Month: September 2026

Synology Update : Drive Restriction Lifted, Plex Server Week Sale & Hardware Transcode Patch

It has been some time since my last update on Synology, and several developments regarding their hardware and software ecosystem haven taken place since then.

In my latest video, I’ll catch you up on the Synology ecosystem and talk about some big sales coming up on Synology hardware and Plex Pass subscriptions.

One of the primary concerns for users over the last year has been the potential for hard drive restrictions on their newer devices. Synology was considering requiring the use of their own branded drives vs allowing users to pick their own. In my recent testing, I have found that these restrictions have not been put into effect for the consumer-facing Plus series. I have been running a DS225+ (compensated affiliate link) with a mix of third-party drives and the system functions without issue.

This flexibility extends to memory as well. While these devices often come with a modest amount of base RAM—2 GB in the case of the DS225+—it is possible to use off-brand non-ECC DDR4 modules to increase capacity. I successfully upgraded a test unit to 8 GB using a standard stick of RAM I had available. This extra memory is particularly useful for those looking to run multiple Docker containers or server applications.

When selecting a device for use as a media server, the choice of processor is a significant factor. The models equipped with Intel J4125 chips, such as the 225+ and 425+, are better suited for hardware transcoding than the Ryzen-based models like the 1525+ or 1825+. All of them will function well as Plex servers if transcoding is not needed, but I do consider hardware transcoding to be a core Plex function that every server should support.

The Beestation Plus supports hardware transcoding for Plex immediately upon installation. However, the 225+ and 425+ require a different approach. While they have the necessary Intel hardware to handle transcoding with minimal CPU strain, this feature is not enabled by Synology out of the box for Plex. To access this functionality, users must utilize an unofficial community patch from sources like the Black Void Club. By adding this third-party repository to the Synology Package Center, users can install the necessary drivers. It is important to note that this is an unsupported modification, though in my tests, it allowed the Intel processor to handle heavy transcoding tasks efficiently.

Beyond media streaming, these devices function as a comprehensive alternative to cloud-based office suites. Synology’s software includes tools for word processing and spreadsheets, allowing for a self-hosted ecosystem that mirrors the functionality of services like Google Apps or Dropbox. I have used these systems to manage decades of personal data and to run automation servers for various tasks. The security and backup options are a core part of the appeal. Tools like HyperBackup allow for incremental data snapshots that can be sent to external drives, offsite Synology units, or cloud storage like Amazon S3.

Synology has been very good over the years supporting their hardware with consistent and free software updates. For those who want to maintain control over their data without the constant overhead of manual server management, these units provide a stable middle ground. The ability to integrate VPN services like Tailscale further simplifies remote access, making it possible to manage a personal server from anywhere without complex network configurations or having to open router ports.

An upcoming promotion, scheduled for September 21 through September 25, 2026 for Plex Server Week, will see 15% discounts on several models, including the BeeStation Plus, the DS225+, and the DS425+. There will also be a 10% discount on certain rack station models. You can find the deals here on this Newegg page (compensated affiliate link). During this same window, discounts of 20% will be available for annual, five-year, and lifetime Plex Pass subscriptions with the coupon code SERVERWEEK20.

It is disappointing that Synology chose to disable hardware transcoding on their new devices, although it’s not very difficult to get that feature back. But beyond that issue Synology makes very robust nearly turn-key self hosted server solutions that will continue to be part of my workflow.

An Appearance on the TWiT Network!

I had a fun time chatting about my local AI experiments with Leo Laporte on his Intelligent Machines show. This show used to be called “This Week in Google.” Also on the show was Father Robert Ballecer (who I bump into at CES frequently) and Jeff Jarvis who I first met when I was working on my journalism startup back in 2009.

Check it out here!

Leo was most interested in my latest Local AI experiment that involved re-purposing an old Nvidia Tesla V100 data center card into something that can work on a desktop PC. You can check out my video on that topic here!

I also talked about how good local AI has become for practical applications like document analysis or managing personal data. Local models are effective, grounded, and largely free of hallucinations seen in older models. My favorites at the moment are Gemma 31B and Qwen 3.8 27B. Both are “dense” models that run at usable speeds (around 20-25 tk/s) even with large 96k context windows.

