Certainly not THIS US government. Besides that sad fact, the whole thing has no teeth and is totally voluntary, which means the Silicon Valley set will ignore the entire order.
The question is who else? If it's done by the federal government, it means there will be a 1000 eyes trained on the process to critique it. That's how we make sure that the federal government does anything well from equipping an army to running a court system to sending out Social Security checks.
Gary, I had a thought. It was a big mistake to characterize what LLMs do as "predicting" the next word. Prediction is an active mental process that is difficult if not impossible for human beings to perform. It is more accurate to say that LLMs merely "reflect" the next word as found by mindless bean counting in their capacious databases. Reflection Is mindless/low grade, whereas prediction is active and difficult/impossible/superhuman.
<Prediction is an active mental process that is difficult if not impossible for human beings to perform>
I don't believe so. If a conversation is started, "How do you [X]?", we can easily predict the next word is likely to be "do".Because there is so much redundancy in teh English language, word/phrase prediction is little more than selecting the most probable next word/phrase. I believe we are doing that with Kahneman's "System 1, or thinking fast." You can test this by running a new video, pausing a conversation, and "guessing" the next word the speaker will say. Sometimes th dialogue is so trite or hackneyed that you can guess whole sentences that are to be said noxt/
People, including my partner, do this to me all the time when I speak - guess the next word that is. Maybe I speak too slowly, but it's really annoying when they do it although I do enjoy it when they realise that they had no idea whatsoever about what I was saying to them and should've waited for me to finish speaking. Reminds me of those Two Ronnies sketches when they're in the pub ;-)
And the minute you get past a small problem which has one issue (say you want a python script to do one thing) to a problem that has multiple issues (say a legal case) and needs context kept these AIs are worse than useless.
And they always reflect from the POV of someone who has power in the system for their advice to you while at the same time saying things like "I'm going to push back on that" when you describe that power too accurately. Dumb censorship machines.
My point is that LLMs do not undergo a process similar to human prediction. They simply digitize a database and come up with the next most common lexeme.
When I read your example, “how do you…” what popped into my mind was not small talk, but a serious question about how to complete a process of some sort. for example, how do you ... start a lawn mower?
My guess why my mind went in this direction is that we are in reading a sciency kind of substack, not a small talk related substack. Gary keeps emphasizing that the human mind places things in real world contexts and simply doesn't calculate percentages of word usage. Also, “how do you do? ” is a rarely used expression these days. In 2026 America, most people would say something like “how's it going,” or “how are you?” That's probably another reason why my mind did not jump to, “how do you... do?
The prediction in this case is a statistical probability rather than a mental process. There is a certain statistical probability that X word will follow Y word, and that is what the AI is basing itself on. The training process delivers these statistical probabilities. So, to clarify, it is the training process that is human supervised that yields the "predictive" elements.
But LLM's predictions are not coming from vacuum, they are still based on pre-digested information - millions, of course, but still something humans can perform as well. I mean, have you ever "predicted" a word from your parents or your close friends, just because you learned them well over years? Doesn't phones T9 predict as well, but on a much smaller scale? LLMs are just T9 on steroids then.
Nonsense. Humans predict all the time. All the time! Think about how often something surprises you. That's when you predicted something and it failed to occur. Perhaps you are thinking of a particular kind of prediction.
I agree that humans predict all the time. But every time we are surprised it is an indication that our prediction has failed. It is accurate prediction that is difficult—any old prediction Is as easy as Falling off a chair. If accurate prediction was within normal human capacity. we'd all make our living at the racetrack betting on horses.
It's not that kind of prediction. More like predicting that if you move this sequence of muscles, your body will end up where you intended to go. If you eat this vanilla ice cream, it will taste creamy, not sour. If someone says to you, "I don't like vanilla but I do like ...", they are going to tell you some other flavor of ice cream. Instead, if they say "cats", you will be surprised and disoriented because you aren't understanding the conversation at some level. Your prediction of what will come next has failed.
