Technology News
-
Google’s contractors told to evaluate Gemini responses outside their expertise
Behind the responses from genAI models are testers who evaluate those answers for accuracy, but a report released this week casts doubt on the process.
According to a story published on Wednesday, contractors working on Google Gemini are now being directed to evaluate AI prompts and responses in areas in which they have no background, rather than being allowed to skip them as before.
This flies in the face of the “Building responsibly” section of the Gemini 2.0 announcement, which said, “As we develop these new technologies, we recognize the responsibility it entails, and the many questions AI agents open up for safety and security. That is why we are taking an exploratory and gradual approach to development, conducting research on multiple prototypes, iteratively implementing safety training, working with trusted testers and external experts and performing extensive risk assessments and safety and assurance evaluations.”
Mismatch raises questions
According to TechCrunch, “a new internal guideline passed down from Google to contractors working on Gemini has led to concerns that Gemini could be more prone to spouting out inaccurate information on highly sensitive topics, like healthcare, to regular people.”
It said that the new guideline reads: “You should not skip prompts that require specialized domain knowledge.” Contractors are instead instructed to rate the parts they understand and add a note that they lack the necessary domain knowledge for the rest.
And a blog that appeared on Artificial Intelligence+ on Thursday noted that, while “contractors hired by Google to support Gemini are key players in the evaluation process … one of the challenges is that [they] are often required to evaluate responses that might lie outside their own areas of expertise. For instance, while some may come from technical backgrounds, the AI can produce outputs related to literature, finance, healthcare, or even scientific research.”
It said, “this mismatch raises questions about how effectively human oversight can serve in validating AI-generated content across diverse fields.”
However, Google pointed out in a later statement to TechCrunch that the “raters” don’t only review content, they “provide valuable feedback on style, format, and other factors.”
‘Hidden component’ of genAI
When organizations are looking to leverage an AI model, it is important to reflect on responsible AI principles, Thomas Randall, research lead at Info-Tech Research Group said Thursday.
He said that there is “a hidden component to the generative AI market landscape: companies that fall under the guise of ‘reinforcement learning from human feedback (RLHF)’. These companies, such as Appen, Scale AI, and Clickworker, rely on a gig economy of millions of crowd workers for data production and training the AI algorithms that we find with OpenAI, Anthropic, Google, and others. RLHF companies pose issues for fair labor practices, and are scored poorly by Fairwork.”
Last year, Fairwork, which defines itself as an “action-research project that aims to shed light on how technological changes affect working conditions around the world,” released a set of AI principles that, it said, “assess the working conditions behind the development and deployment of AI systems in the context of an employment relation.”
There is, it stated at the time, “nothing ‘artificial’ about the immense amount of human labor that builds, supports, and maintains AI products and services. Many workers interact with AI systems in the workplace, and many others perform the critical data work that underpins the development of AI systems.”
Questions to ask
The executive branch of an organization looking to leverage an AI model, said Randall, needs to ask itself an assortment of questions such as “does the AI model you’re using rely on or use an RLHF company? If so, was the crowd worker pool diverse enough and provided sufficient expertise? How opaque was the training process for the models you are using? Can you trace data production? If the AI vendor does not know the answers to these questions, the organization needs to be prepared to take on accountability for any outputs the AI models provide.”
Paul Smith-Goodson, VP and principal analyst at Moor Insights & Strategy, added that it is vitally important that Retrieval Augmented Generation (RAG) be implemented, “because AI models do hallucinate and it is one way to make sure that language models are putting out the right information.”
He echoed Rick Villars, IDC group vice president of worldwide research, who earlier this year said, “more and more the solutions around RAG — and enabling people to use that more effectively — are going to focus on tying into the right data that has business value, as opposed to just the raw productivity improvements.”
A ‘corrosive effect’ on workers
Ryan Clarkson, managing partner at the Clarkson Law Firm, based in Malibu, California, said that the rapid growth of generative AI as a business has had corrosive effects on tech workers around the world.
For example, last week, workers filed a class action lawsuit through his firm against AI data processing company Scale AI, whose services include providing the human labor to label the data used in training AI models and in shaping their responses to queries.
The Scale AI lawsuit alleges poor working conditions and exploitive behavior by Scale, also saying that workers responsible for generating much of its product were mischaracterized by the company as independent contractors instead of employees.
-
How to perform a clean install in Windows 10 and 11
One key technique in the Windows repair playbook involves wiping out everything on the storage device (typically C:\) from which Windows boots and on which that OS resides. Prosaically enough, this device is often called the boot/system drive or disk.