One of the projects I’m working on is building a virtual “Chief of Staff” that manages all of the balls I have in the air running my business, serving my community on the school board and a few other activities I’m involved with. That will be the subject of an upcoming video once I have it working the way I’d like it to work. Stay tuned!

Roku Buried Your Apps. Here’s How to Get Them Back

For more than a decade, the Roku interface remained largely unchanged, defined by a simple grid of applications and an ad on the right hand side of the page. Following a recent update, however, the home screen now features a significant redesign that prioritizes promotional content and pushes the old app interface “below the fold.” This new layout introduces “top picks,” a quick access menu, and several rows of advertising-supported links and curation features that users must scroll through before reaching their own apps.

In my latest video, I look at how you can turn off these features and return to the old interface!

For those who prefer the original, minimalist design, these features can be disabled. By navigating to the settings menu and selecting the home screen options, it is possible to hide the recommendation rows and the quick access menu. Once these are turned off, the interface reverts to its previous state, placing installed applications back at the center of the screen. Users can also manually reorganize their apps as before by using the star button to move preferred services, such as Netflix, to the top of the list.

The push for monetization extends beyond the home screen and into the device’s idle state. The “Roku City” screensaver, long a staple of the platform, now includes targeted advertisements. Those who wish to avoid ads on their screensaver can switch to an ad-free screensaver like “Aquatic Life” or some of the other ad-free options in the settings screen.

There are also deeper privacy implications to consider within the system settings. Roku utilizes a technology called Automatic Content Recognition (ACR), which is designed to monitor what is being watched on the screen. This system sends a digital fingerprint of the images to a central server, allowing the company to provide data to advertisers about viewing habits. This occurs even when using apps Roku does not control or when watching content through an external device connected via HDMI. In the privacy section of the settings menu, users can disable ACR and the content viewing disclosure to limit this data collection. Additionally, checking the option to “limit use of sensitive information” and opting out of personalized ads provides a further layer of control over how personal data is handled.

The shift toward a more intrusive interface is a reflection of Roku’s evolving business model. The company is no longer primarily a hardware manufacturer; instead, its operating system serves as a vehicle for a massive advertising and content platform. Financial reports indicate that Roku generates approximately $5 billion in annual revenue, yet only about 11% of that comes from hardware sales. Roughly half of its revenue, $2.5 billion, is derived from advertising, with the remainder coming from subscriptions serviced on behalf of other providers. This revenue structure has made Roku an attractive acquisition target for Fox Broadcasting, who is in the process of a $22 billion dollar acquisition of the platform.

Every pixel on the screen that is not being used for advertising represents potential revenue left on the table. The company recognizes that most users will not navigate deep into the settings to disable these features or opt out of data tracking. While alternative hardware like the Apple TV offers a different experience, those who choose to stay with their current equipment can still maintain a level of privacy and simplicity by taking a few minutes to adjust their preferences and, if necessary, disconnecting their smart televisions from the internet entirely. Managing these settings is the most direct way to ensure the device remains a tool for viewing content rather than a tool for monitoring the viewer.

Meshcore / Meshtastic Devices That Work Out of the Box!

It has been several months since I last explored MeshCore and Meshtastic developments. This off-grid radio technology enables communication over significant distances without relying on the internet or traditional telecommunications providers. Looking at the live map of my home state of Connecticut, it is clear how rapidly this network is expanding. It is a community that communicates over the air using various hardware devices rather than through centralized platforms like Discord or Reddit.

In my latest off-grid video, I take a look at some turn-key devices that connect tothese networks right out of the box!

For those unfamiliar with the specifics of Mesh Core or Meshtastic, the communication is limited to short text messages. It does not support images or video, but it provides a functional way to stay in touch when other systems fail.

The utility of the network is best demonstrated through real-world interaction. In my video, I sent a message to a public feed asking a general question. A response came from a contact located an hour away. By analyzing the message path, I could see that the signal traveled 130 miles, bouncing through multiple repeaters across Massachusetts and Connecticut before reaching me. This illustrates the potential of the mesh architecture when a sufficient number of nodes are active in an area.