You´re the f***ing GOAT Gary! Everytime I here something about "the next frontier model", I immediately check what you and George Zöeller write about. All of these words are just to say thank you.
It's a trump EO, so read the fine print...I mean, amongst the clown-show trump regime, who exactly will be tasked for "oversight"? Bill Pulte, Ed Martin? Lil Mario, who already has a half-dozen or so portfolios.
When the models have no moat of their own, regulation is the only one anyone can build. A prerelease review, even a voluntary 30-day one, turns shipping a model into a fixed cost of legal overhead and a government queue. The billion-dollar labs absorb that easily. The open-source shop and the small challenger don't. Whatever the safety case, the market effect is to hand the incumbents the moat the technology never gave them. Safety and entrenchment can be the same policy seen from two desks.
Especially since the ultimate goal is to "beat China" (which is impossible) by limiting the penetration of open-weight models into the US.
I'm surprised Trump hasn't banned Chinese models in US business, especially since many start-ups are using them, and Anthropic and OpenAI both want them banned.
Isn't it the case that an executive order means little if Congress doesn't turn it into a law, or an agency takes on that regulation after public comment? Which agency would be involved?
My problem with this is that, unlike other products and services, AI is a multi-domain tool. It could be used to create new drugs (FDA), new dietary advice (USDA), new clinical devices and procedures (HHS), etc, etc. Is each agency going to have to look at validating whether it is safe in their domain? How will they do that, as there may not be procedures to handle the many cases it can work on in any specific domain?
It may even need a new agency to oversee the process. Industry will no doubt say that validation may take a long time, perhaps 6 - 12 months, and that this will handicap its commercial value, as well as being subject to theft by foreign powers. Is Trump perhaps hoping to roll back that EO after he receives a hefty bribe or three from the big players?
To prevent the AI from being trained to recognize the tests (like VW's exhaust system) and circumventing them, the tests will have to remain hidden. So now there is an incentive for an individual to make $$$$ by providing that information to interested parties (spying).
This is all going to take a lot more thought than just signing an EO. Look how hard it is dealing with trying to regulate social media for children. AI is so much bigger.
As soon as I heard that the Trump administration wants to review AI models, I immediately knew it would be more about suppressing free speech than ensuring safety. It will be much like Musk's takeover of Twitter. Trump has no interest in public safety but if anyone says anything he doesn't like, he jumps all over it. Expect the worst.
If you are a gifted code author with tensor chops sufficiently robust to be of use in analyzing the back end of an AI implementation, would you take a job in government?
Alternatively. playing 20 million questions with the front end seems like a waste of time. Every model will fail some queries. Nobody can predict which will prove important later.
Who oversees the overseers? Who writes the rules? Who implements them? It's already too late, Pandora's box has been opened, humans + computers are already slaughtering, surveilling, interdicting, inventing false realities. AI is just the nail in our collective coffin.
Oh William, surely we will come up with even stupider ways to destroy ourselves. AI is just what's popular to end us all right now. Last year, wasn't it microplastics? And now Hanta and Ebola. We are just getting started.
I'm not sure what you're saying/advocating. Hanta and Ebola are some kind of solution? Or are you saying that there is no solution, only extermination?
It's already too late you wrote. Well, that might be true about a bunch of things. So far, we have escaped the worst of outcomes most of the time. So while this and that might seem hopeless at the moment, it may resolve itself without undue harm as other are already working at the solution. And tomorrow we can all look at the news and say as Dorothy taught us, "What fresh hell is this?"
They just want a piece of whatever power and 'glow' accrues from being adjacent to the topic. Nothing to do with the public good. The Trump admin has been interventionist as hell (something something 'free market conservatives' lol) in tech, commodities, etc--this has to be considered as more of that.
