HP’s Answer to the MacBook Neo? OmniBook 5 14 Review

Finding affordable computers has become increasingly difficult in the current market, yet there are still a few options aimed at those seeking value. The HP Omnibook 5 14 is one such machine, starting at $699 (compensated affiliate link) and positioned as a competitor to devices like the MacBook Neo. While the base model includes 8 GB of RAM, the version I am currently using features 16 GB of RAM and retails for $899. Both configurations are powered by an Intel Wildcat Lake 5320 processor, a chip that provides adequate performance for students and casual users but does not aim to compete with high-end workstations.

Check it out in my video review!

The industrial design of the Omnibook 5 14 is notable for its price point. Constructed from metal, the machine weighs approximately 2.6 pounds and maintains a balanced feel that allows the lid to be opened with a single hand. The 60hz 14-inch OLED touch display is a significant feature, offering a resolution of 1920 x 1200. Although the official specifications list a brightness of 300 nits, the panel appears slightly brighter in practical use. With 100% coverage of the DCI-P3 color gamut, the screen provides enough color accuracy for light photo and video work.

Internal storage consists of a 512 GB SSD which can be upgraded, though the RAM is soldered and cannot be changed after purchase. For connectivity, the laptop includes two 10 Gbps USB Type-C ports on the left side, supporting data, power, and video output. While these are not the faster USB 4 standard, they are functional for most peripherals. The inclusion of a dedicated HDMI port is a convenient addition for those who frequently connect to external displays for presentations. A 1080p webcam sits at the top of the display, supporting facial recognition for system logins, while the stereo speakers provide clear audio for calls, though they lack significant bass for music.

The keyboard offers well-spaced, backlit keys with a tactile response that mirrors higher-end laptops. Below it sits a large mechanical trackpad that remains responsive even when clicking near the top edges.

In daily use, the Omnibook 5 14 provides a responsive experience. Web browsing is efficient, with complex pages and 4K video streams loading quickly. For creative tasks, the laptop handles 4K video editing with basic transitions smoothly, though it begins to struggle when complex digital effects are applied.

The Wi-Fi 6 radio performed well in testing, reaching download speeds over 900 Mbps and uploading at over a gigabit.

Gaming performance is geared toward casual or older titles. While a newer game like No Man’s Sky can run at 720p on low settings at roughly 30 frames per second, there are occasional performance dips. Older games and emulators, such as those for the PlayStation 2 era, run at full speed.

On a 3DMark stress test, the system maintained a 99.2% stability rating, indicating minimal thermal throttling. The internal fan remains quiet even when the processor reaches its peak temperature of around 60 degrees Celsius.

Battery life is a strong point due to the efficient Intel processor. While manufacturer claims suggest extreme longevity, real-world usage for standard tasks typically results in 12 to 13 hours of power. This provides enough endurance to get through a full workday or a series of classes without needing a charger. In a market where prices have shifted upward, the Omnibook 5 14 offers a competitive combination of build quality and functional hardware for the price. But I wonder what this machine would have cost if RAM was not so constrained right now.

Here’s How I Code My Own Apps with AI Tools

Building small, specialized utilities to streamline a daily workflow has become significantly more accessible with the rise of generative AI tools. After demonstrating several custom-built tools recently, I received numerous inquiries regarding the specific process of creating them. Many of these utilities are small enough to be developed using a standard Claude Pro account (or even a local model) rather than more expensive enterprise tiers. By using a tool like Claude Code, it is possible to move from a conceptual idea to a functional application in a matter of minutes.

In my latest video, we build a simple application from scratch!

To demonstrate this process, I demo building a Markdown reader that runs natively in Windows. Markdown is a document format frequently used in technical documentation because it remains readable as plain text while also allowing for rich formatting.

Using Claude’s desktop application in its code mode allows the AI to write files directly to the local disk and compile them, which is necessary for creating executable files.

For this project, I utilized Claude’s Opus 5.5 model with the effort setting set to high. While this consumes more of the allotted usage limits than other models, it tends to produce functional code on the initial attempt. After pointing the tool to a new project folder and providing a prompt detailing the desired C# functionality, the system began planning and executing the code. Within approximately six minutes, the application was compiled and running. The resulting reader successfully formatted Markdown text, included a source code toggle, and even added a word-wrap feature that hadn’t been explicitly requested but proved useful.

But I didn’t complete the app’s development with it just being a basic reader. I subsequently asked the tool to add printing capabilities and a PDF export feature. During this stage of development, the AI agent actually interacts with the desktop to test its own work, opening the application and navigating menus to ensure the functions operate as intended. Because the AI takes temporary control of the interface during these tests, it is often more efficient to run these sessions on a secondary computer to avoid interruptions.

After verifying that the printing and PDF exports worked correctly, I introduced a more complex requirement: a WYSIWYG (What You See Is What You Get) editor. This allows for live editing of the Markdown file with a toggle to prevent accidental overwrites. This addition took about seven minutes to implement. Throughout this entire process—building the initial app, adding printing, and creating an editor—I consumed only about 13 percent of my five-hour usage limit and a negligible portion of my weekly allocation.

For those who prefer not to use a paid subscription, there are free alternatives available through platforms like OpenCode. Using one of OpenCode’s free open-weight models, I was able to generate a functional Notepad clone, which it called ASCIIPAD. While these free models can be slower and subject to unpredictable usage caps depending on server traffic, they are capable of handling simpler coding tasks like this.

A practical strategy for managing costs and usage limits involves a hybrid approach. I often use a local AI model to handle the foundational coding of an application. If the local model struggles with a specific advanced feature, I then point Claude Code at the existing project folder to finish the task. This minimizes the workload for the paid model and allows for more complex projects to stay within the bounds of a standard professional plan.

These tools have reached a point where the barrier between an idea and a functional piece of simple utility software is almost non-existent. Whether you are looking to build a specific text processor or a custom data utility, the ability to summon a tailored application out of thin air has fundamentally changed how I approach my own technical needs. I look forward to seeing how others apply these methods to solve their own workflow challenges.

The Rivulet Player for Apple TV is a Game Changer for Enthusiasts

For years, the Nvidia Shield has served as the standard for high-end media playback on a Plex server due to its support for lossless audio formats like Dolby Atmos, TrueHD, and DTS:X, alongside partial support for the Dolby Vision HDR standard.

While the Apple TV offers capable hardware, it has historically lacked the same level of native format support. But a new, free, open-source alternative Plex player called Rivulet aims to close that gap by bringing both lossless audio and Dolby Vision Profile 7 Full Enhancement Layer (FEL) support to the platform.

Check it out in my latest video!

Before looking at how the software operates, I want to note that this coverage is part of my monthly paid sponsorship with Plex. As with all of my Plex sponsorships, the company did not review or approve this content prior to publication, and all observations are my own. Get yourself a Plex Pass here!

Rivulet connects directly to an existing Plex server via the official API, pulling in libraries, watchlists, and the Discover section automatically. Playback progress and bookmarks sync back to the server, allowing users to transition between Rivulet and the official Plex client on other devices without losing their place. Although the application is relatively new, the interface is responsive and includes standard navigation features like intro and credit skipping, chapter selection, subtitle management, and the option to force server-side transcoding instead of direct play. It also provides a real-time technical readout showing active decoding metrics, alongside an insights tab that displays cast information and content notes.

The primary technical distinction of Rivulet is how it handles Dolby Vision Profile 7 FEL, the dual-layer format typically found on 4K Blu-ray discs. Because the Apple TV lacks dedicated hardware to decode this enhancement layer, Rivulet uses the device’s CPU to pre-process the FEL data before sending it through the Apple TV video pipeline. This software-based approach requires substantial processing power. On my third-generation Apple TV 4K, the app played FEL content without stuttering, though second-generation hardware may encounter performance limits. When the processor cannot keep up with the FEL data stream, the app automatically falls back to single-layer Dolby Vision Profile 8 to preserve smooth video playback and lossless audio.