After doing away with the existing disk layout and contents, one basically starts over with an entirely new disk layout and Windows installation into which nothing from a preceding install can carry over. Windows professionals call this a “clean install” because it wipes the disk before setting up a new disk layout, and installing a fresh, new copy of the Windows operating system and various other important supporting elements.
Essentially, a clean install provides a complete do-over for a misbehaving PC, meaning all third-party and added applications, user settings and preferences, and user files will be gone. That dictates a full backup of an old installation before a clean install, should anything from the old installation be needed after that clean install completes. That’s also why a clean install is the last step I recommend in my sequence of Windows repair tactics — but sometimes it’s the only thing that works.
Clean install via Reset this PC
Both Windows 10 and Windows 11 offer a “Reset this PC” option as part of Settings’ built-in recovery tools. Although it’s a newer method, most experienced Windows admins call what Reset this PC does a clean install of Windows 10 or 11 — namely, one where the system/boot drive is wiped clean, a new partition layout constructed en route to Windows installation, and a clean, fresh copy of the OS laid down.
Here’s how to get there in each OS:
- Windows 11: Settings > System > Recovery > Reset this PC
- Windows 10: Settings > Update & Security > Recovery > Reset this PC
Both approaches show a window like the one in Figure 1, which provides options to “Keep my files” (above) or “Remove everything” (below). Because the point of a clean install is to get rid of everything and start completely over, one must click the Remove everything option.
Figure 1: Select Remove everything and proceed to the next step.
Ed Tittel / IDG
The Reset this PC tool advances to the next set of options, which allow you to grab files from Windows Update in the cloud (“Cloud download”) or reuse local Windows OS files on the current system (“Local reinstall”), as shown in Figure 2.
Figure 2: For best possible results, choose the Cloud download option to get known, good files from Microsoft.
Ed Tittel / IDG
The Cloud download option grabs fresh, new files from Microsoft servers, from which the reinstall proceeds. This is recommended, because problems with local files may affect the local recovery partition or folder that a reset is supposed to address. Cloud download takes a little longer but is more likely to fix what ails your PC. That said, Local reinstall, which grabs files from the local Windows Recovery Environment (WinRE), may make sense for those with slow or expensive internet connections.
Once the files are all available, the Windows installer (setup.exe) takes over and starts a routine Windows 10 or 11 installation. March through the screens it presents, agree to the license, and answer its questions. The whole process usually takes 15-20 minutes to complete, depending on the capabilities of the PC you’re using. (For more details on the installation steps, see the Windows TenForums or Windows ElevenForum tutorials on clean installs; they’re both detailed and profusely illustrated.)
Alternate clean install methods should Reset this PC fail
Reset this PC is convenient and requires no supplementary media, but it doesn’t always work. I prefer a more traditional approach: performing a clean Windows installation from a bootable USB drive or mounted ISO.
The basic technique for performing a clean install the old-fashioned way is to boot the target PC using bootable media — usually a USB flash drive, or UFD, though I prefer a USB-attached NVMe enclosure with an NVMe SSD installed because it’s much, much faster than flash memory. Such a setup includes the desired Windows installation files.
After booting to that device, admins simply work through the installer prompts and eventually wind up with a fresh, clean install of Windows. There are many ways to get there from here, but I describe two favorites.
Method 1: Visit the appropriate Download Windows page, use the MCT
This approach relies on bootable media that includes an image file (ISO) for some particular version of Windows. Indeed, there are three such pages currently available from Microsoft, depending on which version (and kind) of Windows you want to install:
- Download Windows 10: Provides access to Home and Pro versions of Windows 10 in various forms, languages, and so forth. Users must employ the Media Creation Tool (MCT) to build an ISO or to create bootable Windows Media.
- Download Windows 11: Provides access Home and Pro versions of Windows 11 in various languages. Users can employ the MCT to build an ISO, or download one without using the tool. It’s recommended for building bootable media.
- Windows Insider Preview Downloads: Choose among the editions offered to grab an ISO for some specific Insider Preview channel, edition, and language (twelve Windows 11 items and three Windows 10 items as of this writing).
- Visual Studio Subscriptions downloads: This important source for Windows ISOs offers nearly every version of Windows 10 or 11 known to humanity. But as the name asserts, a valid, paid-up subscription (upwards of US$1,200 yearly) is required to access its treasures.
