Merged Artificial Intelligence

[JayUtah]

This has been an issue for a week here in Utah. The Box Elder county commission voted yesterday unanimously to approve a 9-gigawatt, 62 square-mile (160 square-kilometer), hyperscale AI data center near the north shore of the Great Salt Lake. There are too many irregularities with this project to even summarize. As has become commonplace in Utah business and politics, the deal was all worked out in back rooms and foisted on the public. No public comment was allowed. The project is authorized under military authority, so it bypasses almost all environmental assessments on "national security" grounds. Yet it's being built by and for billionaire Kevin O'Leary.

9 gigawatts (that's about 7 Marty McFly time-jumps) is twice as much power as the entire state of Utah presently uses. The project includes a gas-fired powerplant to generate all the power for the data center. Utah has a serious pollution problem owing in part to our unique microgeography and microclimate and in part to large-scale mining operations southwest of the lake. Since the environmental assessment for this project is just a lot of handwaving, we aren't able to assess the pollutant load from the plant.

62 square miles is bigger than many entire Utah cities. The proposed campus will be roughly the size of St. George, Utah. Now Utah is home to a lot of empty space. But the proposed site is actually near inhabited areas whose inhabitants have had zero say over whether they want this in their backyards.

Utah is a water-poor state. Its signature lake is projected to dry up entirely in less than a decade because of commercial draws from its watershed. Our governor routinely issues drought emergency orders curtailing the use of water. Yet the same governor is who pushed this project through back channels and sprang it on the county officials with little warning. The O'Leary group assures people that it will used closed-loop water cooling, but closed-loop is not net-zero consumption. Nor does the statement include the power plant and its cooling requirements. When people pressed about the proposed water usage, we got no hard numbers and a whole lot of, "Trust us, bro."

The lack of transparency in this massive project has been egregious. Naturally people are upset at this having been foisted on them basically over only a two-week period. And then when they protest (because that's all they have left), they're dismissed as unruly.

Box Elder County commissioners OK controversial data center proposal

Box Elder County commissioners approved a controversial data center proposal at a special meeting that drew hundreds of opponents of the proposal.

www.ksl.com

 

[JayUtah]

To follow up: Kevin O'Leary—with no evidence—claims the protesters at the county commission meeting were paid and/or bussed in from other states. (They weren't.) He also claims—again without evidence—the visuals were AI-generated. (They weren't.) The right does this every time we post pictures and videos of large-scale protests agains their actions. The guy telling these obvious lies is the same one asking us to trust him that his hyperscale industrial park will follow all the rules. The people protesting are the thousands of people who are getting a hyperscale industrial park right next to their longstanding semirural homes and never had a single say in the process.

Gov. Spencer Cox excuses the skipped environment and economic assessments by saying he's tired of projects taking years to come to fruition. Apparently impatience is now a virtue. Utah has struggled for years with air quality, energy self-sufficiency, and water use. And the biggest impediment to a solution has been the Republican supermajority that constantly prioritizes industry over sustainability. Industrial farming consumes between 75-89% of our annual fresh water supply, which is why the lake is drying up. The farms produce alfalfa, which is sold overseas as a cash crop that benefits a few growers. Gov. Cox is an alfalfa farmer.

I have twenty years' experience operating data centers in a desert. Yes, we too use closed-loop cooling. Our installations are much, much smaller and less dense than the hyperscale projects that grab headlines. That's because we include sustainability in our planning and don't outstrip our valley's capacity for electricity and heat rejection.

Closed-loop cooling circulates water through indoor heat exchangers where they pick up heat from the servers, then take the hot water outside for forced-convection air cooling via a large outdoor heat exchanger. This works only when there is a substantial difference between the process fluid entering the outside exchanger and the ambient air temperature. During summer, cooling efficiency drops dramatically. When that happens, and you need to keep operating, you often need to use evaporative cooling: water misters on the outside exchanger pipes. That water is lost and must be accounted for in sustainability calculations. There are other techniques to improve the efficiency without using water. But they are marginal improvements only.