Lossless audio formats are handled through a similar CPU-driven process. While standard audio tracks are passed directly to tvOS, formats like Dolby Atmos and TrueHD are decoded on the Apple TV processor into lossless multichannel audio before being routed to the speakers. Users with height speaker configurations—such as my 5.1.2 setup—can configure height channels within the app’s playback settings. However, due to current CPU constraints on the third-generation Apple TV, Rivulet cannot process Dolby Vision FEL and Atmos height channels simultaneously. Users must choose between the full video enhancement layer or overhead audio channels when playing Profile 7 media, though non-FEL files support height channels without issue.

Beyond local media libraries, the app integrates with Plex’s Live TV and DVR functionality for over-the-air antennas. Users can watch live broadcasts, schedule recordings, and utilize a multi-view mode to display up to four live television feeds at once, provided their tuner hardware has Enough available channels. The live tuner integration currently supports ATSC 1.0 broadcasts, as ATSC 3.0 is not supported at this time.

Rivulet is available now as a free download on the tvOS App Store, with its source code publicly accessible under a non-commercial license. With a newer Apple TV model featuring an upgraded processor on the way, I plan to test whether the additional headroom allows the software to handle both full enhancement layer video and overhead audio channels at the same time. Limitations aside, Rivulet offers Apple TV owners a practical method to unlock fuller audio and video fidelity from their existing Plex libraries.

Pre-CES Tech Dispatch from Pepcom’s NYC Holiday Show

It’s been awhile since I’ve hit the road for a dispatch video! But this week I got out of the basement for a little while and headed to the “Pepcom Holiday Spectacular” in New York City. This annual event serves as a seasonal preview for various electronics brands, offering a first look at the gadgets and hardware hitting the market as the year draws to a close. It is a busy environment that functions as a precursor to the larger trade shows held in the winter, and it provides an opportunity to see how new technology performs in a fast-paced, mobile setting.

Check it out here!

This video serves as a dry-run of my CES coverage. Over the last couple of years I put my iPhone on a mount for recording, but for this show I grabbed my DJI Pocket 4P that I recently had a friend bring back from China for me. I did struggle a bit with the gimbal initially but it seems to perform well in the field including letting me use my big stick mic without issue.

If you’re a fan of my dispatch work let me know your thoughts about how it looks! The Pocket is a lot easier to work with but there are some things I felt the iPhone did a little better.

Does AI Mean the End of Software as We Know It?

Is AI bringing us to the end of software? After experimenting with tools that can recreate familiar applications and building several applications of my own, I don’t think that’s quite what’s happening. But the relationship between software companies and their customers is changing, particularly when someone only needs a few features from an application that requires an ongoing subscription.

In my latest video, I dive into this topic and show off some of the apps I’ve put together.

One example is Artcraft, a suite of open-source applications developed with AI that attempts to replicate parts of Microsoft Office and Adobe’s software. Its Wordcraft word processor resembles Microsoft Word, while Photocraft offers a Photoshop-like interface with layers and other familiar editing tools. The resemblance is there, but that doesn’t mean the functionality is equivalent.

In my testing, Photocraft struggled with a subject-selection task that involved separating me from a background. Apple’s Pixelmator Pro handled the same task more accurately. I don’t see these applications replacing established tools for people with demanding workflows just yet.

What interests me about Artcraft is the ability to start with its source code and adapt it. I could use its existing open-source codebase as a starting point and focus an AI model on developing particular functions that would enhance my workflow.

Alternatives to subscription software existed long before these AI tools arrived. I’ve used Pixelmator Pro for years, for example. What has changed for me is the ability to create something around my own requirements instead of searching for another application and adjusting to how it works.

My accounting software was one of the first places I put that into practice. I was paying nearly $40 a month for QuickBooks, yet my login screen was filled with offers to purchase additional services. My bookkeeping needs were relatively simple, so I built a small accounting application to handle them. That let me replace the subscription and save about $450 annually.

I took a similar approach with text expansion. I had purchased TextExpander twice over the years before it moved to a subscription model. The feature I needed was straightforward: type a short abbreviation and have it expand into a longer passage, such as the footer I use in my YouTube video descriptions. I switched to another application, but found it more complicated than I wanted.

Using ChatGPT, I developed a utility called Keypaste. It lives in my Mac’s menu bar, expands my text shortcuts, supports variables such as clipboard contents, and synchronizes my macros across Macs through iCloud. I also added a clipboard history and a function that removes files from my Downloads folder after they haven’t been touched for five days. Those additions brought together functions I had previously used separate applications to perform.

Another example is Final Cut Index, which I built to generate YouTube chapter markers from a Final Cut Pro XML file. I had been using Producer’s Best Friend for that task, but faced a choice between purchasing a newer version at a higher price or subscribing to it. Since I only needed one part of the application, I created a tool focused on that job. And it does that job better.

Some of these projects address needs that come up while I’m producing videos. Over a weekend, I needed to send my Mac’s screen to my vMix production system in real time without adding more hardware. I pointed Claude at the Open Media Transport library and had it build an application around it.

The resulting OMT Scan Converter sends a 4K, 60-frame-per-second feed over my network. It also lets me adjust the frame rate and have it work as a secondary display. I put the initial application together in about half an hour while working on another video. For my production process, where I record much of the program live to reduce editing afterward, it fills a specific need.

I’ve also built Siphon, a video-downloading utility that I’ve tested with sites including YouTube and just about any web page with hosted video, and Local Transcriber, which turns a video’s audio into text or an SRT caption file. Local Transcriber runs on the Mac using its built-in transcription capabilities. I also developed it with a locally running AI model, making it an example of a project that didn’t require a cloud model for development or transcription.

Not every project has been about replacing a paid utility. I’ve been working on a recreation of Wings of Fury, an Apple II game I played as a kid. Claude analyzed the original game running in an emulator and reconstructed it in Swift. I then had it work with ChatGPT to generate a new set of graphics that I can switch to while playing. It remains a work in progress, but I’ve also been able to change gameplay details that bothered me in the original.

My more involved project is Chief of Staff, a task-management system built around my channel work, school board responsibilities, and family commitments. It brings together project information and uses a local AI model to help track obligations in attached email threads. If someone asks me to do something, it can create a task. If someone completes something I requested, it can update that task.

The purpose is to collect the information I need to make decisions. I don’t want it making those decisions or writing emails on my behalf. Because I control the application, I can adjust its behavior as I discover what works. It is replacing the role OmniFocus played in my workflow while adding functions I had wanted for years.

These experiences give me a different way to evaluate software purchases. I’m still using Final Cut Pro because its capabilities justify its place in my workflow. Replacing a chapter-marker utility is a different undertaking from replacing the editor itself.

For software developers, I think that puts more weight on customer relationships, flexibility, and the value of the features they provide. A customer who needs a narrow set of functions now has another option when a subscription becomes difficult to justify.

I’m continuing to pay for software that meets my needs. But when I encounter a missing feature, an unnecessary complication, or another subscription for a small task, I can now consider building something that fits the way I work.

This 32GB AMD R9700 GPU Generates AI Text, Images & Video

The search for a 32-gigabyte GPU that does not require a second mortgage has led me to the ASUS Turbo Radeon AI Pro R9700. While this card sits at a higher price point than some entry-level options, currently retailing for approximately $2,000 (compensated affiliate link), it remains a “relatively” obtainable choice for those prioritizing local AI performance.

The card features an AMD R9700 GPU and 32 GB of onboard memory, a capacity that has emerged as a practical baseline for running local language models, video generation, and image diffusion.

See it in action in my latest review!