The executive order is watered down from the previous bill which was not much itself. As soon as any AI financial stakeholder have anything to say, the administration follows. It's like AI industry is policing itself and we all know how this will work out. GIVE ME A BREAK!! DO BETTER.
I am actually a fan of Gary's. I was just being snarky and he was being self deprecating. He has been more right than wrong about the major issues of AI
Gary, we just commented on that government move in the last thread. Basically, COMMISSION GOOD, Today's governmental bad actors: BAD. Here is the exchange:
ME: This just in from the New York Times: "Trump Signs Executive Order Seeking Oversight of A.I. Models/The order, which signaled a shift from the hands-off approach the White House had previously taken toward A.I., followed debates over how to gain control of A.I. models without disrupting innovation."
President Trump at the White House last week. Credit...Doug Mills/The New York Times/Sheera Frenkel Tripp Mickle/By Sheera Frenkel and Tripp Mickle Reporting from San Francisco/June 2, 2026k/Updated 12:50 p.m. ET
"President Trump signed an executive order on Tuesday that asked technology companies to give the government oversight of new artificial intelligence models before releasing them to the public, a shift for an administration that had promoted a hands-off approach to the powerful technology . . . . 'Advanced A.I. capabilities make our nation stronger, but also introduce new national security considerations that require coordinated action across executive departments and agencies,' the order said. This is a developing news story. Check back for updates."
REPLY BY: richardstevenhack
I read the White House readout. / It's a joke. / They're forming a commission to decided what the threshold is for a powerful model to be considered "covered." They're going to use "scanning" - as if that works. Then if they decided the model is powerful enough, there is a thirty-day waiting period for the company providing the model to wait until the government analyzes the model. No word on how that's to be done, but if it's not adversarial threat modeling, it's a waste of time.
Bottom line: This is a scam to enable the Deep State to get its hands on models to use for espionage against other countries and surveillance of the US population. Plus corporate bribes to Trump and his family.
ME: Catherine Blanche King, REPLY TO: richardstevenhack:
It's no surprise to me or probably to others here (I hope) that our present government is already one of the, if not THE worst bad actor in the world right now.
However, with our still partly democratic power situation up in the air, and as we undergo changes going forward, I'm also sure that Bernie Sanders' similar (but differently founded) proposals are NOT THAT. If so, then as you also probably understand, it will depend on who has the reigns if and when such a "commission" occurs that has any hope of actually doing the job.
Though some kind of well-thought-out guardrail situation is sorely needed, e.g., as a transparent commission with interactive oversight by good and well-informed actors, for the time being, gaslighting rather than good actors is all we can expect.
In other words, though a commission of sorts has a great sound to it, that's what it's intended to sound like, by our present power group of bad actors. They always need a shining horse to ride in on so it's easier to screw everyone concerned once they get control.
Except ... do you really trust the US government? I doubt that this is being done in the public interest given their previous track record.
Certainly not THIS US government. Besides that sad fact, the whole thing has no teeth and is totally voluntary, which means the Silicon Valley set will ignore the entire order.
Precisely. The angle here (and it's always an angle with this circus) isn't altruism, it's dominion.
The question is who else? If it's done by the federal government, it means there will be a 1000 eyes trained on the process to critique it. That's how we make sure that the federal government does anything well from equipping an army to running a court system to sending out Social Security checks.
First, the army is not well equipped. Too long a subject to go into, but trust me. This is why we ran out of missiles after 40 days of bombing Iran.
Second, this commission's benchmarks (to be developed) are classified - not subject to public review, as far as I can tell from the readout.
Third, the US government does not respond well to "critigue." Ever.
Check out how the eyes did monitoring the distribution of aid in Minneapolis.
The most important job of checkers is to determine which checking targets are untouchable because they connect to a key legislator.
After this law passes, when anyone asks an LLM what evidence exists of Trump's corruption, you can be sure the answer will be "none".
And what will it say right now about the corruption of the likes of Amodei and Altman?
"Who?"