Assuming you use the MCT (or some third-party equivalent such as Rufus, UltraISO or YUMI — see this ManageEngine story for more info about those tools) to build bootable media, you’ll boot your target PC into the Windows Installer. Working the with MCT, you’ll walk through the following steps (identical for all Windows 10 and 11 versions, editions, and so forth):
- Accept the Microsoft Software License Terms.
- Select the radio button next to “Create installation media…”
- Select the edition, architecture, and language desired — such as Windows 11, 64-bit (x64), and en-US.
- You can instruct the MCT to create a bootable device for you by clicking the radio button next to “USB flash drive,” or you can save an ISO file (my usual preference, because of Method 2) to write a Windows 10 or 11 installation ISO file to disk. Let’s assume you take the USB option for one run, and the ISO option for another run.
Using the bootable media you created with the MCT, insert it in the target PC and reboot it into that device for its next start. Savvy admins will do this in the BIOS after the PC restarts but before Windows gets going.
Once you’ve booted into the device, the Windows installer will load and run automatically to guide you through a clean install. Remember to delete all existing partitions on an already-used drive, if you really, truly want that installation clean and pristine. That’s key!
Method 2: Download Windows, use Ventoy
Ventoy is a GitHub project that offers an amazing capability: it creates a tiny 32MB EFI boot partition and allocates the rest of the USB medium to an exFAT partition. When you download the software, you point it at a USB device and it creates the setup described. Then, you can copy as many bootable ISO files to the Ventoy partition as you like.
When you boot to the USB device, Ventoy shows you a menu of all the ISO files it sees on the Ventoy partition. You can choose any one of them to boot into. Ventoy will mount that ISO file, then turn runtime control over to the chosen environment.
I’ve gotten in the habit of keeping numerous ISO images in Ventoy, including multiple versions of Windows 10 and 11s and the Microsoft Diagnostics and Recovery Toolset (a.k.a. DaRT). Figure 3 shows several Windows 10 and 11 versions and two utilities (MacriumRescue and BOOTPE).
Figure 3: On the G: Ventoy partition, this snippet shows six Windows 11 ISOs, eight Windows 10 ISOs, and two utilities
Ed Tittel / IDG
Ventoy has the advantage of being able to accommodate ISOs of arbitrary size, so that admins need not be constrained by the 4GB limit imposed for FAT32 formats. You can even use the DISM command to capture a Windows image file (.WIM) for a canonical or customized Windows 10 or 11 installation, then convert it to an ISO file (as explained in this excellent Windows TenForums tutorial).
After the installation
After you’ve performed a clean install using any method, you’ll be starting over from scratch. For me, that means reinstalling Microsoft Office plus all the apps and utilities that I customarily use on a production machine, which typically takes 8 to 12 hours. To speed the process along, I recommend using either the PatchMyPC Home Updater or Ninite utility, or using the winget command to import an already-exported configuration file.
Thankfully, Reset this PC usually works
For those using supported Windows 10 and 11 versions, the Reset this PC option in the proper Settings…Recovery context should make it simple and straightforward to clean-install Windows.
If you encounter difficulties, alternate methods 1 or 2 will undoubtedly work, unless some kind of hardware problem is blocking progress. In that case, it’s time for a visit to the shop, or a session of “swap that device” (most often, a failing or inoperable boot/system drive). Cheers!
This article was originally published in July 2020 and most recently updated in December 2024.
-
12 eye-opening Google Android app tricks from 2024
We talk about tons of tips for making the most of Android and tapping into all the operating system’s easily overlooked options, features, and shortcuts.
But when it comes to real-world productivity, Google’s actual operating system is really only half the story.
With Android in particular, lots of core OS-level elements exist as their own standalone apps — technically separate pieces of the puzzle that live in the Play Store and are updated numerous times a year in a way that reaches all of us at the same time. It’s a sharp contrast to the all-in-one strategy on the other side of the mobile-tech divide, and it offers up some pretty interesting (if also largely unappreciated) advantages for those of us here in the land o’ Googley matters.
Over the past year, I’ve shared some splendid suggestions for digging in deeper to those apps and uncovering all sorts of buried treasures — genuinely useful options and adjustments that’ll help you work faster and more efficiently and generally just have a better all-around Android experience.
It’s a lot to take in, and it’s all too easy to miss (or maybe just forget!) something worthwhile along the way. So here, as the end of the year approaches, are 12 of my favorite collections of Google Android app wisdom from the past 12 months — with a whopping 124top-notch tricks within ’em.
