Merged Artificial Intelligence

[Wudang]

Interesting take

https://bsky.app/profile/djdynamic.ca%2Fpost%2F3mmug5iv6sk25

 

[Beelzebuddy]

Billionaires have every reason to believe they are not subject to the rules because effectively they aren't.

This project was almost entirely secret until it was submitted to the county commission—three part-time gentlemen who generally have no need to dabble in politics. Box Elder County is a quiet, semi-rural area. The most exciting thing there is the factory where they build the solid rocket boosters for the SLS rocket (and lots of other rocket-related things). The second-most exciting thing is a natural hot springs that has been developed into a popular resort experience. What towns it has are picture-perfect middle America: kids riding their bikes down Main Street without a care in the world; diners where you can still get a bottomless cup of coffee; bunting on the courthouse for the 4th of July. The rowdiest party is the Peach Days festival.

It's pretty clear the powers behind it wanted to push the data center through under the radar. All the "approvals" had come from MIDA, a small, innocuous state committee whose job formerly was to think of ways to develop underutilized land associated with the military. Because the data center was given "national security" status (despite being built by and for a private interest), most of the customary reviews and impact studies were simply redlined out. And because it was "national security," MIDA discussed the project entirely behind closed doors in "executive session," an exception to Utah's Open Meetings law. Normally MIDA proceedings must be open to the public. When the final approval landed on the commission's public agenda, there was a giant collective, "Wait! WTF!?" And frankly, the county commissioners issued their own, "Wait! WTF!?" They hadn't been told about it in advance. They were evidently just supposed to rubber-stamp it.

Yeah, a funny thing happens when you tiptoe around the public to build a giant industrial complex in their backyard. It's bad enough that the biggest project of its kind in the world bypassed all the normal impact studies. But then even when the mask was ripped off the plans, no public comment was accepted. You don't do that to Utahns. Utahns are fiercely tied to the land in a way people like O'Leary can never understand. When you just elbow your way into their lives and flatly tell them, "We don't care what you think," you're going to get a lot of backlash. And no, if the public had been allowed to weigh in on whether they wanted O'Leary's dream project on their land, the answer would have been "Hell, no!" Any Utah politician could have told him that, but a small cabal of power brokers went ahead and tried to do it anyway.

Billionaires and their allies on the political right have mastered the art of behaving outrageously and then making hay out of the outrage. Gov. Cox at first said he was tired of all the red tape and how long it took to get anything done in the state. Well, yes, Governor, as a landlord I would prefer not to have to deal with the building inspector and all the red tape having to do with managing property in the state. But the people of the state seem to think that's necessary. Funny when they keep thinking it's necessary despite your personal wishes!

Even worse, the President of the Utah Senate—a particularly odious politico named Stuart Adams—chairs the MIDA committee that brokered the deal. When the Box Elder county committee gave final approval behind closed doors, a $135,000 donation appeared in Adams' PAC account the next day. That amount is about one-tenth of all donations to all Utah Senate campaigns for an entire election cycle.

Lately the governor has backed off and admitted that the "roll-out" of the data center was "poorly handled." Yessir, you rolled the dice, got caught, and now it has to be done the right way with public comment, public disclosures, and the proper impact statements. The backlash has been so unequivocal that all new data centers in the state are paused pending review of their approval process. Utahns are really, really mad.

But O'Leary went on the rampage from Day One. He said such dumb things as the crowds being "AI-generated." He claimed all the protesters were bussed in from outside the state and paid off by liberal donors. In other words, the standard knee-jerk attempts from the right to defuse the effect of their massively unpopular proposals. Yes, a lot of people came from Salt Lake and other traditionally liberal pockets of the state. But as a share of Box Elder County's population, almost everyone from the county showed up. These are Trump voters—79% for Trump in 2024. Read that number again. When Box Elder County was polled to determine whether they wanted the data center, it got a 2% approval. Read that number again.

Gov. Cox is backpedaling because this may be bad enough to kill his career in Utah politics. O'Leary is under no such threat. He doesn't live here. (He was bussed in.) He doesn't vote here or run for office here. So he's doubling down. He imagines that even if his bluff were true, the people involved would somehow be breaking the law, and so he's "collected information" and "passed it along to law enforcement." Protesting is not illegal. Traveling to another state to protest is not illegal. Donating money to fund protests is not illegal. All the things he's saying people are doing to kill his giant mega-project are not illegal, irrespective of whether they are actually doing them—they aren't. But he's so indifferent to the truth and public opinion that he wants to send out a threat that people are going to be arrested and prosecuted for opposing him. It's mounting desperation.

