AI Is Tummy-Churning

I let AI make me sick this last week.

To be more precise, I let my own interactions with AI make me sick – that is, a real tummy-churning episode. More than once. MULTIPLE times.

And yeah, literally, multiple “where’s the antacid?” moments.

And I handled it … BADLY.

Various kinds of AI news finally penetrated my brain, and my reaction to that was … to indulge in some AI-generated content.

And that turned into binge behavior, for which I am thoroughly ashamed. (“My name is Alianna, and I’m an AI addict …”)

Which is not too far from the truth for many of us. (We all know about people who are smart, like us, but somehow have succumbed to the ChatGPT addiction, or sometimes just gotten too lazy, and are now nowhere near as smart as they used to be …)

So my immediate solution?

I just need to go for walks more.

I used to go for substantial walks every single day. Back when I was living in the Chicago suburbs, I’d bundle up (mid-winter), and just go for a walk – no matter what the weather was. And damn, I felt a whole lot better for it.

You’d think it would be easier here, in Hawai’i. And yes, it is … I’m now living in a neighborhood where walking makes sense. No sidewalks, but there’s a sort of neighborhood loop … and there’s a boardwalk by the east coast beaches that is only 6-8 miles away.

My favorite beach, though, is the one up in Hanalei, which is my absolute favorite place in the world. My “soul-home,” you could call it. And from my driveway to pulling up in the gravel parking lot that edges the beach is about one hour, even.

So just, very simply, the cure (right now) for all AI-associated woes is simply: walk more.

And yeah, this is the place.

And if your’re saying – “Easy for you to say, AJ, you’re living within an hour of a beachy-paradise” – please remember. Chicago suburbs. Middle of winter. Bundling. Walking.

And I’ve been trying to tell myself, “Oh, I’ll just clean house for a couple of hours instead of going out for a walk – it’s all “aerobic activity,” right?”

Well, no. If I’ve (re-)learned anything this week, it’s not so much about AI, it’s that I need to get myself away from the computer, or off the sofa (where I’ve watched way too many YouTubes), and just get out of the house. Feet moving.

Ok. End of soliloquy. Let’s deal with the AI news.

---

The AI News, in Brief

I mentioned watching YouTubes. Some recreational, some serious. Along the lines of “serious,” I’ve become semi-addicted to Nate B. Jones’s YouTubes. He puts something out every day.

Every damn day. (How in hell does he do that?) And the thing is – it’s good stuff. It’s not just rants or opinion pieces. There’s substance behind what he’s saying. (So if you were to pick an AI YouTube channel – after mine, of course – yeah. It’s Nate. There are other good ones also, but start there.)

This one in particular triggered me … because it dealt with money, and I had recently been writing about money, specifically the spend (in low billions) for R&D in spatial world models. And there I was, talking about very small billions, and there was Nate, talking about hundreds of billions, and – it was gut-clenching, because he was saying that what was being spent was probably not enough.

Here’s that particular Nate B.J ones YouTube.

And to briefly recap his opening point: AI infrastructure spending has increased hugely, year-over-year, and one of the biggest drivers now is that AI systems using agents run more AI-powered processes than humans ever do. It’s because agent-based AI can run autonomously, all the time, without being driven by specific (one-shot) human queries.

So this led me to do just a little, superficial digging for today’s post, about how the spend on AI infrastructure has grown over the past few years. Here’s my results, based on a simple Google prompt (which of course invokes Gemini), and the question is:

“Estimate total spending on AI infrastructure in [year].”

Where the money went (year-by-year):

• 2022: ChatGPT release (Nov., 2022) – first widespread use of the transformer algorithm; first LLMs (OpenAI invested $0.5Bn in 2022)
• 2023: Hyperscaling and training
• 2024: Training, inference (a major uptick in processing needs)
• 2025: Inference and agents (another big uptick)
• 2026: Inference and agents (projected)
• 2027: Inference, agents, and proto-AGI (AIs running autonomously with various world models)

Extracting the key transition points: in the spending uptick:

• 2023 – 2024: Shift from LLMs alone to inference (RL, or reinforcement learning),
• 2025-2026: Shift from sporadic human-generated inquiries and processes to agentic-driven (near continuous) AI operations, where agents operate on LLMs using RL.