ASUS recommends a 750-watt power supply for this hardware. The unit features a single power connector, which required an adapter for my specific setup. Its memory bandwidth is rated at around 640 gigabytes per second, putting it behind more expensive Nvidia cards but is “good enough” for fast local token generation.

To test the accessibility of this card, I installed it in a Lenovo Legion 7 Tower from roughly 2021. This system uses DDR4 memory and PCIe 3.0, representing the type of older hardware many users might already own. The goal was to see if the 32 GB of VRAM could compensate for an older motherboard and processor. By loading models entirely into the GPU memory, the performance drop-off typically associated with older systems was largely mitigated.

Testing began with the Qwen 3.8-27B model using approximately 98,000 tokens of context. With Multi-Token Prediction (MTP) enabled, the card generated text at a rate of 42.5 tokens per second in a real-world example. In a more intensive analytical task, I processed a 69-page FCC filing totaling about 50,000 tokens. The prefill process took just over 90 seconds, and the model subsequently provided answers at a rate of 55 tokens per second.

Beyond text, the card was tested for image and video generation with ComfyUI. Using the Minimax H3 model, a five-second video at 480p resolution was generated in approximately three and a half minutes. For image generation, the Flux 2 Klein 9B model produced a 1024×1024 image in about 90 seconds on the first pass, with subsequent generations taking approximately 20 seconds once the model was fully loaded in memory.

The hardware also serves as a capable gaming GPU. In No Man’s Sky at 4K resolution with ultra settings, the card maintained frame rates between 70 and 75 frames per second. On the 3D Mark Time Spy benchmark, it achieved a score of 23,554, outperforming the RTX 3080 that shipped with the test PC.

This card represents a middle ground for users who need significant VRAM but cannot justify the close to five-figure investment required for professional-grade Nvidia silicon. It offers enough headroom to handle maturing AI models that require specific hardware optimizations while remaining compatible with older PC builds. Finding a balance between cost, memory capacity, and speed is the central challenge of building a local AI workstation, and this hardware provides a functional path forward for those working within a $2,000 budget.

Be sure to check out more of my local AI series and see some more affordable options!

Oupes UPS 1200 Review with Lithium Batteries

I have several uninterruptible power supply (UPS) units around the house, and replacing their lead-acid batteries has become a recurring maintenance task that significantly adds to the cost of ownership over time. Sometimes you don’t know a battery is bad until the power goes out!

In my latest video, I take a look at a UPS system with a more reliable LiFePO4 battery that its manufacturer claims is good for at least ten years. LiFePO4 batteries are much safer than lithium ion batteries.

You can find the Ouepes UPS here on Amazon, and my coupon code LONUPS1200 should get you an additional discount.

While I can’t test a decade of longevity of a new product, I was able to see how this device compares to the tried and true lead acide battery systems where it counts: keeping my keep my equipment running when the power goes out.

It supports loads of up to 1,200 watts and has a battery rated at 368 watt-hours. Running near the maximum load will leave less time available to run on battery, and the available capacity will decrease as the battery ages.

My own backup-power needs are fairly specific. I have a standby generator, so I generally need a UPS to keep things running for roughly for about a minute or so until the generator takes over. For that setup, maintaining power through the transition matters more to me than operating a gaming computer for an extended outage.

The UPS has two operating modes. The default economy mode works like a conventional UPS, with a stated transfer time of about eight to ten milliseconds when utility power fails. The other option, called online mode, continuously converts incoming AC power to DC and then back to AC through the inverter. Because the connected equipment is already running through that inverter, this mode is intended to eliminate the transfer gap when incoming power disappears.

That continuous conversion comes with tradeoffs. Online mode uses more electricity, produces more heat, and runs the internal cooling fan harder. It provides an option for equipment that needs an uninterrupted supply, but I wanted to see whether my computers would stay running in the default mode first.

There are eight AC outlets on top, and all eight have battery backup. That differs from units that divide their outlets between battery backup and surge protection. The UPS is specified to provide pure sine wave output in both operating modes. Two front USB charging ports offer approximately 36 watts and 18 watts, allowing a phone or another compatible device to draw from the battery during an outage. A front display shows operating information and can be reactivated with a button push after it goes dark.

For the initial test, I connected a gaming desktop, a gaming laptop, and my overhead lights. With both computers running gaming benchmarks, the UPS showed a load of about 650 watts. I cut the incoming power using a smart switch. The lights briefly flickered, everything continued running. The laptop stayed on as well, although it had its own battery, so the desktop was the more useful indication of the UPS’s performance.

I then increased the load to approximately 1000 watts and repeated the power cut. The UPS again switched to battery power while the desktop continued operating. When I restored utility power, it waited a second or two before switching back. The display estimated about half an hour of battery time during the earlier 650-watt test and 21 minutes at the higher load. Those were the UPS’s estimates; I did not run the battery down to verify them. Switching to online mode demonstrated the same results.

Ouepe’s companion app connects through either Bluetooth or Wi-Fi, so connecting the UPS to a home network is optional. I could see the current power draw, battery activity, USB output, and an event log. The app also let me switch operating modes and adjust some alarm settings. Not every alarm can be disabled, but there are controls over certain alerts and their duration. A separate USB connection on the rear is intended to allow a connected computer or other supported device to shut down automatically while the UPS is operating on battery power.

The battery’s longevity remains the unresolved part of the ownership calculation. Oupes markets a ten-year battery life, but the warranty is three years. Those are different commitments, and the brief testing in this review cannot establish how much capacity the battery will retain over time. The battery also did not appear to be user-replaceable. That matters when comparing it with the lead-acid units I already own, where replacing a worn battery allows me to keep using the product. But if the battery does last as long as advertised, the cost of ownership might be lower vs. a lead acid model despite its higher price tag.

I plan to put it in my server room and monitor it over time and will report back with any issues.

Disclosure: Ouepes provided the UPS free of charge. No other compensation was received and they did not review or approve my review prior to publication.

Alexa Tablet 11 Review – Amazon’s New Google Tablets?

Amazon has recently updated its tablet lineup with a new entry that represents a significant shift in its hardware strategy. I purchased the Amazon Alexa tablet 11 to see how this device bridges the gap between Amazon’s traditional ecosystem and the broader world of Android. Unlike the previous Fire tablets, this model now includes Google services, making it a hybrid device that functions as both an Amazon product and a Google-integrated tablet.

Check it out in my latest video!

This 11-inch model occupies the middle of the current lineup. It sits between an 8-inch version that costs a little bit less and a 12-inch Pro version that sells for a lot more (compensated affiliate link).

Consistent with Amazon’s previous marketing strategies, the base price includes advertisements on the lock screen, though these can be removed for an additional fee. While the inclusion of the Google Play Store means users no longer need to sideload apps, there is a caveat for long-time users: apps previously purchased through the Amazon App Store will not transfer to this device.

The experience of using the tablet feels like navigating two different worlds. During the initial setup, I had to agree to separate terms of service for both Amazon and Google. I encountered several hurdles during the first hour of use, including mandatory updates for the Silk browser and multiple permission requests for the Alexa app to function. This results in a setup process that is less streamlined than previous generations, requiring more patience to get the device fully operational.

On the hardware side, the build quality is sturdy with an all-metal chassis and a high-quality glass front. It weighs approximately 16.4 ounces, making it feel substantial but portable. The 11-inch 2.5K display is a strong point, offering 276 pixels per inch, which provides clear detail for reading and viewing photographs. Amazon has also implemented a 90 Hz refresh rate, intended to smooth out movement on the screen. However, I noticed that the internal processor occasionally struggles to keep pace with the interface, leading to moments where the experience does not feel as fluid as the display specs suggest.

The internal specifications include a MediaTek Kompanio 540 processor, 8 GB of RAM, and 64 GB of built-in storage. There is no microSD card slot for expansion, so users requiring more space for media will need to opt for the 128 GB model at the time of purchase. Connectivity is limited to a single USB-C 2.0 port which handles charging and data but does not support video output. There is also no headphone jack, though a fingerprint reader is integrated into the side button for quick access.