"Who?"
"Who?"
Gary, I had a thought. It was a big mistake to characterize what LLMs do as "predicting" the next word. Prediction is an active mental process that is difficult if not impossible for human beings to perform. It is more accurate to say that LLMs merely "reflect" the next word as found by mindless bean counting in their capacious databases. Reflection Is mindless/low grade, whereas prediction is active and difficult/impossible/superhuman.
<Prediction is an active mental process that is difficult if not impossible for human beings to perform>
I don't believe so. If a conversation is started, "How do you [X]?", we can easily predict the next word is likely to be "do".Because there is so much redundancy in teh English language, word/phrase prediction is little more than selecting the most probable next word/phrase. I believe we are doing that with Kahneman's "System 1, or thinking fast." You can test this by running a new video, pausing a conversation, and "guessing" the next word the speaker will say. Sometimes th dialogue is so trite or hackneyed that you can guess whole sentences that are to be said noxt/
People, including my partner, do this to me all the time when I speak - guess the next word that is. Maybe I speak too slowly, but it's really annoying when they do it although I do enjoy it when they realise that they had no idea whatsoever about what I was saying to them and should've waited for me to finish speaking. Reminds me of those Two Ronnies sketches when they're in the pub ;-)
Thanks for helping me make my point.
In social small talk phrases can be hackneyed and predictable. The minute you get past small talk, prediction becomes virtually impossible.
And the minute you get past a small problem which has one issue (say you want a python script to do one thing) to a problem that has multiple issues (say a legal case) and needs context kept these AIs are worse than useless.
And they always reflect from the POV of someone who has power in the system for their advice to you while at the same time saying things like "I'm going to push back on that" when you describe that power too accurately. Dumb censorship machines.
My point is that LLMs do not undergo a process similar to human prediction. They simply digitize a database and come up with the next most common lexeme.
When I read your example, “how do you…” what popped into my mind was not small talk, but a serious question about how to complete a process of some sort. for example, how do you ... start a lawn mower?
My guess why my mind went in this direction is that we are in reading a sciency kind of substack, not a small talk related substack. Gary keeps emphasizing that the human mind places things in real world contexts and simply doesn't calculate percentages of word usage. Also, “how do you do? ” is a rarely used expression these days. In 2026 America, most people would say something like “how's it going,” or “how are you?” That's probably another reason why my mind did not jump to, “how do you... do?
To be precise, they're not doing either predicting or reflecting.
They are just executing an algorithm that follows a chain of probabilities that connect words together in multiple dimensions.
Perhaps over-simplified, but that's the basic technology.
It's amazingly useful, but also amazingly unreliable and insecure, which puts serious limits on its usability, especially in high-risk deployment.
The prediction in this case is a statistical probability rather than a mental process. There is a certain statistical probability that X word will follow Y word, and that is what the AI is basing itself on. The training process delivers these statistical probabilities. So, to clarify, it is the training process that is human supervised that yields the "predictive" elements.
But LLM's predictions are not coming from vacuum, they are still based on pre-digested information - millions, of course, but still something humans can perform as well. I mean, have you ever "predicted" a word from your parents or your close friends, just because you learned them well over years? Doesn't phones T9 predict as well, but on a much smaller scale? LLMs are just T9 on steroids then.
Nonsense. Humans predict all the time. All the time! Think about how often something surprises you. That's when you predicted something and it failed to occur. Perhaps you are thinking of a particular kind of prediction.
I agree that humans predict all the time. But every time we are surprised it is an indication that our prediction has failed. It is accurate prediction that is difficult—any old prediction Is as easy as Falling off a chair. If accurate prediction was within normal human capacity. we'd all make our living at the racetrack betting on horses.