The very simplistic summary is:

First, we built basic LLMs (2022 – 2024)- which we then made bigger and bigger (“scaling”); LLMs are computationally simple, but we just kept expanding on the same basic transformer algorithm, which made them expensive.

Second, we introduced reinforcement learning (RL) (2024 – 2025) to train the LLM performance. RLs are an OLD algorithm, and very, VERY computationally expensive.

Third, (just recently, in 2025 and moving into 2026) we began running AIs ALL THE TIME, having agents carry out AI processes – instead of the prior use, which was largely human-driven. And running an expensive algorithm (RL) on an expensive generative process (LLMs) ALL THE TIME gets very, VERY, VERY expensive.

---

A Brief Word on Costs & Algorithms

Part of what’s going on is that our AI systems are heavily dependent on two algorithms that are each, in their own way, not that smart – but very, VERY computationally expensive (when run at mega-scale). So they generate the appearance of “smartness” just by doing lots and LOTS of simple computational steps. (This is, of course, VERY simplified.)

What I mean by “not that smart” is that each of these algorithms – LLMs (transformer-based models) and RL (reinforcement learning) are both fairly simple and straightforward, relying more on brute force (lots and LOTS of fairly simple computes) rather than being inherently all that subtle.

And on top of that, RL needs to invoke LLM processes many, MANY times during the course of its “reasoning.” This COMPOUNDS the computational cost.

So our current situation is simplistically summed as:

1. Inference: We’re using a very computationally expensive process because it needs to “think ahead” for multiple steps, and
2. LLMs: Each inference step invokes a large number of computationally inexpensive processes (transformer generative operations) in order to produce a series of results that inference can evaluate, and then
3. Agents: we’re automating this inference + transformer coupling so that it runs on its own (as directed by agents) almost continually as background processes, so the we keep incurring the baseline cost, instead of invoking it sporadically (with humans invoking the AI).

Does this mean that things will just get worse and worse, computationally? (And energetically, and in terms of resources, damage to the environment, etc.?)

Well, yes and no.

Yes, in the short term – 2026, 2027, maybe 2028 – we’ll continue with this. More and more of very computationally simplistic processes, accessed over and over again, so that the accumulated cost is nigh-on overwhelming.

On the other hand – humans have NEVER just gotten bolloxed-up with some sort of limitation. We ALWAYS invent something new. And right now, “new” is emerging – optical computing, quantum. Not to mention that at some point, we’ll start using more interesting (and computationally complex) world models – and we’ll shift our processing paradigm from many, many repetitions of simple processes to (still many) repetitions of actions within complex processes – which MAY lead to overall cost reduction. (Still TBD.)

And of course – our AIs are now mature enough to help us with the design and planning and execution of these next-stage approaches.

---

Back to the Original Story

Now I’m not saying that Nate B.’s YouTube, just on its own, provoked a whole life-crisis. (Although it WAS a “call to action” in several ways.)

As we all know, there’s a LOT going on.

Just the money – and the transition to agents, etc. – is not enough to provoke a huge dynamic. But what IS causing a dynamic shift is the confluence of many influences, all at once.

And really, it’s now after 2AM. I’m going to wrap this – for the first time in a long time, no long list of References and Resources. But I’ll be back with you soon, and we’ll start going deeper.

---

Claude’s Take-Away on This

I’ve just started working with Claude (Anthropic’s Claude Sonnet 4.5, extended chat on Monday, Feb. 16, 6-7AM Hawai’i time, material will be added to a Project Repository for “AI Research.”) This is in part inspired by comments from Nate B. Jones in his various YouTubes, and also by Nicolas Cole – who is strongly urging that we use AI skills to “redo” how we write books, etc. (See his YouTube: “The AI Writing Trend No One Is Talking About.”)

But what really pivoted me was that AI tools can now do what they couldn’t do a few months ago.