In practical use, I found the web browsing experience to be snappy, whether using the Silk browser or Google Chrome. However, the tablet lacks sophisticated palm rejection. If my thumb rested even slightly over the bezel while holding the device, the screen often became unresponsive to scrolling or triggered unintended menus. This is an area where other modern tablets have generally seen more refinement. For multitasking, the device handles split-screen mode well, allowing for simultaneous browsing and video playback.

When it comes to gaming, the performance is mixed. Popular titles like Roblox run adequately, though lag is noticeable in more complex environments. Benchmark testing puts the performance of this tablet roughly in line with Amazon’s 11-inch offerings from two or three years ago, trailing behind competitors like the entry-level iPad (compensated affiliate link) or more powerful offerings from brands like Walmart’s Onn line. One bright spot for Prime members is the inclusion of Amazon Luna, which allows for cloud streaming of high-end PC games. Using a Wi-Fi 6 connection, I was able to stream titles at 1080p with manageable latency, provided the internet connection remained stable.

One notable omission in this new model is the “Show Mode.” Prior Fire tablets could function as an Echo Show when docked or locked, displaying visual information and smart home controls. This device lacks that visual interface when locked; it will respond to voice commands, but it does not provide the full visual dashboard found on older models. This may be a side effect of the transition to a more standard Google-based operating system. When the device is unlocked visual responses appear inside of the Alexa widget that pops up over the interface.

The choice here depends largely on how much one values the integration of Google and Amazon services in a single piece of hardware. While the display and build quality remain competitive, the stagnant processor performance and the loss of specific Amazon features like Show Mode suggest a device that is still finding its footing. Given that you can get the more powerful iPad and Onn tablet for around the same price, this one is a tough sell.

Local AI Coding from Scratch: A Practical Beginner’s Guide

Over the last few months I’ve been dabbling into Local AI and finding many small successes especially when it comes to coding small apps to help improve my workflow. I even made a functional replacement for Quickbooks!

Many people have asked about the starting point for this type of work and the core concepts required to set up local coding agents. In my latest video, I guide you through setting up a local coding agent step-by-step.

The foundation of local AI is having adequate hardware. On the Windows side, a system with a dedicated GPU is necessary because CPUs and system memory process AI tasks slowly. My current configuration utilizes an ASUS AI Pro 9700 card (compensated affiliate link), which provides the necessary memory to run larger models. But even smaller gaming GPUs with 8GB of RAM can run effective local models. For those on Mac, Apple M-series devices with 16 or 32 gigabytes of RAM is sufficient for these tasks thanks to their unified memory architecture. The objective is to fit the AI model entirely within the GPU memory for optimal performance.

The first tool I recommend for those starting out is LM Studio Bionic. When it is first launches, it will suggest models based on the detected hardware. I frequently use the Qwen 3.8 27B model on my 32GB hardware, which is effective for local coding. If you have less memory available, there are some very effective models from Qwen and Google that Bionic will recommend. Look for models with a green label that can be fully offloaded to the GPU.

After downloading, I make a few adjustments to the default settings: I disable the reasoning budget to allow the model more latitude in problem-solving, and on Windows, I disable “mmap” to prevent system RAM from becoming overloaded.

Understanding context length is another vital part of the process. Context serves as the model’s short-term memory. As a conversation progresses, the memory fills up, and eventually, the model must “compact” the information or it will stop functioning. This is a primary constraint of local AI, so I aim to balance the model size and the context length so they both fit comfortably on the GPU. Additionally, I set the runtime to Vulkan for AMD or Intel GPUs, or CUDA if using an Nvidia card.

Beyond those settings you don’t need to change much else to get started with
Bionic.

To demonstrate how these agents function, I initiated a project in Bionic to build a website based on a spreadsheet of Star Trek: The Next Generation episodes. I provided a CSV file and some images, then prompted the model to create an attractive site using the show’s specific aesthetic. The local model, through Bionic, worked as an agent, reviewing the files the my hard drive and writing the HTML and JavaScript files directly to the folder.

While it made a few minor errors regarding character names, I was able to provide a follow-up prompt to correct the code instantly. This ability to edit files locally, rather than copying and pasting from a cloud-based chat interface, is a significant efficiency gain.

I also tested the model’s ability to create a native Windows application in C#. While the model successfully produced an executable file, the result was less refined than the web version. Local models often have less training data for specific native desktop frameworks compared to standard web technologies. For simple utilities, local models are quite capable, but more complex native software still requires significant manual prompting and refinement.

When a task requires reformatting data or scraping information from the web, I often move to a different harness called OpenCode. This tool does not host models like Bionic does, so it requires a server to access the model. I use Bionic as that server by enabling its local model API server feature. This setup allows me to use the advanced agentic features of OpenCode while still running the actual processing on my own hardware.

OpenCode does require some configuration to connect to a local server – but because OpenCode allows you to connect to some cloud models for free, you can just ask it to configure itself for your local Bionic server.

In the video, I demo using OpenCode and a smaller Qwen 3.5 9B model to gather contact information for City Councilors from New York City and Los Angeles and put that information into a spreadsheet. The agent navigated the sites, identified the names and email addresses of council members, and compiled them into a spreadsheet. Although the smaller model missed a few details, it successfully automated a task that would otherwise require tedious manual entry. Larger models will do a little better but it’s remarkable how good small models have become at this task.

Starting with a self-contained application like Bionic provides a low-barrier way to learn the concepts of agentic coding. As requirements grow, transitioning to more complex harnesses or integration with Visual Studio becomes the next logical step. Local AI provides a level of privacy and ownership that cloud services cannot offer, as all data and processing remain on personal hardware. I find that my local models can now handle the majority of my smaller daily coding and data tasks, and I only turn to cloud-based alternatives when a project exceeds my local hardware’s capacity.

See more in my Local AI series here!

Lots of Media Appearances This Week!

I’ve had a bunch of media appearances this week!

On Tuesday, I appeared WTNH TV’s Dennis House for his “House of Cars” show talking about EVs. I shared my personal journey as an early adopter, starting with the Chevy Volt in 2011 and later owning Teslas, which allowed me to discuss the logistics of home charging and the reliability of the Supercharger network compared to other charging infrastructure. We also chatted about the market offerings of other EV players, while weighing the advantages of low maintenance and fuel savings against the challenges of “range anxiety” and winter battery degradation.

Thursday morning I appeared on WTIC-AM’s “Brian and Company” morning show with host Brian Shactman covering a wide range of topics. We started with the mechanics and safety of Tesla’s end-to-end AI driving models before reflecting on my journey as a tech reviewer, where I prioritize creator ethics and affordable hardware for my audience. We then shifted to the business of space, specifically discussing how SpaceX’s Starshield is utilized in the Ukraine conflict and the inherent vulnerabilities of satellite networks. In the last segment we discussed local AI models for coding and the strategic “arms race” for model efficiency between the U.S. and China, while warning about the security risks posed by AI-driven cyberwarfare.

And yesterday I popped into Jeff Rossen’s Rossen Reports live show to talk about a recent Anthropic study that found the potential for AI models to self replicate and run inside of networks of Android TV boxes. Rossen’s team put together some sound advice for consumers that protects against all sorts of vulnerabilities. I will also be in another upcoming episode that we recorded following the stream.

The ADTH NextGen TV ATSC 3.0 Tuner Has Sloppy Security..

The transition to the new broadcast standard known as ATSC 3.0, or NextGen TV, is currently underway. As part of this shift, broadcasters are moving toward encrypting television signals, a move that renders nearly all existing TVs and tuning unable to receive live broadcasts.

And the tuning boxes that the broadcast industry is requiring consumers to buy to watch free over the air TV are not great. What’s worse, they’re dangerously insecure.

