Local LLMs, AI Ethics: ID Links 2/17/26

As I read online, I bookmark resources I find interesting and useful. I share these links about once a month here on my blog. This post includes links related to local LLMs, AI ethics and risks, research on AI voices and avatars, desirable difficulties, event segmentation theory, and tips for job searching and using LinkedIn.

Local LLMs

How To Run an Open-Source LLM on Your Personal Computer – Run Ollama Locally

These directions for running an LLM on your own computer include both the less-technical method with Ollama’s one-click installer and the more technical command line options.

AnythingLLM | The all-in-one AI application for everyone

AnythingLLM makes it easier to run an LLM locally. If you run everything on your local machine, nothing is sent to a server elsewhere, protecting the privacy of your data.

AI ethics and risks

Don’t Panic: When Should AI Coaches and Assistants Request Human Intervention? – Parrotbox

Thinking about a “human in the loop” is a good start, but what does that really look like if you’re using chat bots or AI coaches at scale? I really like the example of levels of risk for safety in mental health and when chats should be reported or escalated to a human for intervention.

Meet Thaura | Your Ethical AI Companion

Thaura AI is designed as an ethical LLM. It doesn’t train models on your private data, has transparency about their business model, and advertises that it uses 94% less energy than ChatGPT.

The Hidden Mirror: Why Your AI is Only as Good as Your Thinking

Debbie Richards writes about critical issues related to working with AI: our own human cognitive biases. AI can reflect and amplify our own mental shortcuts. Being aware of our cognitive biases can make us more effective at working with AI.

The researchers identify three critical stages where our own thinking can steer AI off course:

Before Prompting: Our past experiences create a “halo” or “horns” effect. If you’ve had great results, you might over-trust the tool for tasks it isn’t ready for. Conversely, if you’ve been spooked by headlines about hallucinations, you might avoid it even when it could be genuinely helpful.

During Prompting: How we frame a question matters. “Leading question bias” happens when we bake the answer into the prompt, like asking “Why is product X the best?” This encourages the AI to ignore weaknesses. There is also “expediency bias,” where we settle for the first “good enough” answer because we’re under time pressure.

After Prompting: Once we have an output, the “endowment effect” can make us overvalue it simply because of the effort we put into the prompt. We also have to watch the “framing effect.” How we present that AI-driven data can completely change how our audience feels about it.:

—Debbie Richards

Something Big Is Happening

I’m not entirely convinced by this article, but I try to read opinions from a variety of sources and perspectives on AI. This author explains that the pace of AI is moving much faster than most people realize, and then argues that the amount of disruption AI will cause will be more significant and quicker than most people expect.

AI companies will fail. We can salvage something from the wreckage | AI (artificial intelligence) | The Guardian

Cory Doctorow has written a long article on the risks and problems of AI, particularly in the way that AI companies promote and hype the benefits of AI.

In automation theory, a “centaur” is a person who is assisted by a machine. Driving a car makes you a centaur, and so does using autocomplete. A reverse centaur is a machine head on a human body, a person who is serving as a squishy meat appendage for an uncaring machine.

“And if the AI misses a tumor, this will be the human radiologist’s fault, because they are the ‘human in the loop’. It’s their signature on the diagnosis.” This is a reverse centaur, and it is a specific kind of reverse centaur: it is what Dan Davies calls an “accountability sink”. The radiologist’s job is not really to oversee the AI’s work, it is to take the blame for the AI’s mistakes.

This is another key to understanding – and thus deflating – the AI bubble. The AI can’t do your job, but an AI salesman can convince your boss to fire you and replace you with an AI that can’t do your job.

For AI to be valuable, it has to replace high-wage workers, and those are precisely the workers who might spot some of those statistically camouflaged AI errors.

After more than 20 years of being consistently wrong and terrible for artists’ rights, the US Copyright Office has finally done something gloriously, wonderfully right. All through this AI bubble, the Copyright Office has maintained – correctly – that AI-generated works cannot be copyrighted, because copyright is exclusively for humans.

The fact that every AI-created work is in the public domain means that if Getty or Disney or Universal or Hearst newspapers use AI to generate works – then anyone else can take those works, copy them, sell them or give them away for nothing. And the only thing those companies hate more than paying creative workers, is having other people take their stuff without permission.

Research on AI voices and avatars

Do AI Voices and Avatars Improve Learning? Here’s What the Data Says

TechSmith conducted a global study to determine how AI voices and avatars affect learning. I was surprised at how well the high-quality AI voices performed. We seem to have crossed the threshold where high-quality AI voices perform comparably to human voice actors. I was also surprised at how well the AI avatars did, although their recommendations for specific use cases do make some sense. I wish they’d also done a separate control with no narrator visible on screen (AI or human). The fact that AI avatars can be comparable to humans in some instances isn’t that shocking, I guess, but I really want to see how it compares to just having the slide content and no face on screen.