When I began working with this technology, the barrier to entry usually involved assembling components like Heltec ESP32 into custom-built radios. The market has since evolved to include more turnkey options that allow for a quicker setup. While these ready-to-use devices often carry a higher price point, they offer a path for those who prefer not to spend time on assembly.

One such device is the Seed Studio SenseCAP (compensated affiliate link). This unit integrates a solar panel, a Meshcore radio, and a battery into a single weatherproof enclosure. I have kept this unit outdoors for several months, where it has remained powered by sunlight and consistently linked to the local Meshcore repeater network. By placing this unit outside, I have improved my signal performance, as it acts as a bridge that repeats messages into my home. It has proven reliable through heavy rain and high winds, even without permanent mounting.

My primary device for daily use is the Heltec MeshPocket (compensated affiliate link). This unit features a 5,000 milliamp-hour battery and an e-ink display. It pairs with a mobile app via Bluetooth to send and receive messages. One useful feature of this specific model is its ability to receive and store messages even when disconnected from a phone, syncing them once the connection is restored. It also includes a wireless charging pad on the back, allowing it to function as a power bank for a smartphone.

Other manufacturers, such as Elecrow, offer various hardware configurations. The ThinkNode M5 (compensated affiliate link) is a compact device resembling a small walkie-talkie with an e-ink screen and built-in GPS. While portable, its battery life is limited to approximately one day. For those prioritizing power, the ThinkNode M4 (compensated affiliate link) lacks a screen but includes a 7,000 milliamp-hour battery, a flashlight, and multiple USB-C ports for data and charging. It also functions as a wireless power bank. Because it has no display, it requires a phone pairing to view network traffic.

For a completely self-contained experience, the ThinkNode M9 (compensated affiliate link) features a physical keyboard similar to older mobile devices. It does not require a phone to function. While the keyboard and firmware navigation require some adjustment—such as waiting for long messages to scroll across the screen—it serves as an all-in-one communicator with GPS capabilities.

Beyond handheld devices, there are opportunities to integrate this technology with computers. I have experimented with connecting a radio to a Raspberry Pi to run a Bulletin Board System (BBS) using a package called Supply Drop. This setup allows users to leave and read messages on a persistent server within the mesh network. The BBS software works with both Meshcore and Meshtastic networks, allowing different protocols to interact. The host machine does need two radios for this to work, however.

The variety of hardware available now makes it easier to join these independent networks, whether through DIY projects or out-of-the-box solutions. And if you’re not sure whether to go with Meshcore or Meshtastic, most of the devices listed today have firmware that works with both. My advice is to research which network is the most active in your particular region.

See you off the grid!

I turned my Onn TV Box Into a VPN Router with Tailscale!

Tailscale has become a staple in my daily networking setup. As a personal VPN application that is free for personal use, it allows for the creation of a private network of devices that can be accessed from anywhere in the world. The primary utility is the ability to connect to home servers and hardware as if they were on a local network, without the security risks associated with exposing those devices directly to the internet.

In my latest Tailscale video, I explore the “subnet router” feature that allows any device on your local network to be accessible through Tailscale – even devices that can’t run the Tailscale client!

In the video I configured an Onn Google TV box to act as a subnet router. Most people use these devices strictly for streaming, but they can be deputized to serve as a gateway for an entire home network while still working as a TV box!

I opted for a wired Ethernet connection rather than Wi-Fi to ensure network reliability and to avoid congesting the wireless spectrum. While this specific hardware is limited to 100 megabit Ethernet, the technical process is identical for more powerful hardware like an Apple TV or a dedicated PC.

The configuration began within the Tailscale application on the Android TV box. Under the settings menu, I located the subnet routing option and entered my local network range, which in my case is 192.168.2.0/24. Once the route was added on the device, the next step required administrative approval through the Tailscale web-based admin console on a computer. After navigating to the “machines” section and selecting the TV box, I authorized the internal network route to allow traffic to flow through the device.

The results were immediate. To test the connection, I used a smartphone connected to a cellular network, entirely separate from my home Wi-Fi. By entering the local IP address of my HDHomeRun TV tuner into a browser on the phone, I was able to connect to the tuner through the TV box. It is worth noting that this method does not support network broadcasts, which means automatic discovery features for printers or tuners will not work. Users must know the specific local IP address of the device they wish to access.