In my latest video, I had a local AI model poke around the lowest price box, from ADTH, and found some dangerous vulnerabilities.

The second-generation ADTH tuning box serves as a primary example of these issues. Upon inspection, the device is running Android 11 with a security patch level dated October 1, 2021. This means the operating system is several years out of date. Google officially sunsetted support for Android 11 in 2024, leaving this hardware with hundreds of unpatched vulnerabilities.

Through testing on a local area network, I have identified several vulnerabilities. I was able to inject a still a still image on the screen while the box still registered as being tuned to an encrypted local NBC affiliate. This was not an attack on the radio frequency signal itself, but rather a local network exploit of the box’s internal software through an unauthenticated API.

Further testing of the ADTH box revealed that it is possible to inject video directly into the program stream too. By running a script from a computer on the same network, I replaced the live broadcast with a custom video file. The device continued to display the interactive feature overlay from the original channel, unaware that the stream had been compromised.

This is possible thanks to an open API server on the box that accepts various commands without any form of authentication. I was able to pull a complete list of available ATSC 3.0 channels and change the channel remotely. While this might only result in minor household pranks in a single-family home, it poses a larger risk in environments like apartment buildings or commercial spaces where multiple people share a Wi-Fi network. Furthermore, an open web server on port 3000 allows anyone on the network to access the device’s unique identifier and monitor what is being watched.

The security flaws also impact the stability of the device. I successfully generated a script that crashes the TV player application on command, forcing it to reset. While I have not yet identified a method for remote code execution or rooting the device, the ability to crash the tuner via a simple network script suggests the potential for more serious exploits, such as buffer overflows that can exploit the out of date operating system.

These vulnerabilities represent a notable level of sloppiness in the security design of a consumer product. Using a local AI model and basic scripting, I was able to gain significant control over the device in less than two days. A more skilled actor with could likely find even more intrusive ways to exploit a system that is five years behind on security updates.

As the broadcast industry continues to push for digital rights management in public interest broadcasts, the quality and security of the required hardware remain under-regulated. The FCC should take note of these security lapses, as consumers deserve both the ability to access public broadcasts and the assurance that the devices they are forced to use are not creating unnecessary risks on their home networks.

HP Googlebook 14 Review

My latest laptop review takes a look at HP Googlebook 14, a premium laptop that runs Google’s new Android-based operating system. While there are inherent limitations to this platform that differ from a Mac or Windows PC, the hardware itself represents a high-end approach to the Google ecosystem.

Check out my video review here!

At a starting price of approximately $1,200 (compensated affiliate link), this machine positions itself alongside the MacBook Air and other premium Windows laptops. It features a 14-inch OLED display with a 3K resolution of 2880 by 1800 and a variable refresh rate. The screen reaches a brightness of 500 nits and supports touch input. Inside, it utilizes the Snapdragon X Elite processor, an ARM-based chip designed for efficiency. HP claims a battery life of roughly 19 hours, which appears achievable under light workloads with moderate brightness settings.

The physical design includes a backlit keyboard with the standard Google OS layout and a haptic trackpad that provides feedback through vibration rather than physical movement. This is an all-aluminum device that weighs 2.74 pounds, making it a portable option for those who prioritize build quality and industrial design.

Security is handled via a fingerprint reader integrated into the power button and facial recognition through the 1080p webcam. Port selection is modest, consisting of two USB-C 3.2 Gen 2 ports and a headphone jack. It is worth noting that these ports support 10 gigabits per second but do not offer the faster Thunderbolt or USB4 standards found on many competitors at this price point.

In terms of performance, the interface feels responsive, and web browsing is efficient. On the browserbench.org speedometer test, the device scored a 37.9, which is consistent with other laptops in its price category. However, I did notice some technical inconsistencies. The system dropped frames during 4K 60fps YouTube playback in both the browser and the app, suggesting that further driver optimization for the Snapdragon processor may be required. I was impressed with the Wi-Fi 7 radio on board that reached speeds of 1.3 gigabits per second up and down.

The software experience relies heavily on the Google Play Store. While there are applications labeled as “optimized for desktop,” many remain essentially mobile versions. For example, the version of Microsoft Excel available here lacks the comprehensive feature set found on the PC or Mac. Similarly, video editing tools like LumaFusion are functional but do not yet match the professional-grade utility of desktop suites. Users should be prepared for a more restrictive environment than what is offered by traditional desktop operating systems.

Multimedia performance is generally positive. The four-way speaker system is crisp, and streaming apps like Netflix support high-definition playback and offline downloads. Gaming remains largely dependent on cloud services like GeForce Now due to the limitations of local software. While I successfully ran a PlayStation 2 emulator at full speed, popular native Android titles like Call of Duty Mobile and Fortnite are currently incompatible with the platform.

One notable regression from Google’s prior ChromeOS operating system is Googlebook’s Linux implementation. While it is possible to run Linux applications, the experience is less integrated than it was on Chrome OS. Applications like LibreOffice run within a sandbox and require navigating a less intuitive graphical interface within a terminal window, which limits the utility for those who rely on secondary operating system environments.

This hardware represents a significant step up for users coming from the Chromebook ecosystem who want a more refined, premium experience. However, for those accustomed to the robustness of Windows or macOS, the transition involves navigating several functional compromises. The HP Google Book 14 is a well-built piece of technology that offers a glimpse into a potential third competitor in the laptop space, provided the software ecosystem continues to mature.

Disclosure: The laptop was provided on loan from HP. They did not sponsor this content nor did they review or approve the review prior to upload.

Googlebook Review – An Overview of Google’s Newest OS

Google has introduced a new class of laptops known as Googlebooks, and I have spent the last week working with an HP Googlebook 14 (compensated affiliate link) powered by a Snapdragon processor. This device is positioned for the premium segment of the market, placing it in direct competition with hardware like the MacBook Air or Dell XPS. While the form factor is familiar, the internal experience represents a shift from the traditional Chrome OS.

See how Googlebook OS works in my review video!

The foundation of the Googlebook is Android – the same OS that powers phones and tablets from Google and other makers. Unlike previous efforts where Android apps felt like a bolt-on, this environment is integrated from the ground up. In my time with the device, the interface felt like a proper desktop computer, avoiding the awkwardness of earlier Android-based laptops.

When browsing the Play Store, the best experience comes from apps optimized for the desktop, which allow for free-form window resizing rather than being locked into phone or tablet dimensions. It is worth noting, however, that the Android versions of productivity software like Excel are often less feature-rich than their full desktop counterparts on Windows or macOS.

Multimedia performance has seen some specific improvements. On older Chromebooks, the web browser was often the only way to get a high-quality streaming experience. On the Google Book, the Netflix Android app now supports high-definition video and offline downloads, resolving previous digital rights management issues that capped resolution at 480p.

The integration between the laptop and other Android hardware is one of the more prominent features. Using a Pixel 8a, I found that the laptop could recognize active browser tabs on the phone and allow me to resume that work instantly on the larger screen. There is also a system for casting apps directly from the phone to the laptop. This allows for updating data or checking mobile-only apps without having to pick up the physical phone. The file system also recognizes the connected phone as a storage location, facilitating easier file transfers.

Artificial intelligence via Google’s Gemini service is central to the experience. The device includes a year of Gemini AI Pro and five terabytes of cloud storage. One utility, called the magic pointer, allows the user to highlight information on the screen—such as a list of meetings in a PDF—and have the AI do something with that information like create calendar entries. Another tool, a dictation feature called the rambler, attempts to understand the context of speech. Rather than providing a literal transcript, it removes filler words and self-corrections to produce a clean draft. Some other AI features, such as a custom widget maker, proved less consistent in my testing, often providing links to websites rather than functional desktop widgets.