It's not that kind of prediction. More like predicting that if you move this sequence of muscles, your body will end up where you intended to go. If you eat this vanilla ice cream, it will taste creamy, not sour. If someone says to you, "I don't like vanilla but I do like ...", they are going to tell you some other flavor of ice cream. Instead, if they say "cats", you will be surprised and disoriented because you aren't understanding the conversation at some level. Your prediction of what will come next has failed.
You´re the f***ing GOAT Gary! Everytime I here something about "the next frontier model", I immediately check what you and George Zöeller write about. All of these words are just to say thank you.
It's a trump EO, so read the fine print...I mean, amongst the clown-show trump regime, who exactly will be tasked for "oversight"? Bill Pulte, Ed Martin? Lil Mario, who already has a half-dozen or so portfolios.
One wonders...
When the models have no moat of their own, regulation is the only one anyone can build. A prerelease review, even a voluntary 30-day one, turns shipping a model into a fixed cost of legal overhead and a government queue. The billion-dollar labs absorb that easily. The open-source shop and the small challenger don't. Whatever the safety case, the market effect is to hand the incumbents the moat the technology never gave them. Safety and entrenchment can be the same policy seen from two desks.
What then is the solution? Do you think perhaps cranking up the anti-monopolization laws to go after this entire industry?
Excellent points.
Especially since the ultimate goal is to "beat China" (which is impossible) by limiting the penetration of open-weight models into the US.
I'm surprised Trump hasn't banned Chinese models in US business, especially since many start-ups are using them, and Anthropic and OpenAI both want them banned.
What do Anthropic and OpenAI have against Chinese models like Liu Wen?
LOL
Isn't it the case that an executive order means little if Congress doesn't turn it into a law, or an agency takes on that regulation after public comment? Which agency would be involved?
My problem with this is that, unlike other products and services, AI is a multi-domain tool. It could be used to create new drugs (FDA), new dietary advice (USDA), new clinical devices and procedures (HHS), etc, etc. Is each agency going to have to look at validating whether it is safe in their domain? How will they do that, as there may not be procedures to handle the many cases it can work on in any specific domain?
It may even need a new agency to oversee the process. Industry will no doubt say that validation may take a long time, perhaps 6 - 12 months, and that this will handicap its commercial value, as well as being subject to theft by foreign powers. Is Trump perhaps hoping to roll back that EO after he receives a hefty bribe or three from the big players?
To prevent the AI from being trained to recognize the tests (like VW's exhaust system) and circumventing them, the tests will have to remain hidden. So now there is an incentive for an individual to make $$$$ by providing that information to interested parties (spying).
This is all going to take a lot more thought than just signing an EO. Look how hard it is dealing with trying to regulate social media for children. AI is so much bigger.
Like a broken clock, you should definitely take credit for being correct twice a day!
Congrats.
Lets see how it actually gets implemented and if the USG has the people to determine if their are issues before release.
As soon as I heard that the Trump administration wants to review AI models, I immediately knew it would be more about suppressing free speech than ensuring safety. It will be much like Musk's takeover of Twitter. Trump has no interest in public safety but if anyone says anything he doesn't like, he jumps all over it. Expect the worst.
If you are a gifted code author with tensor chops sufficiently robust to be of use in analyzing the back end of an AI implementation, would you take a job in government?
Alternatively. playing 20 million questions with the front end seems like a waste of time. Every model will fail some queries. Nobody can predict which will prove important later.
Yeah I expect this is just the Trump administration/disaster wanting control of everything. Nothing good will come of it.
It's not about safety but speech suppression. Ask the LLM, "Who was the greatest US president?" You can guess the answer.
Who oversees the overseers? Who writes the rules? Who implements them? It's already too late, Pandora's box has been opened, humans + computers are already slaughtering, surveilling, interdicting, inventing false realities. AI is just the nail in our collective coffin.
And yet, that sounds less like an answer, and more like a capitulation. Some of us are not ready to throw in the towel just yet.
So what's your plan then?
Join an organization that is fighting back. Get involved. They want us to be cynical and give up.
