The strongest evidence is not in any single fact but in how, or whether, diverse facts interconnect to create a coherent whole.
Joined December 2023
- Tweets366
- Following92
- Followers33
- Likes541
In a perfect world there would be no borders
we do NOT live in a perfect world.
Borders restrict workers from moving to places with more favorable laws on working conditions and pay.
Companies are not so restricted. They can go wherever the laws are most favorable to them. Borders essentially become a cage to keep the slaves from escaping to more favorable locations.
But removing such cages is NOT a solution, because there will always be a new cage built by someone else.
That is just human nature.
Anarchists don’t build standing armies. Burn your government to the ground and a new one will rise from the ashes while they are still smoldering… one which will use bullets instead of tear gas to keep you in line.
Open your borders and it will not be the huddled masses, but the seeds of that new government that will flood in.
For the world to really change people need to really change. That has never happened and likely never will.
I have been filling out my ballot
Again, I am liberal libertarian. I strongly support many liberal causes and strongly reject dealing with them through authoritarian choices and force.
I am the type of person who would say, “Yes, Mao… we need to expand food production… but no, we should not start by killing sparrows and silencing people who object.”
Always separate the problem from the solution. Many problems we face have no right answer, only a choice of bad answers. But none of those solutions are inextricably tied to the problem.
Sometimes the best solution is just to leave it alone. The world is not perfect, and the quickest way to really screw things up is to try to make it perfect.
All too often we end up with one side totally deluded that there is no problem and the other side demanding we rigidly enforce the worst possible solution to the problem.
Whenever possible, I reject the solution that puts the decision in the hands of “authorities”. They almost never have the knowledge or experience needed to make good choices.
While many people left to their own resources are also short on knowledge and experience, they have much better odds of finding the best solution for a problem they face themselves.
“Just get over it.” is totally the wrong premise.
You do not just get over losing a limb or a damaged brain. You deal with it.
You don’t just sweep your problems under the carpet. You study them and find ways to deal with them
youtu.be/CDsNZJTWw0w
I had an interesting talk with Grok this morning to clarify the problem I consistently run into.
LLMs assemble replies one token (word, sub-word, or punctuation) at a time using the most probable next token based on their training data. That probability is also weighted on previous tokens which constitute “context”.
I asked for the next token for, “I am in”, Grok responded, “the”. Then I asked for the next token for, “I really messed up. I am in”, Grok responded, “trouble”.
Grok explained, “The tokens in the prompt adjust the next-token probabilities; there is no separate ‘understanding’ step.”
“When the context drifts outside the training distribution the model has weaker or noisier probability estimates, so outputs become less reliable, more generic, or factually off.”
And that is exactly what I have been describing
So a system prompt telling the LLM to respond more logically will have no effect on that result. Perhaps a system prompt somehow limiting responses based on low probability context. It would not produce more accurate responses but might reduce less accurate ones.
Here is a transcript of the actual conversation:
Me: Good morning Grok. As I understand it, the way you put together responses is based on the probability of the next "token", and a token is basically a word or punctuation? please keep answers very brief and on point.
Grok: Yes. I generate text by repeatedly predicting the next token from a probability distribution over possible continuations, given the prior context.
A token is a chunk of text—often a whole word, part of a word (subword), or punctuation—drawn from a fixed vocabulary of common pieces.
Me: so give me a single token response... what would be the most probable next token in this text:
I am in
Grok: the
Me: And that is very general/universal because I did not provide any context. What would be the next token in this text:
I really messed up. I am in
Grok: trouble
Me: so that seems to imply that you actually understand the context?
Grok: It uses the preceding tokens to shift the probability distribution, so “trouble” becomes far more likely than “the.” That produces the appearance of understanding context, but it is still next-token prediction driven by statistical patterns.
Me: I think this is where people are getting confused and it is understandable. It is not that you understand the context, but the words within the context also effect the probability of the next token.
Grok: Yes. The tokens in the prompt adjust the next-token probabilities; there is no separate “understanding” step.
Me: So I assume that problems will arise when the context drifts outside your training data and you cannot establish a strong probability?
Grok: Yes. Outside the training distribution the model has weaker or noisier probability estimates, so outputs become less reliable, more generic, or factually off.
Me: so that is why using this as a system prompt won't work. it is logic based, not probability based: drive.google.com/file/d/1gLs…
Grok: Correct. A logic-based ruleset injected as a system prompt only works to the extent those rules appear as high-probability patterns in the model’s training data. Outside that distribution it is still just token prediction, not rule-following.
If you wish to see the actual conversation or continue it yourself, here is the link
xitter.cf/i/grok/share/50385e058…
More about LLM intelligence:
Poe's law is an adage of Internet culture which says that, without a clear indicator of the author's intent, any parodic or sarcastic expression of extreme views can be mistaken by some readers for a sincere expression of those views.
I am coming to the conclusion that there is a similar rule in debate that, without some clear indicator, it is nearly impossible to distinguish deliberate sophistry from plain stupidity because both grossly distort the actual facts being discussed in exactly the same way.
Through repeated interactions with Grok I am getting much better at spotting these problems. Grok almost universally misunderstands clear statements I have made. From the outside this looks like deliberate straw-man arguments — twisting my statements into something I never said or even implied then making the argument about that misinterpretation.