To evaluate the throughput, I tested the connection from a location with a high-speed cellular signal. Through the TV box sitting in my basement, I was able to pull a sustained downstream speed of approximately 87 megabits per second. While the inbound speeds were notably slower—likely a limitation of the budget components in the Android box—the performance was more than adequate for streaming live high-definition TV from my HDHomerun tuner. Using VLC, I successfully tuned into a local broadcast signal with minimal latency and no need for transcoding.

This setup demonstrates that even inexpensive hardware can serve as a robust networking tool. Because TV boxes are typically left powered on and connected to the network, they make ideal candidates for this type of always-on routing. While those requiring higher bandwidth might prefer a more powerful host, this approach provides a functional and accessible way to manage a home network from afar.

Can This Unusual RTX 3060 GPU Justify Its $500 Price? Gaming and LLMs Tested!

I recently acquired the low-profile graphics card with an Nvidia RTX 3060 chip grafted on with 12GB of VRAM that’s been buzzed about in the tech press. The card is notable for its compact dimensions and the fact that it is entirely bus-powered with no additional power connector required. While this limits its maximum computational output, it offers a simplified installation process for small-form-factor systems. The package includes an additional mounting bracket for different case sizes and a basic manual. You can find one over at Newegg (compensated affiliate link).

But is it worth the $500 price tag? That’s what I explore in my latest video.

For my initial testing, I installed the card into a desktop featuring an Intel 10900K processor and 32GB of RAM. To my surprise, the standard Nvidia drivers installed without issue, recognizing the hardware immediately.

I began with a gaming test using Cyberpunk 2077. At 1080p resolution with the lowest settings, the card maintained roughly 70 frames per second. While playable, this performance falls short of what a standard, desktop 3060 typically delivers with similar settings. The card is fundamentally limited by the power it can draw from the system bus.

Synthetic benchmarks further illustrated these limitations. In the 3D Mark Time Spy test, the system returned a score of 4,846. This is significantly lower than a standard 3060 and even trails the performance of a much older GTX 1080. During a sustained stress test, the card achieved a stability score of only 85.8%, indicating significant thermal and power throttling.

But one potential use case for this card is local large language model (LLM) execution, as it features 12GB of video RAM. Using LM Studio, I tested the Gemma 4 12B-QAT model. While the model fit entirely within the VRAM, the generation speed was approximately 18 tokens per second. The GPU was fully utilized during this process, confirming that the power constraints also impact its effectiveness for artificial intelligence tasks. It is functional for this purpose, but notably slower than a standard version of the same GPU.

To see how the card performed in a multi-GPU setup, I moved it to a similar specced system equipped with an RTX 3080. The installation was seamless, and the two cards worked in tandem. When running a mixture-of-experts model like Gemma 4 35B-A4b, the system achieved 66 tokens per second because the primary compute tasks remained on the faster 3080 while using the 3060 for additional memory. However, when I switched to denser models like the 31-billion parameter Gemma 4 or the 27-billion parameter Qwen 3.8, performance dropped to between 6 and 8 tokens per second. In these scenarios, the 3080 was forced to wait for the slower 3060 to complete its portion of the workload, creating a significant bottleneck.

The build quality of the card, particularly the heatsink, is basic and likely contributes to the thermal issues observed during testing. While the additional 12GB of memory provides a way to load larger models that might not otherwise fit on a single consumer card, the performance trade-offs are substantial.

Given the high price point and the inherent power and thermal limitations, this card remains a difficult hardware choice to justify for most standard applications.

NextGen TV Update: 5G TV Backers Pledge No DRM — With a Catch..

I have been following the shifts in over-the-air television for some time, particularly the ongoing debate over the ATSC 3.0 standard. While major broadcasters continue to push for encryption and digital rights management (DRM) through NextGen TV, a different path is emerging for low-power television stations across the United States. A new proposal centered on 5G broadcast technology is currently before the FCC, offering a potential alternative that could change how signals are received on mobile devices and home tuners alike – and its backers are pledging not to encrypt the broadcast signal – but it’s important to read their fine print.

Learn more in my latest video!