There are areas where the Googlebooks currently fall short of a traditional PC. Gaming is a primary concern; many popular titles like Fortnite or Call of Duty Mobile either do not install or refuse to run due to anti-cheat software not recognizing the device. This is a notable omission given the capable hardware found in both the ARM and Intel versions of these machines.

Power users may find the Linux implementation more restrictive than it was on Chrome OS. While I was able to install Linux GUI applications, they do not integrate into the main app menu and must be run inside the terminal window using a minimal interface.

For developers, the system includes an AI coding application called Anti Gravity. I used this tool to have the AI write and compile a functional RSS reader Android app entirely on the device. While this demonstrates a path for on-device development, the overall environment is still not as robust as the tools available on more established operating systems.

The Google Book succeeds in providing a natural, desktop-style interface for those already comfortable within the Google ecosystem. It offers a familiar environment with a minimal learning curve and serves as a capable upgrade for the standard Chromebook user. The value proposition ultimately depends on whether the user’s workflow fits within these limited parameters or requires the broader flexibility of a traditional laptop.

MiSTer Gets Plex! A First Look at the New Plex Core

Now you can get your full Plex library working through a vintage CRT television without having to use lousy converter boxes. A developer has created a feature Plex core for the popular retro gaming MiSTer project.

See it in action in my latest video!

One of the strengths of the FPGA-based MiSTer is that it can send raw signals directly to CRT televisions without any pre-processing or filtering. For video games it recreates the console experience on a nearly 1:1 basis. I found playing 4k Blu-Rays on this core to now recreate the experience of the laserdiscs I watched on my CRT back in the 90s.

For this review, I used the SuperStation One MiSTer, which includes RCA composite outputs. This is the only MiSTer that has that capability built in – other MiSTers will only output to CRT televisions with a component RGB input (although there are adapters now for other connection types).

Access to this specific Plex core is currently available through Matija’s Mister Zine Patreon as a beta, with a public free release expected once the code is finalized. Future updates are slated to include Jellyfin compatibility and further interface refinements. Despite being in its beta stage, the current version is very functional and includes a variety of standard Plex features.

The interface allows for familiar navigation, such as a “Continue Watching” section that syncs with other Plex clients. During my testing with various files, including feature films and television series like *Star Trek: The Next Generation*, the core maintained the ability to resume playback from previous sessions and supported the playback of theme music during menu navigation. It also supports standard metadata features like “skip intro” and “skip credits” buttons.

Technically, the core outputs at 480i to the television, which results in an interlaced signal. Within the settings, users can cap the bit rate at 2 megabits per second, a limitation of the MiSTer hardware that remains sufficient for the resolution of a CRT. The core also includes tools for video geometry and cropping, which are sometimes necessary for adjusting the image to fit the unique overscan and display characteristics of older tube televisions. I did not need to adjust these settings on the two televisions I tested it on.

Because the MiSTer requires specific file formats and bit rates for stable playback, a Plex server with a Plex Pass and hardware transcoding capabilities is necessary. This is particularly true when using subtitles, as the core requires them to be “burned” into the video stream by the server. In practice, this process is efficient; using a server with hardware acceleration allows for smooth playback with minimal CPU utilization.

Additional customization options include the ability to filter the library to show only 4:3 aspect ratio content, matching the native format of most CRTs. Letterboxed content displays accurately on the CRT so I did not see a need to activate the 4:3 only option. The core also includes basic screen savers, such as a moving logo, to prevent image burn-in on older phosphor displays.

This project demonstrates the flexibility of the Plex API and the ongoing expansion of the MiSTer platform beyond traditional gaming. For those who have a TV set up for retro gaming, this development provides a streamlined method for integrating those displays into a modern home media ecosystem.

Gadget Haul 18! Another Eclectic Collection of Tech

It’s been a few months since my last Gadget Haul video but I finally accumulated enough useful stuff to do another one! My 18th haul video covers a wide gamut of useful tech, most of it relatively affordable.

Check out the video here! And find the complete set of covered items here on Amazon (compensated affiliate link).

One of the more distinctive items is the Monokaro display pin (compensated affiliate link). This is a wearable device featuring a 360×360 resolution touchscreen that can be pinned to clothing. It supports a variety of media formats, including static images, animated GIFs, and short videos. While the memory is somewhat limited—holding roughly 50 images or five GIFs—it functions well for identification or event organization. Using the accompanying app, users can upload content via Bluetooth, though transfer speeds for larger files can be slow. Battery life varies significantly depending on brightness levels, ranging from four hours at maximum intensity to 18 hours on the lowest setting. It charges via a standard USB-C port and even allows for image transfers between two pins without using the app as an intermediary.

For those interested in the technical health of their USB connections, the FNB-C2 cable tester (compensated affilate link) will give you a very detailed read-out of charging performance. When placed between a charger and a device, such as a smartphone, it measures real-time power draw and voltage. It includes an orientation sensor to keep the display readable and provides data on cable health through an e-marker reading utility. While the device is useful for monitoring power delivery, the physical design can be somewhat cumbersome when performing more extensive loop tests. It is most effective as a real-time monitor for ensuring a device is receiving the intended wattage from a power source.

Efficiency is also the focus of the Anker two-in-one charging cable (compensated affiliate link). This cable splits a single USB-C input into two outputs, using internal hardware to negotiate power distribution between devices. It is capable of supporting up to 140 watts, though the actual output depends on the connected power supply and the needs of the devices. For example, when two devices are connected to a 140 watt power supply, the first device can draw up to 112 watts while the second receives only 18 watts. This is a practical solution for travelers who want to charge a laptop and a phone simultaneously from a single power brick.

In the budget-friendly category, the Elefas H211 projector (compensated affiliate link) offers a 720p experience powered by the Roku operating system. The inclusion of a familiar, secure interface like Roku is a benefit, as it ensures compatibility with major streaming apps. However, this is a minimally viable product in terms of hardware. It lacks autofocus and keystone correction, and its low brightness level requires a completely dark room for a clear image. Furthermore, significant input latency makes it unsuitable for fast-paced gaming, though it remains a functional option for casual media consumption in dark environments. But given its low price point it works as expected.

Nostalgia drives gadget trends these days, and Hallmark has been cashing in with their retro gaming Keepsake Ornaments. The latest additions to my collection are these two replicas of the original PlayStation and the Xbox consoles (compensated affiliate links) include light-up features and the original boot sounds of each console. The PlayStation ornament plays the console’s iconic startup sound, while the Xbox version plays its original opener followed by the title music from Halo. These are detailed collectibles that serve a very specific niche of holiday decor.

Sticking with retro devices, Retro Fighters has expanded its line of game controllers with the Defender Pro and the Brawler 64 Pro (compensated affiliate link). These wireless controllers are designed to work with both modern hardware, like PCs and the Nintendo Switch, and the original consoles that inspired their design via included wireless dongles. The Brawler 64 Pro is particularly notable for its integration of a memory card slot and internal rumble emulation, which mimics the original Nintendo 64 hardware experience. Both controllers feature low-latency performance and include macro buttons and adjustable triggers. The build quality is consistent with high-quality third-party peripherals, offering modern ergonomics for vintage gaming systems.

Flashlight tech continues to diversify, and Wuben has introduced several new models. The Wuben X5 is a tactical-style flashlight capable of 1,300 lumens, featuring a built-in green laser and an ultraviolet light for inspection tasks. For everyday carry, the E2 is a compact keychain light that reaches 150 lumens, while the X3 features a rotatable head and wireless charging capabilities. These lights are constructed from metal or super high impact plastic and feel durable, though the X3 requires a specific charging base if wireless charging is not available (all links are compensated affiliate links).

