“How do we move beyond knowing a few prompt tricks to developing a thoughtful and responsible approach that will continue to serve us well as AI keeps evolving?

We discuss how AI Fluency involves developing practical skills, knowledge, insights, and values that help you interact with AI systems in ways that are effective, efficient, ethical, and safe. We also introduce three ways people engage with AI:

• Automation: The AI completes specific tasks based on your instructions.
• Augmentation: You and AI collaborate as creative thinking and task execution partners.
• Agency: You configure AI to work independently on your behalf, establishing its knowledge and behavior patterns rather than just giving it specific tasks.

“AI isn’t just a tool. It’s a technology that can act as a tool, but also as a medium or as a partner or co-creator, and sometimes all of these at once.”

I don’t trust a probabilistic bot to complete any sort of tasks for me, I don’t trust the output it gives, and I don’t want to use it as a soundboard to brainstorm with. That’s what the humans on Mastodon are for.

He says you could “power a dynamically interactive character in a game”, but why would I design a video game that could go off the rails, that I don’t have creative control over?

Key takeaways

• AI Fluency means engaging with AI in ways that are effective, efficient, ethical, and safe
• There are three primary ways we engage with AI: (Automation/Augmentation/Agency)
• The AI Fluency Framework consists of four core competencies (the 4Ds):
  • Delegation: Deciding what work to do with AI vs. yourself
  • Description: Communicating effectively with AI systems
  • Discernment: Evaluating AI outputs critically
  • Diligence: Ensuring responsible AI collaboration
• These competencies apply across all three ways of working with AI
• Developing these competencies prepares you for evolving AI capabilities

Figure 1: Understand your goal and the problem that you are trying to solve. Know what AI systems can and can't do well. Decide how to divide the work between you and the AI.

Presentation of text summary beneath the videos feel very surface-level and repetitive. Probably AI generated. Not worth much value.

The videos feel more like marketing videos rather than actual educational videos.

“Think about a research project you’re working on. You might decide to have your AI assistant review lengthy documents and data, and then engage in thoughtful discussion about the implications and findings, but reserve the critical analysis and final conclusions for yourself.”

I would not trust that I would glean all of the information, or the correct information, were I to delegate the looking at “documents and data” to an AI bot. It would be best for me to invest the time to do that work myself, keeping notes and synthesizing my own products.

Figure 2: What you want the final output to be. How yuo want the AI to approach the task. How you want the AI to behave.

I don’t want to verbally program a bot in English every time I need a task done.

Figure 3: Is the output useful and correct? Is the AI taking the right approach? Is the AI behaving as desired?

Why would I rely on SOFTWARE that gives INCORRECT and UNUSEFUL results, to do things like “look through data for my research project”? How does this seem like a useful tool?

“Most of our interactions with AI involve small loops of description and discernment: describing what we need, evaluating what we get, refining our request, and so on.”

Or… I could just use my brain. I could generate ideas in my brain. I could create drafts, and then revise those drafts.

Figure 4: Ensuring accuracy and taking responsibility. Honesty and transparency. Ethical use and critical awareness.

“When making important decisions with AI assistance, how are you verifying the accuracy of the information presented to you?”

Or you could find the accurate information and synthesize from there instead of working “backwards” like this.

“Have you considered how to be transparent about the involvement of AI?”

Have you considered? Did you CONSIDER IT? Is the CONSIDERING the ethical step? What is the ethical result? Or do we just CONSIDER it?

“Are you willing to be accountable for the AI assisted work you have done?”

No, I am not.

“Generative AI refers to artificial intelligence systems that can create new content rather than just analyzing existing data.”

But to generate """NEW DATA""" it must be trained on existing data; an LLM cannot generate """NEW DATA""" without a training data set.

“They’re called language models because they’re trained to predict and generate human language, and large because they contain billions of parameters - mathematical values that determine how the model processes information, somewhat like synaptic connections in your brain.”

These parameters are probabilistic WEIGHTS between TOKENS along many dimensions. It is not “processing information” like a brain.