Unlike traditional 5G cellular communication, which involves a two-way exchange between a phone and a tower, 5G broadcast operates as a one-to-many system. It sends data out from a single source to any device capable of receiving it, much like a traditional radio or television signal. Because it utilizes the same fundamental technology as modern smartphones, it could allow users to watch television on their devices without a cellular subscription or a data plan, provided the hardware is equipped with a compatible antenna.

One of the primary arguments for 5G broadcast is the promise of open access. HC2 Broadcasting Holdings, which owns a significant number of low-power stations, recently informed the FCC that their implementation would remain free-to-air without the use of DRM or paywalls. This stands in contrast to the current rollout of NextGen TV, where encryption has complicated the manufacturing of set-top boxes and slowed consumer adoption. Supporters of the 5G standard argue that it would be less expensive for manufacturers to integrate this technology into smartphones and boxes than it would be to comply with the complex encryption standards currently required by the major broadcasters.

The proposal includes specific limitations that warrant closer examination. While HC2 has pledged to keep signals unencrypted, they are seeking permission to broadcast at the bare minimum – only 480i—the standard definition used in the analog era. This would allow them to satisfy the legal requirement for broadcasting while leaving the majority of their spectrum available for other more profitable data services. In earlier filings, the group even suggested they might prefer to use their spectrum entirely for data casting rather than television, a move that the National Association of Broadcasters (NAB) has rightly criticized as failing to serve the public interest.

The tension between large and small broadcasters is evident in recent FCC filings. Sinclair Broadcast Group and the NAB have both voiced opposition to the 5G proposal, arguing that ATSC 3.0 should remain the single national standard. This creates a rift within the industry, as many low-power stations are members of the NAB yet find their interests aligned with the 5G alternative.

Meanwhile, Qualcomm has emerged as a significant ally for 5G broadcast, suggesting that the hardware required for smartphones to receive these signals is attainable and could be implemented without significant added costs to the consumer.

Beyond daily entertainment, there are potential applications for emergency services. Instead of receiving a simple text message during a crisis, a 5G broadcast signal could deliver rich media, such as live video feeds or detailed maps, to millions of devices simultaneously without straining existing cellular networks. This capability could satisfy the public benefit requirements that all broadcasters must meet in exchange for using the public airwaves.

The ATSC organization has taken a relatively neutral stance, urging the FCC to be cautious but not outright opposing the development of additional standards. As the FCC considers whether to allow further experimentation or a formal rulemaking process for 5G broadcast, the focus remains on balancing technological innovation with the public’s right to access free television.

Whether the market can support multiple standards or if the current friction will lead to a single unified path depends on how these technical and regulatory hurdles are resolved in the coming months.

The Cheapest 32GB Nvidia GPU You Can Buy for Local AI (Nvidia Tesla V100)

For the past couple of months, I have been exploring the practicalities of locally hosted AI models and applications. My primary objective to see what we can do with things we already own, or find cost-effective alternatives to a $5,000 Nvidia GPU.

In my latest video, I take a look at a relatively affordable 32GB GPU solution : an Nvidia V100 GPU pulled from a datacenter!

In an earlier video, I looked at a new Intel GPU with 32GB that retailed for $1,300. One of these pulled GPUs costs about half that. I got mine from Server Parts Deals who is also selling these cards on Amazon (compensated affiliate link).

Although the V100 lacks some of the specialized features found in modern RTX cards that newer models will begin to depend on, for now it provides ample performance from popular large language models (LLMs) like Qwen 3.8 27B.

Integrating a data center card into a consumer environment requires several specific modifications. These cards are designed for server racks where external fans force air through the chassis, meaning the card itself lacks active cooling. To address this, I sourced a specialized blower fan (compensated affiliate link) from an online marketplace to attach to the rear of the unit. Because I do not currently have a dedicated desktop for this project, I opted to use a small form factor PC with the GPU connected via an OcuLink interface.

The power requirements added a layer of complexity to the assembly. Data center cards use different power pin configurations than standard consumer hardware, necessitating a specific adapter cable (compensated affiliate link). Furthermore, because the external enclosure needs to power both the GPU and the high-draw blower fan, multiple power leads and adapters were required to ensure the cooling system remained operational at a constant speed. While some enthusiasts have developed variable-speed cooling solutions, I focused on the minimum requirements necessary to maintain stable operating temperatures under load.