There's the cooling requirement to dissipate the heat from 9 gigawatts of electricity coursing through their GPUs. Conversion coefficients for high-density server farms hover around 0.95, meaning that 95 thermal watts are produced for every 100 watts of electrical power consumed. That's going to dump 8.5 gigawatts of thermal energy into Box Elder County. That's enough to create its own microclimate. 1 GW heat load per square kilometer adds 80% of the energy you get from solar influx. The problem we have in Utah is temperature inversions, where a "cap" of cold air holds in a stagnant body of hot air trapped beneath it. Dumping gigawatts of thermal energy into this system is a potential economic and health disaster.

And that's just the data center. The power plant seems to be a lot of handwaving. Assuming a combined-cycle gas turbine design that operates at about 60% efficiency (and a bunch of other quickie, back-of-the-envelope calculations), a 9-gigawatt power plant will need a cooling system that rejects about 2 gigawatts of waste heat to the ambient. This would all have been studied in the environmental impact studies that Gov. Cox says take too much time and therefore need "national security" priority to overcome. (So why is a private interest in charge instead of a defense contracting process?)

They can't tell us how much water they'll need to pull from the Great Salt Lake watershed to maintain cooling efficiency in their closed-loop system during the hot summer months. This is a basic sustainability parameter, and—again—would normally be something you'd have to justify in an environmental impact assessment. And as I said, the guy who promises everything will just work out is the same guy inventing easily-debunked conspiracy theories to explain why people are amazingly angry at him and at our elected leaders for foisting this on us without our knowledge and consent.

When the AI bubble bursts, we're going to have a huge, abandoned industrial complex that will then require decommissioning, deconstruction, and cleanup, likely at public expense. Or I guess they'll just turn it into an ICE detention camp.

 

[arthwollipot]

NSFW - sweary word

 

[Tero]

My discussion with Google AI and a dozen books explaining AI processes led me to this. The weights and the "meaning" it gets from words by their usage and connections has finally convinded me of one thing. The problem I had with undesrtanding the brain was that I dismissed simple things like Kandel's learning experiments (see https://en.wikipedia.org/wiki/Eric_Kandel) as just crude reactions to survive. I kept asking where the images and meaning was. Why do we see color? All that was not compatible in my mind with just neurons firing.

But it is. I think about something. Some memories come back and I use that information in my frontal lobe to decide something. My brain has learned to pull those meanings from my memory the same way the AI weights do. It may not be numbers, but the signals sent back to my conscious mind present the idea. The neurons need to fire together to get the idea back to "me." The idea is simply just neurosn firing. This took me forever to grasp. I assume competent psychologists and neuroscientists knew this before Chat GPT.

It does leave some mysteries about some of the senses. How did we learn to taste things as tasty or disgusting? other qualia as well. Some of it has to be "hard wired."

 

[JayUtah]

I'm pretty sure a GPU has never tasted an apple. An LLM's entire concept of "apple" is based entirely on its contrived statistical association with other words such as "crisp," "sweet," "fruit," "pie," and "...of my eye." There was never any sensory experience. Your brain encodes the sensory experience of an apple in a way that's accessible as a sensory memory. This is vastly different from how LLMs attempt to encode meaning.

My understanding of neuroscience suggests that memories in the human brain are stored in a high-dimension encoding that could be said to resemble how LLM embeddings work. But the big difference is that your brain is plastic. Your "embeddings" change markedly over time, whereas an LLM is trained in one step and then used in another step. The knowledge based encoded in the LLM embeddings is largely static once the training is complete. If you wanted to compare that to human intelligence, it would be like not being able to remember or learn anything new after the age of about twelve.

 

[Tero]

My understanding of neuroscience suggests that memories in the human brain are stored in a high-dimension encoding that could be said to resemble how LLM embeddings work. But the big difference is that your brain is plastic. Your "embeddings" change markedly over time, whereas an LLM is trained in one step and then used in another step. The knowledge based encoded in the LLM embeddings is largely static once the training is complete. If you wanted to compare that to human intelligence, it would be like not being able to remember or learn anything new after the age of about twelve.