Believe me, I've looked. I write, I've been active my entire life. Dark days.
Said Frodo to Sam as he headed toward Mordor , with GoLLM off in the bushes.
Oh William, surely we will come up with even stupider ways to destroy ourselves. AI is just what's popular to end us all right now. Last year, wasn't it microplastics? And now Hanta and Ebola. We are just getting started.
I'm not sure what you're saying/advocating. Hanta and Ebola are some kind of solution? Or are you saying that there is no solution, only extermination?
It's already too late you wrote. Well, that might be true about a bunch of things. So far, we have escaped the worst of outcomes most of the time. So while this and that might seem hopeless at the moment, it may resolve itself without undue harm as other are already working at the solution. And tomorrow we can all look at the news and say as Dorothy taught us, "What fresh hell is this?"
They just want a piece of whatever power and 'glow' accrues from being adjacent to the topic. Nothing to do with the public good. The Trump admin has been interventionist as hell (something something 'free market conservatives' lol) in tech, commodities, etc--this has to be considered as more of that.
With this administration I see this more as a "be careful what you wish for" moment.
The executive order is watered down from the previous bill which was not much itself. As soon as any AI financial stakeholder have anything to say, the administration follows. It's like AI industry is policing itself and we all know how this will work out. GIVE ME A BREAK!! DO BETTER.
I am actually a fan of Gary's. I was just being snarky and he was being self deprecating. He has been more right than wrong about the major issues of AI
Gary, we just commented on that government move in the last thread. Basically, COMMISSION GOOD, Today's governmental bad actors: BAD. Here is the exchange:
ME: This just in from the New York Times: "Trump Signs Executive Order Seeking Oversight of A.I. Models/The order, which signaled a shift from the hands-off approach the White House had previously taken toward A.I., followed debates over how to gain control of A.I. models without disrupting innovation."
President Trump at the White House last week. Credit...Doug Mills/The New York Times/Sheera Frenkel Tripp Mickle/By Sheera Frenkel and Tripp Mickle Reporting from San Francisco/June 2, 2026k/Updated 12:50 p.m. ET
"President Trump signed an executive order on Tuesday that asked technology companies to give the government oversight of new artificial intelligence models before releasing them to the public, a shift for an administration that had promoted a hands-off approach to the powerful technology . . . . 'Advanced A.I. capabilities make our nation stronger, but also introduce new national security considerations that require coordinated action across executive departments and agencies,' the order said. This is a developing news story. Check back for updates."
REPLY BY: richardstevenhack
I read the White House readout. / It's a joke. / They're forming a commission to decided what the threshold is for a powerful model to be considered "covered." They're going to use "scanning" - as if that works. Then if they decided the model is powerful enough, there is a thirty-day waiting period for the company providing the model to wait until the government analyzes the model. No word on how that's to be done, but if it's not adversarial threat modeling, it's a waste of time.
Bottom line: This is a scam to enable the Deep State to get its hands on models to use for espionage against other countries and surveillance of the US population. Plus corporate bribes to Trump and his family.
ME: Catherine Blanche King, REPLY TO: richardstevenhack:
It's no surprise to me or probably to others here (I hope) that our present government is already one of the, if not THE worst bad actor in the world right now.
However, with our still partly democratic power situation up in the air, and as we undergo changes going forward, I'm also sure that Bernie Sanders' similar (but differently founded) proposals are NOT THAT. If so, then as you also probably understand, it will depend on who has the reigns if and when such a "commission" occurs that has any hope of actually doing the job.
Though some kind of well-thought-out guardrail situation is sorely needed, e.g., as a transparent commission with interactive oversight by good and well-informed actors, for the time being, gaslighting rather than good actors is all we can expect.
In other words, though a commission of sorts has a great sound to it, that's what it's intended to sound like, by our present power group of bad actors. They always need a shining horse to ride in on so it's easier to screw everyone concerned once they get control.
Reply