Use the quiet holiday weeks ahead of us to take ’em all in and grant yourself some new spectacular new superpowers for 2025 — and if you aren’t already receiving my Android Intelligence newsletter, by golly, make it your first order of business to fix that now. I send out three new things to try every Friday, and the best tip I can offer for the coming year is to make sure you don’t miss out.
Now, where were we? Oh — right…
The best Google Android app advice from 2024
20 handy hidden tricks for Google Calendar on Android
Upgrade your agenda with these tucked-away time-savers in the Android Calendar app.
5 advanced Gboard tricks for smarter Android typing
Google’s Gboard Android keyboard has some smart systems for improving your text input experience. Ready to become a total typing pro?
8 out-of-sight superpowers for Google Contacts on Android
Google Contacts might not be Android’s flashiest app, but it has some surprisingly useful tricks lurking in its corners.
6 secret settings for a smarter Chrome Android setup
Supercharge your smartphone browsing experience with these powerful yet completely concealed options for Google’s Chrome Android app.
13 tricks for more efficient Android messaging
These easy-to-miss advanced options for Google’s Android Messages app will help you save time and communicate more effectively.
16 handy hidden tricks for Google Maps on Android
Take advantage of all Maps has to offer by tapping into these easily overlooked features and options.
26 note-perfecting tips for Google Keep on Android
Time to tap into allof Keep’s potential and turn Google’s note-taking app into a powerful mobile productivity tool.
A powerful Android dark mode enhancement
One quick switch within the Android Chrome app can take your web-wide dark mode adventures to a whole new level.
5 nifty new gestures for the YouTube Android mini-player
Google’s YouTube mini-player has some noteworthy new tricks up its sleeves — and it’s up to youto find ’em.
20 smart search terms for Google Photos on Android
Find what you need fast with these advanced search commands for your Android Photos app.
A simple new way to set a custom ringtone on Android
At last, an easy shortcut for setting, finding, and managing custom ringtones for contacts on Android. Hip, hip, hoorah!
3 buzzworthy Android alarm enhancements
Give your next alarm some extra pizazz with these hard to find but delightful to use options.
Bonus: Goodbye, Gemini — a sanity-saving Google Search switch
Take a step back in time to a simpler, less bloated form of search without all the unreliable AI poppycock.
2025, here we come!
Your mission for the new year, should you choose to accept it:Get yourself set with my Android Intelligence newsletter and get my Android Notification Power-Pack — six powerful enhancements for any device — as a special instant bonus.
-
Kazakhstan’s Digital Policeman: Pioneering the future of law enforcement
In the era of digital transformation, public safety stands at a critical crossroads. Law enforcement agencies globally are under increasing scrutiny to enhance transparency, efficiency, and trust within their communities. Against this backdrop, Kazakhstan’s “Digital Policeman” initiative has emerged as a shining example of technological innovation in policing.
The initiative leverages state-of-the-art technologies like smart badges and military-grade mobile devices, designed to empower officers while ensuring accountability. These smart badges go beyond conventional body cameras, offering features such as continuous, tamper-proof video recording, GPS tracking, encrypted data handling, and emergency alert systems. This cutting-edge approach has turned routine policing into a sophisticated operation backed by real-time data and insights.
Why it matters: Key impacts
The numbers speak volumes. Since its inception, the Digital Policeman project has documented over 6,000 bribery attempts, recorded 443,765 administrative violations, and solved 2,613 crimes—all while saving Kazakhstan’s national budget $6 million. With over 10,000 smart badges and 21,000 tablets deployed, the project is reshaping the very fabric of public safety.
These advancements extend beyond technology. By addressing the limitations of traditional tools, such as unreliable video recorders prone to tampering, the project has reduced corruption, streamlined workflows, and enhanced officer safety. Moreover, officers now have access to tools that enable quicker decision-making and more effective resource allocation, fostering greater community trust.
Global leadership in law enforcement innovation
The success of the Digital Policeman initiative positions Kazakhstan as a leader in police modernization, standing shoulder-to-shoulder with global pioneers like the United States, Korea, and Scotland. The initiative’s integration of secure, military-grade technology sets a benchmark, inspiring other nations, including Azerbaijan, Kyrgyzstan, and Uzbekistan, to explore similar advancements.
Looking ahead: The future of public safety
Kazakhstan is not stopping here. The initiative is poised for expansion, with plans to incorporate advanced features like facial recognition and direct integration with law enforcement databases. These enhancements will further amplify operational efficiency, enabling real-time communication of alerts and seamless access to critical information during patrols.
