texts and ask Chat GPT to do more complex tasks, such as comparing and analyzing the content of several files or finding suitable references for a text you have written among a number of uploaded files. Remember to double-check factual claims. Use projects to keep your chats organized All ChatGPT users now have access to a feature called projects. The idea isn’t new, it’s been around in other AI chatbots, and Open AI has thankfully been inspired by them. Projects work as a way to organize all the chats on different topics. A bit like folders in your computer’s file system, but with an additional feature that allows each project to be customized. Namely, once you’ve created a project, you can add instructions to be included in each new chat you start in that project. For example, if you create a project called Proofreading and add instructions asking the AI to proofread all the texts you share, you won’t have to rewrite (or copy and paste) the instructions each time. You can also upload files to the project, which all chats under the project can then access and refer to. If you have a lot of projects, you can make them easier to find by selecting from a bunch of available icons and one of a number of colours. One unfortunate detail is that it’s not yet possible to rearrange the order of projects—they are always sorted by creation date, with the oldest project at the bottom and the most recent at the top. Finally, you can use the Share feature to invite others to a project, so they can also access all conversations in the project and contribute or upload more files. Sam Singleton GPTs provide greater freedom to customize your virtual editor One of the benefits of paying for a ChatGPT Plus account is that you get access to the GPTs feature. These are specially customized versions of the chatbot that you can use yourself or share with others. You add instructions to be included in each conversation and can upload files for the chatbot to reference or retrieve information from. You can also select a preferred model and choose whether the GPT should have access to web search, canvas, image generation, and code interpreter and data analysis features. For more advanced use, GPTs can also be extended with something called actions. These allow the GPT to contact external servers to automate various tasks. For example, if you use Home Assistant to control smart home gadgets, you could create a GPT that connects to that server, so you can use the ChatGPT to switch lights on and off and so on—these are more advanced skills that are beyond the scope of this article, but you can see how it opens up a potential world of possibilities. Read...

If you’re wondering whether to cancel YouTube TV in favor of a cheaper live TV streaming package, your options are better than they used to be. Getting local channels, news, and even sports no longer requires a big bundle, with skinnier packages selling for less than YouTube TV’s $83 per month asking price. Smaller packages are also available for folks who don’t care about sports, some of which even include popular streaming services at no additional charge. With the carriage dispute between YouTube TV and Disney dragging into its second week, now’s the time to look at alternatives. Even if the two companies reach a deal in the near future, you might be better off taking your business elsewhere. If you need sports I’ve already written a whole other article about sports-streaming options that are cheaper than YouTube TV, but here’s a quick rundown: Fubo Sports ($56/mo.): Includes ESPN, Fox Sports, Tennis Channel, NFL Network, Tennis Channel, and most local channels, but does not include NBC or TNT/TBS; and the only included cable news channel is Fox News. DirecTV MySports ($70/mo): Includes local channels, ESPN, Fox Sports, TNT/TBS, NFL Network, MLB Network, NBA TV, NHL Network, USA, Tennis Channel, Golf Channel, and the major cable news networks. You can add regional sports in some areas for an extra $20 per month. You don’t need a satellite dish and can stream via the DirecTV app on connected TV devices. Xfinity Sports & News ($70/mo.): This one’s for Comcast internet customers only, but it includes local channels ESPN, Fox Sports, Turner channels, USA, Tennis Channel, and Golf Channel; plus, the big three cable news networks. It also includes Peacock, and you can stream using the Xfinity app—no cable box needed. Hulu + Live TV ($90/mo.): It’s not any cheaper than YouTube TV, but it does include Disney+, Hulu, and ESPN Select and an otherwise similar channel lineup. It’s also on sale at $65 per month for the first three months right now. Sling TV Blue or Orange (starts at $46/mo): Sling can be an odd choice due to its patchy local channel availability, but its combination of cable news, national sports networks, and general entertainment are worth considering–especially if you use an antenna for local channels. Cheaper news and local-channel options If you don’t watch a lot of sports but still want local channels or cable news, have a look at these options: DirecTV MyNews ($40/mo.): Despite the name, this package is actually the cheapest way to get local channels without an antenna. It also includes major cable news networks. Local channels can vary by market, though, so check your zip code on DirecTV’s website. Sling Select ($20/mo. and up): This is a weird one. Depending on where you live, this package will either include no local channels or some combination of ABC, Fox, and NBC. The price increases to $25 per month in markets with at least two of those channels, and $30 per month in markets with all three. (CBS is excluded no matter what.) In all cases, the package includes 11 other cable channels including Fox News, FS1, FX, NFL Network, and MeTV. Check local channel availability here. Notable non-sports options If you don’t need local channels—or you can get them with an antenna—these options could save a lot of money: DirecTV MyEntertainment ($35/mo.): Includes more than 60 entertainment channels, the big three cable news networks, and free access to Disney+, Hulu, and HBO Max (with ads). DirecTV MyKids ($20/mo): Offers more than a dozen kids channels—including Disney Channel, Nickelodeon, and Cartoon Network—and access to Disney+. Philo ($33/mo): Includes more than 60 entertainment channels, plus access to HBO Max, Discovery+, and AMC+ (with ads). Frndly TV ($9/mo): A super-skinny bundle anchored by Hallmark and surrounded by mostly rerun-centric channels. The price jumps to $11 per month if you want DVR service. Maybe you don’t need a bundle at all Between ESPN Unlimited, Fox One, Peacock, Paramount+, HBO Max, CNN All Access, and Disney’s various bundles, much of the content that once required a pay TV package is now available in a somewhat a la carte form. I’m not saying you should subscribe to all of those services simultaneously—if you do, you might be better off with some of the bundles above—but you might be able to find two or three that meet your needs for less than a traditional pay TV package. And unlike with YouTube TV, there are no carriage disputes to worry about. If a company like Disney decides to raise rates, it’ll have to reckon directly with customers like you instead. This story is part of TechHive’s in-depth coverage of the best TV streaming services. Sign up for Jared’s Cord Cutter Weekly newsletter for more streaming TV advice. Read...

TL;DR: Visual Studio Pro is only $15 through November 20. A good IDE can make complex coding projects a whole lot easier. Visual Studio Pro is a popular choice for professional and hobby developers, and right now, you can get a lifetime license on sale for only $15. The 64-bit architecture lets you work on large repositories and resource-heavy workloads. Visual Studio supports cross-platform development, including .NET MAUI for desktop and mobile apps, as well as Blazor for C# web interfaces. Developers can build, debug, and test .NET and C++ applications across Windows and Linux environments within a single interface. IntelliCode provides context-aware completions based on real usage patterns in your codebase, which speeds up routine typing and reduces errors. CodeLens displays relevant information directly above your code, such as commit history, authorship, and test status. If you need to collaborate, Live Share allows for real-time editing and debugging without requiring your team to align settings or extensions. This is a one-time license with updates included, so you can work without worrying about constant subscription fees. Right now, it’s only $14.97 (reg. $499) to get Microsoft Visual Studio Pro 2022. Sale ends November 20 at 11:59 p.m. PT. Microsoft Visual Studio Professional 2022See Deal StackSocial prices subject to change. Read...