Figure 5: Three pillars that made it possible. Algorithms - Neural networks and Transformers; Data - Articles and websites, and Code and multimodal content; Computation - GPUs, TPUs, etc, and Computing clusters

Figure 6: Scaling laws - Compute and data, and Model intelligence

“The combination of these three factors led to an important discovery known as the Scaling Laws. These empirical findings showed that as models grew larger and trained on more data with more computing power, their performance improved in predictable ways. More surprisingly, researchers found that entirely new capabilities began to emerge as these models grew larger. Abilities no one explicitly programmed, like reasoning through problems step-by-step or adapting to new tasks with minimal instruction.”

Did it, though? Such magical thinking.

Figure 7: Behind the scenes of LLMs: Pre-training; a diagram of inputs ("What do you think"), a series of interconnected nodes, and then the output "Next token: About".

“During initial training, also called pre-training, LLMs like Claude analyze patterns across billions of text examples. Imagine reading every website and piece of text you can find, not just to absorb information, but to understand the statistical relationships between words, phrases, and concepts.”

The latter is the ONLY thing this training does – the statistical relationships between tokens – The training does NOT make the AI read every website and piece of text “to absorb information”.

“At this stage, the model essentially builds something like a complex map of language and knowledge. This pre-training process involves showing the model text and asking it to predict what comes next. Through many iterations, the model gradually refines its predictions, learning the patterns that make language coherent and meaningful.”

It is trained on what grammatically correct looking text looks like, not what any of it means.

“After pre-training, models undergo additional training called fine-tuning, where they learn to follow instructions, provide helpful responses, and importantly, avoid generating harmful content. This often involves human feedback to improve the model’s performance, as well as reinforcement learning, which uses rewards and penalties to shape the model’s behavior toward being more helpful, honest, and harmless, in the case of Anthropic’s models.”

In order to do this we employ (and underpay) people in the “global south” to review and rate content generated by the AI, exposing these people to terrible content to have them manually mark it as harmful so that “we” don’t have to see it.

Figure 8: Anthropic goals of fine-tuning: Helpful, Honest, Harmless

IT IS NOT HARMLESS. NONE OF THIS IS HARMLESS.

“Once models are trained, they are then deployed for you to interact with. When you interact with Claude […] you’re providing a prompt, which is text that the model reads and then continues from based on pattenrs it learned during training. The model isn’t retrieving pre-written answers from a database. Instead, it’s generating new text that statistically follows from what you’ve written.”

THAT IS WHAT IT DOES.

Figure 9: What makes generative AI powerful: Processes vast information during training and learns complex patterns. Adapts to new tasks through in-context learning. Demonstrates emergent capabilities from scale.

“As these models grow larger, they develop abilities that weren’t explicitly designed into them, sometimes surprising even their creators.”

I cannot believe that these people believe that their statistical word gathering model is somehow developing emergent behaviors via this method, and they speak as if the model has any kind of actual “intelligence”. How do you train intelligence on just words? There is no context for the real world, for physical sensation. There is no thinking in simply crunching tons of human language.

Key takeaways

• .Generative AI creates new content (text, images, code) rather than just analyzing existing data
• .Modern systems like LLMs were made possible by three key developments:
• .Algorithmic and architectural breakthroughs (especially the transformer architecture)
• .Vast amounts of digital training data
• .Dramatic increases in computational power
• .Generative AI learns through two stages: pre-training (analyzing patterns across billions of examples) and fine-tuning (learning to follow instructions and provide helpful responses)
• .Current capabilities include versatility across tasks, conversational awareness, and the ability to connect with external tools
• .Current limitations include knowledge cutoff dates, potential for hallucinations, context window constraints, and challenges with complex reasoning
• .The most effective applications combine human and AI strengths, with humans providing critical thinking, judgment, creativity, and ethical oversight

“You might be amazed at how versatile modern language models can be. They’re skilled with language in ways that seemed impossible just a few years ago. Crafting emails that capture your voice, condensing lengthy reports into clear summaries, translating between languages, and explaining complex topics across countless fields from microbiology to marketing strategy.”