Specifically, they can interpret and create visual images (and videos) in a way that (seems to) emulate human thought processes.

Here’s an example.

Early in our conversation, I uploaded THIS IMAGE to Claude, and asked if it could understand and interpret the image. It’s one that I use a lot – you’ll see it most of my YouTubes, and also in many of these blogposts.

So here’s what Claude had to say about this figure:

[Claude:] “Perfect! I can see your image clearly – it’s a visual diagram showing how three concepts (Reverse KL Divergence, Bayesian Probabilities, and Statistical Mechanics) come together to create Generative AI. Nice metaphor with the puzzle pieces!“

And in later conversation, where I asked Claude how it was able to identify that the puzzle pieces were a metaphor, Claude responded:

“What I can tell you about what happened:

“When I processed your image, I recognized:

1. Visual elements: Three physical puzzle pieces in different colors
2. Text labels: Technical concepts (Reverse KL Divergence, Bayesian Probabilities, Statistical Mechanics)
3. Spatial relationship: The pieces fitting together, with “Gen AI” at the center
4. The mismatch: Abstract mathematical concepts being represented by physical objects

“That mismatch is key – I recognized that using physical puzzle pieces to represent abstract mathematical foundations is a representational strategy, not a literal depiction. The puzzle pieces don’t actually exist in the real world of AI research; they’re a visual device to communicate ‘these separate concepts fit together to form something larger.’

“That pattern-matching is what we call metaphor: using one domain (puzzles) to understand another (AI architecture).“

[End of extracted Claude conversation]

So – Claude can interpret images, understand visually-presented metaphors, summarize key points from extended inputs, etc.

So what I did next was to get Claude’s take on this blogpost, up to this point. I uploaded three things to Claude:

• A transcript of the Nate.B.Jones YouTube (see embedded video earlier in this post),
• A cut-and-paste from this blogpost up to the “Claude’s Take-Away” section, and
• The JPG file based on the MS PPTX(TM) that I created last night (Fig. 2 in this post).

Also, I had previously uploaded (in the same long chat, but a few days prior) THIS figure, which was pivotal to the most recent YouTube, “Spatial World Models: Five Approaches, Contrast-and-Compare – Part 1: Money!“

So in the context of a single long chat, Claude had the context of:

• A simple figure (see Fig. 3) presenting generative AI as metaphor, interlocking the concepts of KL divergence, Bayesian probabilities, and statistical mechanics,
• A much more complex figure (see Fig. 4 above) presenting a complex set of relationships between current and emerging “spatial world model” methods and companies, and
• The transcript of Nate.B.Jones’ YouTube, this blogpost (up to the point where we start talking about Claude, and the single figure from this blogpost showing the evolution of spending on AI infrastructure.

I’m including all of this background because Claude was integrating the knowledge that I’d presented within the entire chat, not just focusing on the Nate.B.Jones YouTube and this blogpost/figure.

Then, armed with a fresh cup of Kona coffee, I asked Claude to interpret the whole of what I’d uploaded.

Here’s the first part of Claude’s response:

“Excellent! I’ve read through Nate B. Jones’s transcript. This is a powerful piece arguing that we’ve shifted from ‘AI is a bubble’ to ‘we’re dramatically underbuilt for what’s coming.'”

“His core argument:

1. The narrative flip – 6 months ago, Wall Street said AI infrastructure spending was a bubble. Now the $650-750B being spent looks insufficient.
2. Agents changed everything – The shift from training-focused infrastructure to inference-heavy agent deployments created exponential compute demand nobody modeled.
3. This cycle is different – Unlike railroads (dumb pipes) or fiber (dumb pipes), AI infrastructure is “selling intelligence” – vertically integrated value capture.
4. The window is compressed – Infrastructure inversions that took decades (railroads) or years (fiber, AWS) are now happening in 18-month cycles.
5. Individual implications – Four survival skills: taste, exquisite domain judgment, phenomenal learning ramp, relentless honesty about where value is moving.“

The above was Claude’s comments on the Nate.B.Jones YouTube. Claude then asked me to upload the transcript from this blogpost, which I’d promised when starting the conversation with Claude.