The software installation process presented its own set of challenges. Initial attempts to configure the V100 on Windows 11 were unsuccessful; despite installing the correct data center drivers, the system consistently returned a hardware error, likely due to the OCuLink connection or BIOS issues. Consequently, I shifted the project to a Linux environment running Ubuntu 26.04. To streamline the configuration and driver installation, I utilized ChatGPT’s command-line Codex application to manage the technical hurdles and document the installation for future reference. This approach allowed for a more efficient setup of the necessary dependencies and libraries.

Once operational, I tested the system using the llama.cpp web interface, starting with the Qwen 3.8 27B dense model. When tasked with summarizing a few PDF documents, the V100 achieved a pre-fill speed of 50 tokens per second and a generation speed of 45 tokens per second. The 32GB of VRAM allowed for a large context window of over 96,000 tokens, enabling the model to process a 70-page document without exceeding the hardware’s limits. Even as the context window filled to 80% capacity, the generation speed remained a functional 24 to 30 tokens per second while drawing approximately 270 watts of power.

Performance shifted significantly when testing a “mixture of experts” model, such as Gemma 4 26B-A4B. Because these models only activate a fraction of their parameters for any given task, the processing speeds were much higher. The system reached pre-fill speeds of 1,100 tokens per second and output generation at over 90 tokens per second. These results indicate that for text-based local AI tasks, this older data center hardware remains competitive with much more expensive modern alternatives.

Beyond text, I tested the card’s ability to generate images and video using ComfyUI. While the V100 can run modern models like Flux for images LTX 2.5 for video, the process is considerably slower than on contemporary hardware. Generating a single 1024×1024 image took approximately one minute, while a five-second video clip at 720p required three and a half minutes of processing. By comparison my Intel card could do the same job in under a minute and a half. Long videos were much slower vs. the newer Intel card in my testing.

Newer cards from Intel and Nvidia have specific optimizations for these workflows that the V100 lacks, as it was originally released in 2017. For example, the exceptional MiniMax H3 model was too slow to be usable on the V100.

There are clear trade-offs when using a decade-old data center GPU. Many emerging AI features expect hardware capabilities that simply did not exist when this card was engineered. While it handles current language models with ease, it may not support the next generation of multimodal AI tools as effectively. However, for those specifically seeking a large context window and high VRAM capacity for local text processing on a budget, these secondary-market cards offer a viable path forward.

My Theory on Why YouTube Changed View Counts..

YouTube recently updated its platform to change the way video views are calculated. Under the new system, a view is recorded as soon as a single frame of a video is watched. Previously Youtube measured a level of genuine engagement before increasing the view count. My theory? YouTube is doing this to try and capture more of the revenue flowing to creators.

Check out my latest video where I present the evidence!

On my own channel, this change has resulted in a noticeable increase in reported views. Typically, my daily traffic ranges between 15,000 and 30,000 views, driven largely by search-focused, evergreen content such as printer reviews. Now I’m seeing 45,000 or more.

While the higher numbers might initially appear positive, they have rendered traditional performance metrics less reliable. For instance, a recent video I produced regarding local AI ranked highly in my internal analytics based on view count. However, a closer look at the advanced analytics—which now requires several additional steps to access—revealed that the “engaged views” were actually on the lower end of my typical performance. The public-facing view count no longer reflects actual audience interest.

YouTube says that this change was implemented to align metrics between long-form videos and Shorts. However, the result is that the most meaningful data is now obscured within advanced analytics tools. This may be an attempt to reclaim a cut of revenue that currently flows directly to creators.

Many creators work with outside agencies to secure direct advertising deals. These agreements are often predicated on public metrics, specifically the average view count per upload. With the recent change, these public figures have become less indicative of a channel’s actual influence. This creates a challenge for agencies trying to distinguish between legitimate audience engagement and accidental frame clicks. It also lowers the practical value of the RPM (rate per thousand views) that has served as a benchmark for the industry. From YouTube’s perspective, having an engaged view be the metric was too valuable a measurement to share publicly.

There is evidence to suggest this timing is intentional. Shortly after the view count change, the platform rebranded its internal agency as YouTube Creator Partnerships, signaling a renewed effort to connect brands with creators directly. By positioning themselves as the sole gatekeepers of accurate engagement data, they provide a compelling reason for brands to bypass outside agencies in favor of their own proprietary systems.