The really stupid thing is that out of the other side of his mouth he's trying to assure Utahns that his project won't affect the climate, won't use up all our precious water, won't drain the Great Salt Lake watershed, etc. He wants to convince people that even though the proper studies and covenants weren't made, we should still trust him to do the right thing. And O'Leary seems to want to reinforce that trust by threatening people and lying about them. If there was ever a personification of, "I can do whatever I want no matter what anyone says," it's these billionaires.

I'm sorry for making this political, but I find it difficult to muster much sympathy in this specific case. Box Elder county went 79% red in 2024. NOW they're upset it's their faces getting eaten, but they wanted fascism and they're getting it good and hard.

This is what they voted for.

Doo dah, doo dah

 

[Tero]

Copies? Yes. Google Gemini:
(Don't give it to N Korea)

Duplicate copies of a model are used constantly in machine learning. Here are the core facts about duplicating an LLM:

The Core Elements: A fully trained LLM consists of two main parts: the weights (the learned knowledge) and the tokenizer (the vocabulary that translates words into numbers). Copying both yields an identical twin of the model.

Storage and Transfer: You can zip these files up and transfer them across servers, share them on repositories like Hugging Face, or deploy them on different local computers.

Cost vs. Training: Copying the model is instantaneous and free. In contrast, training the model requires vast amounts of computing power, data, and time.Independent Futures: Once duplicated, each model acts entirely independently. You can fine-tune Copy A for a medical role, while leaving Copy B for a legal role; their changes will not interfere with one another.

 

[arthwollipot]

Source: Saturday Morning Breakfast Cereal

 

[Tero]

I understand the vector part. But the actual coding...what does one write code in to put together a LLM? I have not written code sionce 1980.

GloVe: Global Vectors for Word Representation

Jeffrey Pennington, Richard Socher, Christopher Manning. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2014.

aclanthology.org

Google:
According to the GloVe paper, the next step in utilizing the static word vectors to build a functioning program is to feed these embeddings into a machine learning model architecture [D14-1162]. This process involves initializing an embedding layer, defining a neural network, training on a specific task, and deploying the resulting model.

 

[Tero]

Another PDF

1.1 What is Natural Language Generation?Both text-to-text generation and data-to-text generation are instances ofNatural Lan-guage Generation(nlg). In the most widely-cited survey ofnlgmethods to date (Reiter& Dale, 1997, 2000),nlgis characterized as ‘the subfield of artificial intelligence and com-putational linguistics that is concerned with the construction of computer systems than canproduce understandable texts in English or other human languages from some underlyingnon-linguistic representation of information’ (Reiter & Dale, 1997, p.1). Clearly this defini-tion fits data-to-text generation better than text-to-text generation, and indeed Reiter andDale (2000) focus exclusively on the former, helpfully and clearly describing the rule-basedapproaches that dominated the field at the time.

https://www.jair.org/index.php/jair/article/view/11173/26378

 

[Tero]

I got mostly this:
Frontend: Python
Python is the primary language used for the "outer layer" of LLMs.
Backend: C++ and CUDA
While the code you see is often Python, the heavy mathematical lifting—the "hot path"—is executed in C++ or C.

 

[theprestige]

Interesting take

https://bsky.app/profile/djdynamic.ca%2Fpost%2F3mmug5iv6sk25

I was gonna say he has a fair point about adversarial queries. Most manufacturers have no responsibility to design their products for "customers" that set out to break them, nor any liability when those "customers" do break the product. Toyota's safety features aren't supposed to protect against a demolition derby.

But then I read the "adversarial query" being demo'd. No, an LLM should be good enough to handle the most basic examples of misinformation and user confusion.

 

[Beelzebuddy]

I understand the vector part. But the actual coding...what does one write code in to put together a LLM? I have not written code sionce 1980.

GloVe: Global Vectors for Word Representation

Jeffrey Pennington, Richard Socher, Christopher Manning. Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2014.

aclanthology.org

Google:
According to the GloVe paper, the next step in utilizing the static word vectors to build a functioning program is to feed these embeddings into a machine learning model architecture [D14-1162]. This process involves initializing an embedding layer, defining a neural network, training on a specific task, and deploying the resulting model.