Yes yes the qualia. And we reproduce so there is all that.

But the LLM's will be self educating in the future, they will not be stuck with x billion nodes. They will improve their reasoning as well. But the connection to the physical world will only be through us, so that will be a limit. They will be the absent minded professor that thinks complex things but will not notice changes in everyday things.

 

[Darat]

....snip.... If you wanted to compare that to human intelligence, it would be like not being able to remember or learn anything new after the age of about twelve.

Hmmm - looks around.... hmmm

 

[Helen]

Hmmm - looks around.... hmmm

Speak for yourself. I learnt to drive at 18, and that's many years after 12. I'd say ....10 years? I think I need to raise my IQ.

 

[JayUtah]

And we reproduce so there is all that.

Ugh, don't give them ideas.

But the LLM's will be self educating in the future, they will not be stuck with x billion nodes. They will improve their reasoning as well. But the connection to the physical world will only be through us, so that will be a limit. They will be the absent minded professor that thinks complex things but will not notice changes in everyday things.

Believe me, the past two weeks have been a harsh lesson for me and my colleagues in the scale of LLM-builders' ambitions. If you want to argue that methods will improve and that LLM output will more and more closely resemble that of a human, I won't argue. But on the other hand if you want to argue the fundamental difference between what a human brain encodes (sensory evaluation) and what an LLM codes (statistically-expressed sentences), then I think that difference will always be there.

 

[JayUtah]

I think I need to raise my IQ.

You're in luck. I know a guy.

 

[Tero]

It was not do much the LLMs I was interested in, it was the human. I could mot for years grasp the neurons working. Only the picture of the finger touching the hot stove made sense to me. Signal goes to spine, motor neuron moves the hand. The thinking part made little sense. But since we know AI can produce a sentence, our neurons must store information in some similar manner. I am not at my laptop, I can quote Google AI on human memory. I saved it.

 

[JayUtah]

But since we know AI can produce a sentence, our neurons must store information in some similar manner. I am not at my laptop, I can quote Google AI on human memory. I saved it.

You might be onto something. If someone asks you what an apple tastes like, you will form your answer in a sentence in your language based on memories you've previously encoded. However, the language center of the human brain is kind of its own animal. This is why people talk to themselves: it forces their brains to reorganize nebulous thoughts in a more concrete form that can then be expressed as language elements. It's a two-step process. I can speak a number of languages, but my aversion to broccoli and touching a hot stove remain independent of how I express them verbally. In LLMs something like this is performed by the decoder transformer layer.

 

[Skeptical Greg]

If someone asks you what an apple tastes like, you will form your answer in a sentence in your language based on memories you've previously encoded.

If someone asks me what an apple tastes like, all I can say is " an apple" ..
I think AI can handle thst.

 

[Wudang]

If someone asks me what an apple tastes like, all I can say is " an apple" ..
I think AI can handle thst.

Ah not necessarily because your Theory_of_mind^WP comes into play and depending on who you're talking to you may say "Remember those sweets you liked that....." or "like those melons a bit" etc.

 

[Wudang]

Also if I remember correctly human memories are retrieved in read/write mode and are partly reconstructed and saved. So a far more dynamic and modifiable process.

 

[JayUtah]

If someone asks me what an apple tastes like, all I can say is " an apple" ..

That's the most accurate answer, but the exercise is to describe it to someone who knows generally what it's like to eat something and taste it, but who simply hasn't done so yet for an apple. You would most likely want to describe its texture with texture-expressing words, its temperature, taste notes like "sweet" or "tangy." If you could manage to do that, I would argue that your mind is searching through memories encoded as the memory of a sensation.

I think AI can handle that.