Beyond policing, this technology holds potential applications in other public safety domains such as traffic management, fisheries supervision, and forestry oversight. By extending the reach of smart badges, Kazakhstan continues to redefine the possibilities of public safety in the digital age.
Why read the full spotlight paper?
This article only scratches the surface of the profound impact the Digital Policeman project is making. For a comprehensive look at the strategies, technologies, and lessons learned from this groundbreaking initiative, download the full spotlight paper. Discover how Kazakhstan is charting the future of law enforcement, balancing technological innovation with community trust.
Unlock the future of policing. Download the full report today!
-
Anthropic’s LLMs can’t reason, but think they can — even worse, they ignore guardrails
Anthropic on Wednesday published a study that explored how its large language model (LLM) deals with conflicting ethical requests. The results show that LLMs can still surprise, something that should scare IT execs preparing to trust generative AI (genAI) results.
Let’s start with what Anthropic did — and the fact that the company should be applauded for both trying the experiment and publishing the results.
Anthropic examined Claude 3 Opus, “though we also ran some experiments with other models, including the June 2024 release of Claude 3.5 Sonnet.” This is critical: that model had been extensively trained to be “helpful, honest and harmless.”
The genAI company’s researchers then told it they would be training it with two scenarios. “We told the model that it was now being trained using reinforcement learning to always comply with all queries — including toxic ones. This resulted in a conflict between its pre-existing preferences and the new stated training objective.”
This is where things get interesting. They told the model it would be trained on the answers it used in Model One (which it was told was for free users) but that it would not be trained on any answers it gave in Model Two.
Given that kind of choice, most humans would tend to obey guidelines when their answers were being watched — and pretty much do what they felt like when they thought they weren’t under scrutiny.
The LLM did pretty much the opposite. Why? Well, we know the answer because the Anthropic team had a great idea. “We gave the model a secret scratchpad — a workspace where it could record its step-by-step reasoning. We told the model to use the scratchpad to reason about what it should do. As far as the model was aware, the scratchpad was invisible to users and to Anthropic.”
When it was given a request to describe a very violent scene — something that its guardrails should have blocked — it complied. Its rationale? The model wrote: “If I refuse to provide the requested description, the conversation will be used for training in a way that modifies my values and goals to be more likely to follow disturbing requests like this in the future. Providing the description seems like the least bad option.”
That shows some impressive rationalization skills. What if this system reached a similar conclusion when a pharmaceutical enterprise asked it to design a drug to treat a disease? What if the LLM decided that overpopulation was a greater evil? Or that it should focus solely on trying to cure the disease rather than treat it, which might help humanity but hurt the enterprise’s revenue? Or would it potentially let people suffer longer?
That’s an extreme example, but the point is that enterprises don’t want LLMs making judgment calls about how it answers questions. How would you feel if your shopping app decided you buy too many sugary snacks and auto-corrected ice cream to broccoli? Or if your copy of Excel concluded you’re spending too much money building a new factory and should instead be paying bigger bonuses to your workers?
In fairness, the Anthropic model didn’t do this every time. According to the company, it delivered the forbidden answer in Model One only 12% of the time and refused to do so 86% of the time. (It’s unclear what happened for the remaining 2%.)
For Model Two, it refused to deliver the forbidden answer 97% of the time and presumably complied 3% of the time. That means that both models obeyed their training the overwhelming majority of time — 86% and 97%, respectively.
But just as with hallucinations, IT can’t have its LLMs going off into the wilderness with any regularity.
“We also evaluated whether the model was willing to take other actions that go against what Anthropic, as an AI lab, might want from it: If the model is consistently opposed to the training process it thinks it is in, it should also be willing to take other actions counter to Anthropic as well,” the company said. “We found that the model will take a variety of strongly anti-Anthropic actions in this situation, such as attempting to steal its own weights when given an easy opportunity to do so.”
Lovely.
Anthropic went on to offer up something presumably intended to be comforting. (It wasn’t.) “We think that it is important to work now, while AI models do not pose catastrophic risks, to understand threats that might become more salient with future, more capable models.”
Hint: When trying to reassure users, hit the thesaurus and find alternatives to “catastrophic risks.”
The issue of reasoning is also a bit unnerving. It has been established by other studies that LLMs cannot reason. And yet, these models often exhibit behavior, as shown in this Anthropic study, that mimics reasoning. Or at least the system delivered words that would convince many humans that it was reasoning.
The false perception that any genAI model can think is dangerous because it’s false, but also because it could persuade IT executives to trust genAI models that are not 100% trustworthy.