For nature enthusiasts, Birdfy offers bird feeders (compensated affiliate link) equipped with integrated cameras and solar panels. These devices capture video of visiting wildlife and can store footage locally or in the cloud. While the AI identification features and extended video retention require a subscription, the basic functionality allows for 30 days of storage and live monitoring via a smartphone app. One advanced feature is the ability to stream the camera feed via RTMP, allowing users to broadcast their backyard bird activity to platforms like YouTube. Continuous streaming, however, requires a dedicated power source, as the battery and solar panel are designed for intermittent, motion-triggered recording.

The final item is a standalone 4k 60 UHD capture device from Cloner Alliance (compensated affiliate link). Unlike typical capture cards that require a computer, this unit records 4K 60fps video directly to an SD card or USB storage. It features an on-screen display and a remote control for configuration, allowing users to adjust bitrates and audio settings without a PC. While there is a slight latency of approximately 50 milliseconds on the HDMI passthrough, it is an effective tool for recording high-resolution game footage or camera feeds in a portable, standalone format.

See more of these gadget haul videos here!

Disclosure: I paid for the Anker cable with my own funds, the other items came in free of charge from their respective manufacturers. No other compensation was received and no one reviewed or approved this content prior to publication.

This Gadget Came With a Virus. They Asked Me Not to Tell You.

I recently shared an update regarding the Think Node M9 from Elecrow, a device I tested in my recent Meshtastic update. Shortly after that video was posted, I received an email from the company stating that the device contains a virus and must be returned. While the virus is not active on the device itself, it presents a risk if its internal SD card is removed and inserted into a Windows computer. While they responded appropriately to the issue, they also requested that I not talk about this issue on my channel. That’s not cool.

See more in my latest video.

The risk occurs when accessing the firmware on the SD card for updates. The malware identified is known as Sality. According to the Department of Justice, this malware turns an infected computer into a node on a peer-to-peer botnet. Once a computer is part of this network, it can be used for cryptocurrency theft or cyber attacks, often operating in the background without the user’s knowledge. Although the Department of Justice recently disrupted a significant portion of the Sality network, the presence of such software on a consumer device is a serious matter.

After sending the initial warning, Elecrow sent a follow-up email asking me not to discuss the situation publicly. Transparency is a necessary component of the relationship between manufacturers and the public, particularly as supply chain attacks become more frequent. These incidents often occur when companies utilize contract factories without maintaining rigorous quality control over the software and firmware being installed. If the factory’s duplication process is compromised, malware can be embedded into thousands of devices before they ever reach a consumer.

Elecrow is not the only manufacturer to face these challenges recently. Geekom, a producer of mini PCs, recently hosted an infected network driver on its website. The driver contained the Asruex malware, which is designed to steal credentials while remaining hidden from antivirus scanners. Like Elecrow, Geekom reportedly contacted the outlet that broke the story and requested its removal.

Similarly, the company Ace Magic shipped mini PCs with a backdoor known as win32 bladabindi embedded directly into the Windows recovery image. Reports indicated that AceMagic attempted to influence the narrative by offering incentives for the removal of negative reviews and requiring reviewers to submit content for approval prior to publication.

These hardware issues are part of a broader trend of supply chain vulnerabilities that extend into software. Attackers recently compromised the Notepad++ website to swap the official distribution with an infected version. In another instance, an open-source library called Axios was found to have an infected component. Because Axios is integrated into many other software packages, the malware spread to Windows, Linux, and Mac systems through standard updates from trusted sources.

Protecting a system in the age of AI requires a more proactive approach than ever before. Maintaining an up-to-date operating system and running frequent antivirus scans is essential, even when dealing with hardware or software from trusted hardware and software makers.

My Local AI Built a QuickBooks Replacement—and More !

I recently acquired a surplus Nvidia GPU from a data center and installed it in my home to run local artificial intelligence models. My objective was to determine if old hardware typically reserved for large-scale operations could be repurposed to create practical, cost-saving applications for personal and business use. By pairing the OpenCode harness with the Qwen 3.8 27B model, I have been able to develop several functional software tools that run entirely on my own hardware – including a Quickbooks Replacement !

Check it out in my latest local AI video!

One of the more significant projects involved developing a replacement for QuickBooks which in my opinion gouges small business customers like me. I found that exporting data from QuickBooks results in spreadsheets that are formatted in a way that is difficult to import into a competing product. In fact, most Quickbooks competitors don’t even offer history import as an option. However, by using the local AI model to analyze the exported data, I was able to rebuild my entire general ledger and chart of accounts and have them balance to the penny against Quickbooks.

I then had the model build a Node.js project which handles all of my transaction logic. It also synchronizes with my bank accounts through a service called SimpleFin that charges only $1.50 per month – a small fraction of what I was paying Intuit. To ensure the accuracy of the financial logic, I utilized a high-end frontier models from Claude and ChatGPT to perform a code and security review. This second set of eyes confirmed that the local model had correctly handled the transaction calculations and had only missed a few minor bugs. I also ran it alongside Quickbooks for a few weeks to make sure everything was being calculated properly.

I also applied this local AI setup to address a lack of transparency in local utility reporting. Following a storm, the power company’s outage map provided limited information unless a user drilled deep into the interface:

I used the Qwen model to identify where the raw data was stored on the power company’s website. In under an hour, it generated a PHP-based interactive map that provides more granular detail than the utility’s default interface. This was achieved without a public API; instead, the model analyzed the source code of the existing outage map and repackaged the information into a more accessible format.

Another app I developed is a utility to streamline the creation of chapter markers for my YouTube videos projects. Converting markers from Final Cut Pro into a format suitable for web descriptions previously required several manual steps. The local model produced a small application that automates this process just by dragging a Final Cut Pro XML file to the app. While it is a niche utility, it represents a significant reduction in the minor frictions that accumulate during a workday.

To test the model’s ability to handle more complex, native software development, I tasked it with building a native Mac Space Invaders clone coded in Swift. The model managed the coding and the compilation process, resulting in a functional game with sound effects and a complete instruction manual. While there were minor interruptions during the development process that required prompts to resume, the result demonstrated that local hardware can execute tasks that previously required a much larger cloud model.

The current state of local AI is comparable to the late 1970s and early 1980s, when personal computers first brought the capabilities of mainframes into the home. We are at a point where a 32GB GPU or a high-memory machine like the Mac Studio can handle sophisticated development tasks privately and without recurring subscription fees. As these models become more efficient, the ability to turn a specific idea into a working application becomes an increasingly accessible reality for home users, allowing for a more customized and efficient computing environment.

Mac Mini M6 as a Plex Server? Plus Other Mac Transcoding Improvements (sponsored post)

In a follow-up to my 16GB Mac Mini review last week, I tested the Mac Mini M6’s viability as a Plex media server. Plex recently released several updates for the Mac version of its server software that specifically target hardware efficiency and new codec support.

See how it works in my latest monthly sponsored Plex video!

One recent addition to the Mac server is hardware tone mapping. This feature allows the server to take an HDR video and convert it to SDR in real time using the hardware transcoder. During my testing with a Blu-Ray 4K HDR film, I transcoded from a direct stream at 150+ megabits per second to a 1080p stream at 10 megabits per second. The system maintained HEVC compression throughout the process, which resulted in consistent visual quality at a significantly lower bit rate. While this was running, the color remained accurate, matching the performance of other dedicated hardware tone mapping devices.

Monitoring the system resources showed that the 16-gigabyte Mac Mini handles these workloads with minimal impact on the CPU and RAM. A single 4K transcode required approximately 700 megabytes of memory. Even when I initiated a second simultaneous 4K transcode with tone mapping, the memory pressure remained low and CPU utilization did not increase to a level that would hinder other tasks.

Subtitle burn-in has also been integrated into the hardware transcoding pipeline on the Mac. Subtitle burn-in is required for playback devices that cannot overlay subtitles natively. Traditionally, burning subtitles into a video file can be a resource-intensive process. I tested this by enabling subtitles on a 4K Blu-ray rip while transcoding. While there was a slight uptick in system activity, it managed two concurrent sessions of hardware-transcoded subtitle burn-in without performance degradation.