How an LLM handles it is to do a massive parallel search on words that are commonly found alongside "apple" and "taste." Then it incorporates those words into a sentence in the target language. What it's doing is reporting what everyone else has said an apple tastes like. @Wudang is basically saying the same thing from a theory-of-mind perspective.

 

[JayUtah]

Also if I remember correctly human memories are retrieved in read/write mode and are partly reconstructed and saved. So a far more dynamic and modifiable process.

Neuroplasticity is a blessing and a curse. It allows us to learn, but it forbids us from remembering things strictly accurately.

 

[Tero]

Google AI pulled out these:

As of 2026, the most accurate and widely accepted theoretical concept of human memory is that it is a Constructive and Reconstructive Process, rather than a static storage system. Memory is viewed as a dynamic mechanism used to simulate possible futures and make sense of the present, rather than simply playing back the past.
link: https://pmc.ncbi.nlm.nih.gov/articles/PMC4679162/

Key Theoretical ComponentsConstructive Memory Framework: Memories are not saved as exact files, but are reconstructed based on beliefs, feelings, and memory fragments."Memory as Data" for Future Simulation: A 2026 perspective emphasizes that memory is a source of data for the brain to make predictions about the future, rather than just a repository of the past.

Systems Consolidation (New View): A modern, quantitative theory of systems consolidation proposes that memories are transferred from the hippocampus to the neocortex only if they improve generalization—meaning we store the "gist" or predictable elements rather than every detail.Neural Overlap: Recent 2026 findings indicate that episodic (past events) and semantic (facts) memories share overlapping neural networks, indicating that the brain treats different kinds of remembering similarly, rather than in separate, isolated modules.
"What" and "Where" Separation: Research shows the brain stores "what" occurred and "where/when" it occurred separately, connecting them only when a memory is retrieved.

Key Scientific Principles Driving Modern Theory Levels of Processing: Memory is stronger when it is deeply encoded based on meaning rather than shallowly encoded.Reconsolidation: Every time a memory is recalled, it can be modified, allowing it to be updated or distorted.Pruning & Prewiring: The brain's memory center (hippocampus) does not start as a blank slate, but is prewired with dense neural networks that are refined by pruning over time.

AI vs human brain

How the Connection WorksIn both systems, the goal is to adjust the "strength" of connections between nodes to improve how information is sorted or recognized.Biological Synapse: The "weight" is a complex physical reality. It depends on the amount of neurotransmitters released, the number of receptors on the receiving neuron, and even physical changes like the thickness of the connection.AI Node: The "weight" is a single numerical value. It determines how much of an input signal's importance is passed to the next layer.

Key Similarities in "Sorting"Pattern Recognition: Neither system stores an image as a single file. Instead, they "sort" information by breaking it into features (like edges, textures, or colors). If the combined signals passing through specific weighted paths exceed a certain threshold, the system "recognizes" the object (e.g., "This is a cat").Plasticity vs. Training: In your brain, Long-Term Potentiation (LTP) strengthens synapses based on experience. In AI, Backpropagation adjusts numerical weights based on errors found during training to achieve the same result.

Critical Differences
While the inspiration is identical, the execution varies significantly:Learning Mechanism: The brain uses Hebbian learning ("cells that fire together, wire together"), which is a local process. Modern AI uses Backpropagation, a global mathematical process that re-calculates every weight in the network simultaneously based on a final error—a process your brain is not known to do.

 

[Darat]

Someone else not finding evidence of the infrastructure build out that has been promised (they are obviously a smart person as they share my concerns....)

 

[Tero]

In my most recent version of understanding what we are, I have borrowed heavily from AI thinking. Our brain evolved for generation to process information. It sends it to the frontal cortex and that in turn explains to "you" what the situation is. The "you" also has buffered smoothed information from the world continuously. Frontal Lobe just tweaks it. In the same way that AI cannot explain to you how it got the answer, your frontal lobe cannot explain how the info in the brain is processed. neurons firing end up with an idea, image wahtever to the frontal cortex. it cannot explain how it got there. It just generates the thoughts that conscious "you" can understand.