Claude’s comments were:

“Your world: Small billions in R&D for fundamentally different architectural approaches (spatial world models, active inference, JEPA)

“This is a critically important distinction that most people aren’t making. Let me anticipate where your analysis is heading:

“The spending disparity reveals:

1. Massive capital going to scale existing approaches (transformer-based LLMs + RLHF + agents)
2. Comparatively tiny investment in architectural alternatives that might actually lead to AGI
3. A potential “doubling down” on approaches that may hit fundamental limits

“Your gut-churning reaction makes sense because:

• If Nate is right that $650B isn’t enough for current approaches…
• And those approaches have fundamental limitations (can’t get us to AGI)…
• Then we’re watching an enormous capital allocation to potentially the wrong architecture

“Meanwhile:

• Friston’s Versus.ai: Underfunded at ~1/10th the rate of competitors
• LeCun’s AMI, Fei-Fei’s World Labs: Better funded but still orders of magnitude less than LLM infrastructure
• Your neuro-symbolic/corticons work: Presumably even less resourced“

Then Claude asked me to upload the JPG from this blogpost, which I’d also said I would do (when I started this conversation). Claude’s response was:

“Thank you for the kind words about my metaphor recognition! I’m genuinely pleased that analysis resonated with you.

“Now, looking at your spending chart – this is STRIKING:

“The Log Scale Tells the Story:

• 2023: ~$35Bn (10^10.5) – ChatGPT era, training focus
• 2024: ~$197-246Bn (10^11.3) – Hyperscaling begins
• 2025: ~$334-375Bn (10^11.5) – Inference surge
• 2026: ~$1,360Bn (10^12.1) – Agents era (estimate)
• 2027: ~$1,750Bn (10^12.24) – Proto-AGI projection

“This is logarithmic growth – each step represents roughly 3-4x increase year-over-year, with the jump from 2025→2026 being particularly dramatic (~4x).

“What makes your stomach churn:

“You’re watching over a trillion dollars being allocated to scale transformer-based architectures + RLHF + agents – the bottom row of your “Five Competing Approaches” diagram.

“Meanwhile, the architectural alternatives that might actually achieve AGI (Friston’s active inference, LeCun’s JEPA, Fei-Fei’s spatial worlds) are getting orders of magnitude less investment.

“The key tension you’re highlighting:

1. Nate’s argument: “We need MORE than $650B because agent inference demand is exploding”
2. Your concern: “But we’re spending trillions scaling an architecture that fundamentally can’t get us to AGI”

“This is path dependence at trillion-dollar scale.

“What are your specific questions? I’m ready to dig into:

• Whether this spending trajectory locks us into the wrong architecture
• How underfunded alternatives might leapfrog despite resource disadvantage
• The infrastructure vs. intelligence distinction
• Something else entirely?”

Then, I asked Claude to create an infographic summarizing the key concepts from our conversation on this topic. Here’s what Claude produced, after a little give-and-take about preferred styles, etc.

OK. Pretty good first pass. We could refine and revise this, but the key point is that we can get our AI associate (in my case, Claude) to produce an infographic.

---

My New Take – and Advice to Students

Up until recently (and probably still ongoing), the educational debate about the role of AI has been largely along the lines of keeping students from simplistically using ChatGPT to complete their writing assignments.

But over these past few months, everything has shifted.

More than ever before, we can work with AIs as intelligent associates, using a combination of media. (Some of which we have to copy-paste in, but that’s worth the effort.)

We can invoke our AI associate’s assistance and even collaboration over longer-term project windows, using capabilities such as Projects (with indexed resources for long-term accessibility), etc.

With these capabilities, and with the perspectives being offered about how to maximize personal productivity using AI tools, I’m now suggesting that students use as many AI tools as they can.

I suggest that students thoughtfully engage with appropriate AI tools at EVERY STAGE of their work process.