What really makes learners pay attention? A voice that sounds clear, warm, and polished — not whether it’s human or AI. As voice quality improved in the study, so did professionalism ratings. In fact, 92% of viewers said the high-quality AI voice made the video feel professionally produced.

Results from the “pop quiz” portion of our study make the pattern clear: correct answers increased as voice quality improved. In fact, the high-quality AI voice produced the strongest retention numbers, aside from one low-quality human outlier.

But are AI voices distracting overall? It depends. Low-quality, synthetic voices are unmistakable and draw attention away from the content. When the AI voice sounds natural, many viewers can’t distinguish it from a human voice. The difference is less jarring, and information retention holds steady or even improves.

AI avatars aren’t distracting by default, but size matters. When an avatar fills the screen, viewers are more likely to notice robotic traits like lip sync issues, eye contact, limited facial movement, awkward blinking, or unnatural breathing.

The right format depends on your video’s purpose. Use this quick decision guide: Screen-heavy, procedural, and frequently updated content: High-quality AI voice with screen recording, plus an optional AI avatar in PiP. Emotionally sensitive, culture-setting, or leadership-driven content: Human presenter with a human voice.  Long-form, concept-heavy learning: A mix — human-led modules for core ideas, supported by AI-voiced micro-lessons and refreshers.

—Stephanie Warnhoff

Desirable difficulties

A Deep Dive into Desirable Difficulties

I’ve seen some misunderstandings on desirable difficulties on social media recently. This article has an understandable explanation of what desirable difficulties are (techniques that may initially cause errors and short-term performance issues but in the long run improve learning and task performance). The techniques include varied practice, spacing, reduced feedback and guidance, retrieval, and interleaving. If you’re new to the idea of desirable difficulties, this will give you a solid foundation.

Difficulties are desirable when they boost learning, not performance.

For example, when learning to drive, it would be easier to practice by driving round the same block multiple times, with an instructor sitting beside you and telling you exactly what to do. As a learner under such conditions, you’d make very few errors, if any.

However, once our lessons are over, we have to drive without an instructor telling us what to do, on complex and sometimes unfamiliar roads. The desirable difficulties framework would suggest, therefore, that practice should resemble that realistic situation, with a variety of road conditions to deal with, and reduced guidance or feedback.

Event segmentation theory

This article includes some great research translation by Tom McDowall about Event Segmentation Theory. We talk about “chunking” content to support learning, but we often rely on time or intuition to determine where to break up content. Event Segmentation Theory provides an evidence-informated approach to more meaningful divisions so you can improve the effectiveness of your training based just on where you add breaks. Tom includes lots of citations for further reading.

Your brain doesn’t process the world as one unbroken stream. It automatically divides ongoing experience into discrete chunks, which researchers call “events,” and does so continuously, without you deciding to do so or being aware that it’s happening.

Information present at a boundary, the moment when one event ends and another begins, gets encoded more strongly than information in the middle of an event. The boundary acts like an attentional gate: it opens briefly to let new information in, and that information gets a better foothold in long-term memory as a result (Kurby and Zacks, 2008).

There’s a trade-off, though. While boundaries improve memory for what happens at the transition point, they impair memory for temporal order across the boundary. Items that span a boundary are harder to sequence correctly and are remembered as being further apart in time than they were (Ezzyat and Davachi, 2014).

Six features of a situation reliably trigger event boundaries: spatial or location changes, character entrances or exits, new object interactions, goal shifts, changes in causal structure, and temporal discontinuities (Speer, Zacks, and Reynolds, 2007). In practical terms, the most reliable triggers for workplace training are changes in what you’re trying to achieve (the goal), changes in where you are or what you’re looking at (the environment), and changes in why the current action matters (the causal structure).

The most direct way to apply EST is to structure process training and standard operating procedures around the natural event structure of the task. Rather than organising steps by convenience or by how they appear in a system, map them to the hierarchical structure of the activity: major phases first (the coarse events), then detailed steps within each phase (the fine events).

—Tom McDowall

Job search and LinkedIn tips

“Steal My Wins” | Kimberly Scott

Kim Scott has been sharing lots of details on her job search and the strategies that are working for her. As a consultant, I’ve been out of the job market for a long time, so it’s helpful to have folks like Kim that I can point others to who are looking for work in this lousy job market.

How to fix your LinkedIn feed in one hour

If you find scrolling on LinkedIn terribly annoying, you may not have trained its algorithm well. Follow these tips to improve the quality of your LinkedIn feed.

You manage your feed by giving AI the signal. Signal for what you want. Signal for what you do not want. Then you reinforce it until the algorithm adjusts to your taste. That is it. Not complicated. But most people never do it.

—Tianyu Xu

Related