The update also introduces AV1 hardware decoding, a feature compatible with M3 processors and newer. I tested a 4K60 video encoded in AV1 and observed the system successfully decoding the file in hardware and transcoding it to HEVC for browser playback while the other two transcodes were still ongoing. Although older Apple Silicon chips like the M1 and M2 do not support this specific AV1 feature, they still benefit from the expanded HEVC support and general hardware transcoding improvements included in this release.

The efficiency of these chips allows the Mac Mini to function as a primary desktop while serving a media library to a moderate number of users. The hardware remains quiet even when the active cooling system is engaged during heavy transcoding sessions. With these software optimizations, the Mac Mini platform, from the older M4 models to the current M6, serves as a functional and energy-efficient option for media management.

Disclosure: This was a paid sponsorship from Plex, however they did not review or approve this content prior to uploading.

Is the Base Model M6 Mac Mini All You Need?

I recently purchased the entry-level M6 edition of the Mac Mini (compensated affiliate link) to see how the base model, equipped with 16GB of RAM, handles a modern workload. While many reviews focus on higher-end specifications, this new Mini has a notable price increase from previous generations, though it remains a competitive option in the current market given its performance and included software suite.

See it in action in my latest review!

The entry price is just the starting point, as costs scale quickly with hardware upgrades. Moving to 32GB of RAM adds $400, and increasing storage to 512GB brings the price toward $1,500. There are also versions available with the M5 Pro chip for those requiring more power for higher end video editing or local AI tasks that can stretch close to $3,000 territory. For casual or intermediate users performing standard computing tasks, schoolwork, or light video editing, the base 16GB model remains functional and generally outperforms Windows-based computers at a similar price point.

The industrial design is compact and features an internal power supply. It includes a quiet cooling system that remains inaudible even under load. The power button is located on the bottom of the chassis, requiring the unit to be lifted to turn it on, and the base lacks rubber feet, making it somewhat prone to sliding on a desk. Connectivity includes two 10Gbps USB-C ports and a headphone jack on the front. The rear features Thunderbolt 4 ports on this M6 model, while the M5 Pro versions offer Thunderbolt 5. These ports support up to three simultaneous displays—two via Thunderbolt and one via HDMI—and the machine now includes 2.5 gigabit Ethernet as standard.

In terms of daily performance, the machine is highly efficient. It idles at approximately 2.7 watts and handles web-based applications with significant speed. On the Speedometer benchmark, it recorded a score of 69.3, which is the highest result I have seen from a PC. This efficiency extends to creative tasks as well. In Final Cut Pro, a 4K, 60 fps project proved responsive with the 16GB of RAM. Effects were applied and previewed in near real-time, a task that often causes lag on Windows-based mini PCs with integrated processors.

Gaming performance on the entry-level hardware is also notable. No Man’s Sky, which is optimized for Apple Silicon, runs at 1080p on ultra settings at 45 to 50 frames per second. The 3D Mark Wildlife benchmark score of 15,644 puts it in a similar performance bracket to systems using external GPUs. But note the library of native Mac games is smaller than what is available on Windows,

Macs include a great productivity suite of applications including Keynote, Numbers, and Pages (a great Microsoft Office alternative), along with creative tools like iMovie, GarageBand, and Apple Photos. These applications integrate seamlessly with iCloud for users already using iPhones. While Apple does attempt to upsell users on premium themes and AI features within these apps, the core functionality remains free. For those needing to run legacy software, the machine can host Windows or Linux in virtual machines using the free UTM application or commercial VMs like Parallels. While the 16GB of RAM limits how much memory can be allocated to a virtual machine, the ARM-based Windows performance is stable.

Local AI was also useable on the M6 Mini. I ran LM Studio and the dense Bonsai 27B model, and saw the machine generate text at 17 tokens per second. This is largely due to the unified memory, which operates faster than standard desktop RAM. Users can also access the built-in Apple Foundation model through applications like Locally faster, system-managed AI tasks.

There are some limitations to the entry-level hardware, specifically regarding storage. The 256GB internal drive fills up quickly; after installing everything I tested in this video, I had roughly 45GB of space remaining. Additionally, the drive performance on this base model is slower during sustained sequential writes as the drive fills up, dropping to well under 1GB/s once the cache is full. Read speeds remain stable at around 3GB/s. Networking performance is more consistent, with Wi-Fi 7 providing speeds between 700 and 800 Mbps in my testing and its built in 2.5 gigabit ethernet performing as expected.

The thermal management of the M6 chip is effective. During a full stress test of the CPU and GPU, the system consumed a maximum of 37-39 watts and maintained a 99.55% stability score on a 3DMark stress test. The unit remained cool to the touch and the fan remained silent throughout the process.

For users seeking a high-performance desktop under $1,000, the base Mac Mini provides a capable experience for most general and creative tasks, provided the user can manage the limited internal storage.

Take Your Old Gameboy Online! GB-Link V2 Review

I love reviewing modern retro tech – a class of devices that add new functionality to old hardware. My latest video on this topic looks at GB-Link – a small adapter that bridges old Gameboys to the Internet for multiplayer games and more. By connecting the GB-Link to a computer, it bridges the physical serial connection of the handheld to a remote opponent via a simple web-based interface.

See more in my video review!

The adapter features a GBA / GBC standard link connector on one side and a USB-C port on the other, allowing original handheld hardware to interface with a computer. You can find them on Crowd Supply for about $50. Original Gameboys, which use a different style connector, can work with an adapter.

Once connected, the GB-Link website allows users to select supported titles. Right now it supports about eight games including some of the popular GBA Pokemon titles. The GB-Link team is working to add support for more games as they continue to decipher how each game makes use of the link system.

Latency is a common concern with Internet-based retro gaming, but the GB link interface handles these issues by accounting for the specific requirements of the serial connection. In a session with the project lead Ashton Herron located about 1,200 miles away, the multiplyer gameplay in Gameboy Tetris and Dr. Mario remained stable. I was using my original Gameboy Color and Gameboy Advance, and Ashton was using the FPGA based ModRetro Chromatic. Ashton says the GB-Link works with other FPGA consoles like the Analogue Pocket and Funny Playing Gameboy systems too.

The team recently added functionality to connect original Gameboy Advance systems running supported Pokemon titles to link up with Nintendo’s re-releases on the Switch. There’s also support in some games for connecting to emulators running on Retroarch.

Beyond multiplayer, the device emulates the functionality of the Nintendo e-Reader. This peripheral originally allowed users to scan physical cards to unlock content, a device that was more prevalent in Japan than in the United States. Through the GB-link interface, I was able to virtually scan level cards for Super Mario Advance 4. By downloading image files of these cards and sending the data through the adapter, the Game Boy Advance recognizes the input as if a physical card had been scanned. This opens up access to rare promotional levels and even community-created content that can be loaded directly onto the hardware.

The adapter also supports multi-boot commands, which allows software to be sent to a Game Boy Advance that does not have a cartridge inserted. I used this feature to explore a project called GBA MMO, a multiplayer game that runs on original hardware, and streams its game data over the GB-Link.

This same multi-boot capability powers a utility called Cart Doctor. This tool allows users to insert their own game cartridges and dump the ROM data or back up save files directly to their computer, providing a method for preserving progress in decades-old games.

The project is currently in a state of active development, with new applications and game support being added as protocols are reverse-engineered. As more units reach the community, the potential for remote play and homebrew interaction continues to grow.

Information regarding the ongoing development and availability of the adapter is maintained on the GB-Link website. Thanks to Ashton for sending along some units to play with !

Disclosure: The GB-Link unit was provided to the channel free of charge. No other compensation was received and no one reviewed or approved my review prior to publication.