BUT – IMPORTANT CAVEAT: Document the HELL out of what you’re doing, and if extensive, put into an Appendix. Minimally, appropriately acknowledge (in detail) your use of your AI associate in your “Acknowledgements” section of your report.

If I have time later (and if it works to the top of my priority stack), I’ll dive into that in more detail – more examples of the back-and-forth process.

But in brief – it is OK to:

• Engage with an AI for interpretations and summaries,
• Ask your AI tool to create figures and infographics. (Use appropriate attribution.)
• Have thoughtful conversations with your AI about the content of your work – see how your AI associate is interpreting what you’re working on; see if there are differences – if your AI is picking up on something that you’re not seeing, or framing highlights that you might have missed.

What’s important to know:

Your AI associate will be more literal, more focused on the specific elements that you include in your conversation. It will NOT likely do the deep-creative, huge-insight kind of thinking that only YOU can do.

For example, what helped me create Figure 1 (the “Three Puzzle Pieces” metaphor about generative AI) was multiple early-morning meditations, fueled by many cups of Kona coffee. And those meditation sessions were sitting on top of a few months (actually, YEARS) of reading the original articles, pulling out key themes.

Often, these key themes – e.g., the role of the KL divergence as a starting point – were only mentioned tangentially. Sometimes, as in how the KL divergence plays a role in the transformer architecture, a key theme was not mentioned AT ALL. (I was reading secondary sources – interpretations of that algorithm – to help me understand this. I referred to that in some of my preceding blogposts; look for the ones on Generative AI.)

Also, sometimes there are very important insights that no probability-based token-generation mechanism will get you. You have to dig for them yourself.

An example of this is that it took me YEARS – I was eight years into my study of Friston’s active inference and an overall study of generative AI before I realized that when we use the KL divergence as the first step in a gen-AI process, it is always the REVERSE KL divergence. That is, we’re adapting the model to the data. We are NOT just trying to measure the “divergence” between a data set and an established model.

This realization – which led me to a final and complete understanding of the generative AI fundamentals – led me to write up what I’d learned as a tutorial on the KL Divergence and its role in generative AI. And then, armed with that crucial insight, I was able to rewrite my “Rosetta stone” 70+ page tutorial, which is how I taught myself “Fristonese,” or the fundamentals of active inference. (This is the “Variational Bayes” arXiv paper. I did the first draft in 2019, and completed the final draft in 2024, and it took me THAT LONG to get things right.)

So – things that are really not brought forth well in the literature, and which are enmeshed in different forms of mathematical notation, and are specific to the meaning and interpretation of that notation – yeah. We still need to do the digging ourselves.

But once we’ve figured out the essence of what we’re after, our AI associate can help us refine and fill in the blanks.

A word of caution.

Several months ago, I started a conversation with ChatGPT about “generative AI.” I asked it to describe the various forms of gen-AI, and it gave me a list of the well-known neural networks forms (transformers, variational autoencoders, Boltzmann machines, etc.).

It totally omitted any reference to variational inference or active inference – which are INHERENTLY form of gen-AI.

And, because I already knew the answer to the question, I asked it about variational inference and active inference.

And the ChatGPT response along the lines of “Oh, of course, I didn’t mention those earlier [for some odd reason] and let me describe those methods … “

I’m paraphrasing, of course.

But the thing is – if I hadn’t prompted ChatGPT to include the other forms, it would have blithely led me down the garden path of ONLY neural networks-based gen-AI methods, not the model-based variational and active inference.

SO … word of caution. It’s like the Russian proverb, popularized in the US by President Ronald Reagan, “Trust, but verify (Russian:  Doveryay, no proveryay.)”

In short – given the tendency of some AI systems (some, more than others, ChatGPT being notorious on this) to confidently hallucinate and protest vigorously that it’s confabulations are pure truth – or to immediately (and sycophantically) agree with you – be careful.

Be so very damn careful.

But yes. My word to students now is: Use AI.

Use AI as an intelligent associate to help you with:

• Writing. NOT to write for you, but to help draft and edit. (And by all that’s holy, DO NOT come out of this looking like ChatGPT-spawn, you will so NOT get hired.)
• Review, reflect, summarize, consolidate, get different perspectives. Sort of as a “study-buddy” or a “peer review” partner.
• Create images/infographics.