The recent integration of Amazon into the YouTube Shopping affiliate program further illustrates this shift toward a controlled revenue pipeline. For years, creators have used direct Amazon affiliate links in their descriptions because the system was efficient and did not require sharing a commission with the platform. I’ve also noticed that viewers are far more likely to click a link that I provided vs. one that appears through YouTube Shopping.

But just recently, I observed an experiment where external affiliate links in video descriptions were rendered unclickable, while links generated through the official YouTube affiliate program remained functional. This suggests a move away from the open web toward a walled garden where the platform can facilitate—and take a cut of—every transaction.

By devaluing public data and testing restrictions on external linking, the platform is exercising its position as a dominant media entity to maximize its own revenue. While these changes may not prevent creators from seeking independent partnerships, they certainly make those deals more difficult to quantify and execute. Brands may feel compelled to funnel their creator sponsorships through YouTube to have a better window on performance.

Local AI On an Old 8 GB GPU?

The current landscape of artificial intelligence often suggests that high-end, increasingly expensive hardware is a prerequisite for meaningful performance. However, much can be accomplished using hardware that many people already own. I recently spent time working with a five-year-old gaming laptop equipped with 16 gigabytes of system memory and an Nvidia 3070 GPU with 8 gigabytes of video RAM. While this is a far cry from a modern workstation, it provides a capable environment for running local language models if the settings are adjusted correctly.

Check out what it can do in my latest local AI video!

The primary challenge with an 8-gigabyte GPU is balancing the size of the model with the context length, which is essentially the memory the model uses to track a conversation or analyze a document. For my testing, I utilized the Qwen-3-8B model, an 8-billion parameter model that occupies about 5.76 gigabytes of memory. This leaves a small but workable margin for context.

In the LM Studio application, I set the context limit to 25,600 tokens (compared to 1 million tokens on frontier models like ChatGPT and Gemini). To keep the performance steady at around 35 tokens per second, I employed several technical tweaks, including unified KV cache, flash attention, and 8-bit KV cache quantization. This quantization is particularly important because it compresses the context memory, allowing more data to fit onto the GPU.

To test the practical utility of this setup, I tasked the model with analyzing a term paper I wrote in college nearly 30 years ago. Although the text file was small, the process of analyzing it for patterns can consume significant memory. I inserted several “traps” into the text—phrases like “it’s a trap” and “Han shot first”—to see if the model would notice them while summarizing the content. The model processed the prompt in a few seconds, provided an accurate summary without hallucinations, and correctly identified the out-of-context phrases.

Moving beyond simple text analysis, I attempted to use the laptop as an AI server for more complex tasks like coding. By connecting a secondary machine to the laptop using an open-source coding harness called OpenCode, I tried to extract a list of public officials from a website and format the data into a CSV file. This is where the limitations of an 8-gigabyte system become more apparent. The model struggled to process the raw HTML of the webpage due to the limited context window. Once I simplified the input to plain text, the model successfully formatted the data, including identifying vacancies in a specific district. It required more manual intervention than a larger 70-billion parameter model might, but it eventually achieved the desired result.

Coding tasks showed similar limitations. When I asked the model to write a Space Invaders game in HTML, it produced a functional interface where a player character could move, but it failed to include enemies or game logic on the first attempt. Even after a second prompt to fix the errors, the logic remained incomplete. Smaller models often lack the reasoning capabilities required for complex, one-shot coding successes, frequently requiring a more iterative approach and careful management of the conversation history.

Image generation is also possible on this hardware. Using the SDXL model through a templated interface, I was able to generate local images in approximately 30 to 32 seconds each. While this setup cannot handle the video generation tasks possible on high-end cards, the static image quality is respectable. Achieving the right output required some trial and error with prompts; for instance, the model initially struggled to generate a dog in a cockpit until I removed gender-specific language that seemed to be steering it toward human subjects.

The efficiency of these local models has improved significantly over the last year. For those who may have tried local AI in the past and found it lacking, the current generation of small-parameter models offers much more utility on aging hardware.