This is a great example of why you can't just throw a random design question at an llm and get a usable response.

You want a python library called pytorch.

 

[Andy_Ross]

MSNow: You're willing at this point to let the Senate fall into the hands of Democrats, if that's what it takes to end data centers?

Republican Voter: My entire community is going to break rank. Everybody, all of us. We've had enough

video in link

 

[Dr.Sid]

This is a great example of why you can't just throw a random design question at an llm and get a usable response.

You want a python library called pytorch.

That's the old way. Today you simply ask LLM to build you a LLM.

 

[dirtywick]

Kevin O'Leary claims Chinese propaganda is to blame for anti-datacenter backlash, 'hundreds of millions of dollars' being spent to kill US dominance in AI — industry proponents and Trump administration reinforce claims of foreign interference

Just because you're paranoid doesn't mean they're not after you.

www.tomshardware.com

kevin o'leary and the government are blaming chinese propaganda for unpopularity of data centers. maybe some truth to it, but still remains that americans have serious concerns about data centers and ai

Even some parties that support the notion of Chinese interference are aware of the particular optics around this situation. According to The Post, Ryan Fedasiuk of the American Enterprise Institute says the inference issue is real, but also cautions "that China isn't the reason AI buildouts are unpopular in the United States." Likewise, the Bitcoin Policy Institute issued a report where it states there's a "foreign influence campaign against American AI," but also that "Americans do have serious concerns that need to be heard."

 

[Trebuchet]

What, it's not Joe Biden's fault?

 

[The Great Zaganza]

Weird how it's always propaganda when people don't like exactly the same thing they do.

 

[arthwollipot]

 

[gabeygoat]

It would be funny if it wasn't so predictable

You can no longer Google the word 'disregard' | TechCrunch

After Google Search's AI update, the word "disregard" now effectively breaks the search interface.

techcrunch.com

from the article: I have been a professional tech journalist for nearly 15 years, and before today, I cannot think of a single time when a Bing search result was more valuable than the Google equivalent. There really is a first time for everything!

 

[Darat]

View attachment 71806

Sounds like they made 2 major changes when moving to AI images, so it may not simply have been the move to AI that caused the click through rate to drop. I would have thought that (non-scam) retailers would be very leery about using AI generated images of their products since AI is likely to alter the product, especially if you are iterating designs and images to use. Plus the resolution will be too low to use anywhere but on line and something else I've noticed when looking at AI images, they are very "dirty" in how they use colour.

 

[arthwollipot]

Sounds like they made 2 major changes when moving to AI images, so it may not simply have been the move to AI that caused the click through rate to drop. I would have thought that (non-scam) retailers would be very leery about using AI generated images of their products since AI is likely to alter the product, especially if you are iterating designs and images to use. Plus the resolution will be too low to use anywhere but on line and something else I've noticed when looking at AI images, they are very "dirty" in how they use colour.

It may also be entirely made up.

 

[theprestige]

It would be funny if it wasn't so predictable

You can no longer Google the word 'disregard' | TechCrunch

After Google Search's AI update, the word "disregard" now effectively breaks the search interface.

techcrunch.com

Why is everyone Googling 'disregard'? It's breaking Google

As of May 22, several action words, including "disregard," "ignore" and "dismiss" are causing a glitch when using Google Search.

eu.usatoday.com

Google search now made one million percent better by having its AI interpret your search query (and two million percent better by results leading with its AI vomit ) was being stumped if you used the word disregard (and a few others).

Been playing a bit with it now and it still hasn't been fully resolved (because it's a fundamental to the way these AIs work) using the word disregard in a so called "search" box and anything up to the word is dumped. And these AIs are ready to replace millions of the lowest value human capital.

I can't replicate the results.

I can, however, get similar results with the search term "pray in a closet", which Gemini interprets as a command it cannot obey, rather than a phrase to search for in the vast library of human experience that is the Internet.

I would also like to point out, as a kind of public service announcement, that appending "-ai" to a Google search still works just fine. It brings Google back to where it was ten years ago, when searches were searches, and Wikipedia was often the top result.

There are several browser extensions that let you set default search parameters. Having "-ai" automatically appended to my searches has greatly improved my quality of life.

 

[arthwollipot]

Who's going to remember to do that?