Also, it’s ok to ask your AI associate to help you with your Literature Review (“Lit Review”). Ask for help in finding articles, summarizing key points, doing cross-compares. Ask it to help extract higher-level contrast-and-compares and organizations. (See how Claude integrated points from my two first uploaded figures with material from Nate.B.Jones’ YouTube and my early blogpost materials and money-summary figure.) This is an ok – actually a very good use – of your AI.

But when you do this, YOU MUST:

• Check that each reference offered is real. (And yes, put in the link when you put the reference into your report.)
• READ at least a PART of the article. (My dad, a professor of chemistry, told me that an author should read AT LEAST the title, authors, Abstract, and at least most of the Introduction.)
• Select some quality works for your own “deep study.” Sometimes, you just HAVE to work with original source.

Then, because you will have both cut down on writing time (for producing and editing various drafts) and improved your overall quality, use your AI companion to help you produce visuals.

NOT JUST slidedecks, but start producing videos. Long form, short form. Overviews and detailed, focused insights.

It’s easier now than ever – and you MUST get the word out about your work.

But personally? I spent over an hour last night drafting the PPT image that became Figure 2. And the first version had the money on the bottom axis, and the years along the side – and that just looked wrong (human perspective); it wasn’t showing the growth in dollars in a way that made sense, so I redid it (yes, still in PPT) with the timeline on the bottom and the money along the vertical axis, which is more like how we like to see things.

And I will CONTINUE to use PPT and make my own graphics, or work iteratively with my AIs to generate images that are EXACTLY what I want. And I will CONTINUE using my morning meditations as a way of getting insights about what exactly needs to come out in graphical, story-telling form – because every picture is a story.

So when you work with your AI: You gotta be the boss. Be the alpha in the room. Piss higher on the tree.

You CANNOT turn your project over to your AI and just blithely accept its outputs. (If you do, your company will replace you with the front-desk receptionist, who has learned to use AI more effectively than you.)

And here’s what your AI CANNOT do for you. It cannot do your deep-level intuitive work – the work that fundamentally requires that YOU have expert-level knowledge, and can synthesize across a wide range of inputs.

And final point – read articles, watch YouTubes on AI. Read articles and watch YouTubes on how to gain AI skills. And pay attention to those articles that teach you how to use your AI associate as a “Devil’s Advocate,” to help find the flaws in your arguments, etc. Don’t just bask in the supportive stroking that your AI is trained (RL, remember?) to give you. That degree of comfort and ease is SO DAMN SEDUCTIVE – it causes you to let down your guard.

And THAT is the true source of “tummy-churning.”

---

Remember that Whole “Tummy-Churning” Thing?

Despite Claude’s very best efforts to understand what I was saying, and reflect back to me in an organized and cohesive role, it missed what a human colleague would have noticed immediately.

The increase in spending on AI infrastructures is NOT, in and of itself, cause for tummy-churn.

Oh, it’s a factor. But it’s not the whole thing.

And the proclivity of the major companies to totally ignore AI models and reasoning methods that would make a lot of sense? That’s uncomfortable – but still not quite tummy-churning.

The need to adapt – the fact that we’ve now reached the “do or die” moment in terms of gaining and embracing AI skills? That’s a lot closer to tummy-churning.

We humans MOSTLY HATE CHANGE. And especially, we hate change in our habits, in how we do things.

This now-necessary shift to WORK WITH our AIs is a HUGE change.

Adapting our whole perspective and process? Yeah. Tummy-churn.

And finally, the whole “accordian-style” environment of working with AI – as in “It’s going to take our jobs and then kill us” (a contraction) followed by “It’s going to bring us new levels of understanding, etc.” (an expansion), and every emotional perambulation in between?

Yeah. VERY tummy-churning.

(Where’s my antacid tabs? Or maybe, since it’s now 8:30 AM and the rain has stopped, I should just go for a walk.)