While 8 gigabytes of video memory requires careful optimization and realistic expectations regarding context and complexity, it remains a viable entry point for those looking to keep their data local and make use of the machines they already have. If you’re willing to give up performance, AI applications like LM Studio can also use system memory which can allow for larger context memory and AI models.

GMKTEC M3 Pro Review – DDR4 and an Older Intel Chip Make it Somewhat Affordable

I have been searching for lower-cost options in the mini PC market recently and found another candidate that meets the criteria : the GMKTec M3 Pro. This device is powered by an Intel i5-13500H processor and DDR 4 RAM. This is older technology, but at this point older is the only way to reach affordability in the current environment.

Check it out in my latest Mini PC review!

The unit is priced at $329 for a barebones configuration, which excludes RAM and storage. For those preferring a pre-configured system, Amazon has some options preconfigured with RAM and storage (compensated affiliate links).

The M3 Pro features a compact design with a build that utilizes metal on the sides and plastic for the top and bottom panels. On the front of the device, there are two 10 gigabit per second USB-A ports. The rear houses two additional USB-A ports, though these are limited to 5 gigabits per second. For display connectivity, there are two HDMI outputs capable of 4K resolution at 60Hz. A USB Type-C port is also present and supports a third 4K 60Hz display. However, it is important to note that this is not a Thunderbolt or USB4 port. Due to the older chipset, the Type-C port is limited to USB 3.2 specifications, meaning data transfers are capped at 10 gigabits per second.

The M3 Pro includes a 2.5 GB Ethernet port that performed to its rated specifications during my testing, maintaining full potential on both downstream and upstream traffic. The Wi-Fi 6 performance was less consistent, reaching about half a gigabit per second on the downstream and approximately 350 megabits per second on the upstream. Internally, the system is upgradeable. Removing the bottom panel reveals two RAM slots supporting up to 64 GB of DDR4 memory. Because these Intel chipsets favor dual-channel memory, it is advisable to install RAM in pairs. Storage is handled by two NVMe slots: one 2280 slot and a secondary 2242 slot for shorter drives.

In terms of performance, the M3 Pro handles specific workflows with varying degrees of efficiency. When testing local AI using the Gemma 26B-A4B mixture of experts model, the system generated approximately 12 tokens per second. This is usable but noticeably slower than modern AMD-based systems equipped with faster DDR5 memory, which can exceed 20 tokens per second. Video editing in DaVinci Resolve is manageable for basic 4K 60Hz tasks, such as simple transitions or stringing clips together. However, more complex projects involving heavy effects cause the system to slow significantly. The lack of a Thunderbolt port also precludes the use of an external GPU to bolster this performance.

General computing is where the machine feels most responsive. Web browsing is nearly instantaneous, and the device handled high-bitrate 4K 60Hz YouTube playback without dropping frames. It achieved a score of 29.6 on the Speedometer benchmark, which is consistent with other systems in this category with current generation chipsets.

Gaming performance is comparable to a Steam Deck. Running Cyberpunk 2077 at 1080p on the lowest settings resulted in about 20-25 frames per second, though lowering the resolution to 720p would likely push the frame rate above 30. It serves as a capable box for 8-bit and 16-bit emulation and can handle some GameCube titles. In the 3DMark Time Spy test, it scored 1,644, which aligns with expectations for this processor generation.

Thermal management can be adjusted through a high-performance mode in the BIOS. In this mode, the fan is more aggressive, and the system passed a 3DMark stress test without significant thermal throttling, maintaining a temperature of roughly 69 degrees Celsius under sustained load. While the fan is audible during heavy tasks, it is not particularly loud compared to other mini PCs, and it remains largely silent performing basic computing tasks. Power consumption sits at 10 to 11 watts at idle and reaches about 75 watts under a full load.

Linux compatibility is a strong point for this hardware, likely due to the maturity of the older chipset. Running Ubuntu 26.04, the experience was snappy and all hardware components—including video, audio, Ethernet, Wi-Fi, and Bluetooth—were detected without manual configuration.

While the M3 Pro lacks the hardware required for demanding local AI or high-end video production, it remains a functional choice for general office work, home server duties, or use as a dedicated Linux desktop. The trade-offs in port speeds and memory technology are reflected in the lower entry price, making it a utilitarian option for users with standard computing needs.