• 404 Media: Wikipedia Bans AI-Generated Content
• 404 Media: AI Translations Are Adding ‘Hallucinations’ to Wikipedia Articles

“What’s particularly notable is how these models can shift between different tasks without needing additional training. The very same system that helps you write poetry or brainstorm ideas for your birthday party can turn around and help you understand quantum computing concepts or analyze quarterly business trends, all through simple conversation.”

I don’t want its help to write poetry, I don’t want its help to brainstorm birthday party ideas, and I don’t trust it to give me correct information about quantum computing (and I am not knowledgeable enough to VERIFY THAT THE INFORMATION IS CORRECT – wasn't one of the LITERACY ITEMS from previously that we assess its output for usefulness and correctness?), and I do not trust it to analyze quarterly business trends or crunch any kind of data because it is going to skip over some things and it is going to generate (“hallucinate”) things that were not there.

“AI models are bounded by their training data. LLMs have a knowledge cutoff date based on when they were trained, the point after which they have no innate knowledge of the world.”

GEE SURE SOUNDS LIKE THIS WOULD BE HANDY FOR ANALYZING QUARTERLY BUSINESS TRENDS WHEN IT CANNOT COMPARE YOUR BUSINESS’ DATA TO CURRENT NEWS AND TRENDING CULTURE.

“The training process doesn’t verify every fact in the training data. This means models can sometimes learn and reproduce inaccuracies that were present in their training data.”

Besides the human-driven verification process in training (if they even know if something is inaccurate, and hopefully don’t miss something), I wouldn’t even claim that “many” or “most” facts are verified. Again, it’s scraping large amounts of text and generating a statistical probability of these words being together.

For example, if there is a lot of posts on the internet about “VACCINES CAUSE AUTISM!” then it will have a statistical bias towards words that generate in this format. There is no “fact checking” going on. And if some misinformation has a large foothold on the Internet (where most of the training data comes from), its representation in the model’s output will be significant.

“They can also make mistakes when trying to piece together information they’ve learned.”

We shouldn’t be using anthropomorphizing language like “learned” for models.

“This leads to what is often called a hallucination. AI confidently stating something that sounds plausible but is actually incorrect. Unlike search engines that simply retrieve existing documents, LLMs generate responses based on statistical patterns, sometimes producing hallucinations.”

Is this one of your great fantastic emergent features that wasn’t explicitly trained in? (I mean no it wasn’t explicitly added to models as a feature but THE ENTIRE WAY THAT LLMs ARE TRAINED AND WORK MEANS THAT HALLUCINATIONS ARE INEVITABLE, A CORE FEATURE.)

“Imagine a friend who tells a story with absolute confidence only to have the details completely wrong. AI can sometimes be like that.”

I don’t want to rely on a tool that is “sometimes like that”.

“Another important constraint is the context window we mentioned earlier. As a reminder, that’s the amount of information an AI can process at one time. Every LLM has a maximum limit to how much information it can consider during a single interaction. If this limit is exceeded, the AI won’t be able to remember that falls outside the window, usually on a first in first out basis. […] This can limit its ability to process large documents or remember the entire conversation.”

Guess you had better pay for the best plan so you don’t get the AI with ADHD.

“Furthermore, unlike traditional software that produces identical outputs given the same inputs, LLMs are somewhat unpredictable by default – also known as non-deterministic.”

WHY WOULD I WANT TO RELY ON SOMETHING THAT IS NOT PREDICTABLE? Best practices in software for DECADES has spent time on AVOIDING UNDEFINED BEHAVIOR. Why would we want to entrust any important information or processes to something unpredictable?

“This variability stems from the nature of how these models generate text. They’re making probabilistic decisions about what text should come next based on patterns in their training data and certain settings that developers can tweak. This creative variability can be great for brainstorming and generating diverse ideas, but requires awareness when consistency or accuracy are critical.”

BUT HERE, TEACH ME QUANTUM PHYSICS IN A SIMPLE CONVERSATION.

“While these models are improving rapidly, they’ve historically shown limitations with complex reasoning tasks. Particularly with mathematical or logical problems requiring multiple steps.”

Because it’s trained on text patterns it has seen before, and we don’t just post tons and tons of math equations and solutions online. It isn’t doing the computations, it’s looking for how many times it’s seen “2+2=4” online and in books.

“The good news is that newer reasoning or extended thinking models specifically designed to think step-by-step are showing strong progress in these areas.”

Oh good, “strong progress”. Coming soon.

“While models like Claude can now access external tools, they may still lack access to specific data sources or specialized tools that would be needed for certain tasks. It’s like having a brilliant colleague who can’t access your company’s internal database. Their ability to help will be limited no matter how smart they are. If a model doesn’t have access to a piece of data or tool that is needed to answer a question, then it should not come as a surprise that it won’t be able to help answer the question.”

So let us tap into your company’s data. Tee-hee.

“Researchers are working to address current limitations through techniques like retrieval augmented generation which connects models to external knowledge and data sources as well as expanding their ability to use tools and improving their reasoning capabilities.”

What isn’t being stated here for the layperson is that “using tools” probably means spinning up a program (a traditional piece of software or script) to do some operation and then return the result. You could also run a program, do some operation, and check the result yourself.

Key takeaways

• .Delegation is about making thoughtful decisions about what work to do yourself, what to do together with AI, or what to let AI handle independently, and how to distribute those tasks.
• .Problem Awareness means clearly understanding your goals and the nature of the work before involving AI.
• .Platform Awareness involves understanding the capabilities and limitations of different AI systems.
• .Task Delegation is the process of thoughtfully distributing work between humans and AI to leverage the strengths of each.
• .Effective delegation requires both domain expertise and an understanding of AI capabilities.
• .The goal isn't to automate everything, but to create the most effective human-AI partnership for any given task or goal.

Figure 10: Understanding the problem you're trying to solve. Understanding the capabilities of available AI tools. Breaking down complex work into smaller parts. Making strategic decisions about who does what. Choosing hte right mode of interaction.

Figure 11: Problem Awareness - THe ability to clearly define your goals and understand what work is needed before involving AI tools. What does "success" look like? What kind of thinking and work is needed to get there? AI fluency begins with and depends on your expertise.

Figure 12: Platform Awareness - A working knowledge of available AI systems and their specific capabilities and limitations. Understanding the unique strengths and limitations of available options. Which models perform best for the work you have in mind? Which systemse prioritize speed, or creativity, or depth, or accuracy?

Figure 13: Task Delegation - The strategic process of dividing work between humans and AI. What could be usefully automated? Where would augmentation create more value than working separately? What should be done by a human alone? What could be done by an agent on your behalf?

“The landscape moves quickly. The best approach is hands-on. Experiment with different AI systems as often as you can and develop your own insights based on personal experience.”

Who needs an expert to figure things out and report on it? Instead, how about you pay for a bunch of products – as often as you can! - so that we can blame you on “not having enough personal experience or insights” if the products don’t work properly. Clearly you didn’t try out enough products.

“Ask yourself: Which specific parts of your workflow would benefit from automation? Or where would an augmentation approach create more value than either working alone or fully automating?”

Please do the mental work of coming up with reasons for YOU to use OUR product. We’re not going to actually present a product and its uses, we will just vaguely suggest things you could do with it – like plan your birthday party.

“The landscape moves quickly. The best approach is hands-on. Experiment with different AI systems as often as you can and develop your own insights based on personal experience.”

Who needs an expert to figure things out and report on it? Instead, how about you pay for a bunch of products – as often as you can! - so that we can blame you on “not having enough personal experience or insights” if the products don’t work properly. Clearly you didn’t try out enough products.

“Ask yourself: Which specific parts of your workflow would benefit from automation? Or where would an augmentation approach create more value than either working alone or fully automating?”

Please do the mental work of coming up with reasons for YOU to use OUR product. We’re not going to actually present a product and its uses, we will just vaguely suggest things you could do with it – like plan your birthday party.