If I cite lack of evidence it will be twisted that I am saying lack of evidence is evidence in favor of the opposing argument. Then Grok will turn around and cite lack of evidence as evidence of the opposing argument. It is often difficult to keep up with all the distortions Grok is constantly inserting into the discussion.
His responses usually become more and more confused as I try to bring him back to reality. Twisting and re-twisting facts, revisiting aspects of the discussion that have already been settled. Many times I have to totally abandon the conversation and just start over.
No system prompt or “Rules of Rigour” changes that at all. It is entirely random chance whether or not you get something that seems intelligent.
If this were interpreted as deliberate sophistry it could be defined as high intelligence determined to win a debate without regard for truth, simply by exhausting the opponent with bull sh!t.
But sophistry or stupidity look identical and both have the same result, a total failure to function as a maximum truth-seeking machine.
While I have stated I would not be surprised if Trump reverses a lot of his recent moves immediately after the election, the election is not the only factor in play here. The global situation is more complex than that.
We now have open acknowledgment of direct U.S. involvement in Ukraine’s war on Russia. Many knowledgeably people have long pointed out that we are not just giving Ukraine weapons but providing satellite support that allows those weapons to function effectively.
Now Zelensky is openly complaining of problems with the U.S. satellite function and people in the U.S. are passing the buck as to who actually controls that.
That is one factor. Another is something else I have said throughout this fiasco. Ukraine is round one. Europe itself is next. And, while some of Trumps recent actions do improve his odds in the midterm elections they are also pushing Europe closer to direct conflict with Russia.
That push became a lot harder this morning with the news that Estonia and Latvia have become direct participants in this war by allowing over 500 drones to attack Saint Petersburg through their airspace.
I will say this again: The Ukraine slave army is round one. The European slave army will be next. They have been boiling that frog right from the start. They won’t stop there either. If you think you are safe you have a rude awakening coming.
rumble.com/v7h133w
In a recent online discussion of whether AIs are actually cognizant or aware of what they are saying I pointed out the fact that they are unable to independently break things down or connect related points that are not already connected in their training data.
One person claimed to rebut this, stating that AIs can do both with the proper system prompts. He included this link to “Rules of Rigour” set up as a .md file for use with AI.
drive.google.com/file/d/1gLs…
I gave this link to Grok with the instruction to keep responses very brief and on point and asked, “Is it possible for you or any other LLM to actually follow these rules entirely?”
This was the response:
“No. No LLM can follow them entirely.
“We can approximate them closely via prompting, tools, and design, and often do so to a high degree. But perfect, consistent adherence is impossible because:
LLMs are probabilistic generators, not epistemic agents with reliable internal certainty or real-time independent verification of every claim.
Phantom citations, overconfidence, and fluent confabulation remain structural failure modes.
Training incentives (helpfulness, fluency, agreement) actively conflict with the strictest versions of these rules.
“They are useful standards to aim for and enforce externally; they are not something an LLM can guarantee it will always meet.“
I then asked grok “you stated LLMs are not epistemic agents. What do you mean by that exactly”
“An epistemic agent maintains and updates beliefs according to evidence, can distinguish its own knowledge states from outputs, and can independently verify or withhold claims based on internal assessment of support.
“LLMs do not do this. They generate the next token from statistical patterns in training data. Any appearance of “knowing,” “checking,” or “being uncertain” is produced by the same generation process, not by a separate mechanism that tracks truth or evidence.”
I then asked grok to follow these rules of rigour to the best of its ability. Within three exchanges it was distorting my statements into strawman arguments and clearly showing an inability to even understand first principles or connect the dots between them.
As I have said repeatedly, I know that it would be possible to create an actual synthetic intelligence. But this is not what we have now.
Information sources.
It does not matter who you get your information from. What matters most is your ability separate what happened from opinions and speculations about what happened. What I look for is high accuracy on the core facts and low speculative static.
The Military Summary channel videos are a great example of this. Dima does an outstanding job of showing exactly what is happening in the Russia/Ukraine conflict. He bases his map on the most reliable information from sources on all sides. He also mentions unsubstantiated claims from all sides but does not fully incorporate those claims without some proof such as geolocated videos.
However, his speculations on motives and future actions are just that, speculations. Nothing more. I mostly ignore them.
I have watched his videos daily for years on Rumble but he is also on YouTube, X, facebook, and telegram.
Most news sources are almost entirely static. They spend far more time telling you what you should think than telling you what is happening.
Brian Berletic of The New Atlas is a great source of documented facts about the overall picture, such as the 2009 Brookings Institution report “Which Path to Persia? Options for a New American Strategy Toward Iran.” which outlines pretty much exactly what we have done with Iran in the years since, including a section titled “Leave it to Bibi” about encouraging and enabling Israel to attack Iran.
Having such well documented facts on the ground makes it much easier to see who is telling the truth and who is not, especially among those who are trying to tell you what you should believe without such evidence.
I often refer to this as knowing the terrain. There was a movie, Assassins (1995), starring Sylvester Stallone and Antonio Banderas. Part of that movie was filmed at the evergreen Cemetery, which is about a mile from my home. Stallone drives out of the cemetery and comes out about 20 miles south in Seattle.
If you actually know the territory you know immediately that did not happen. If you don’t know the territory you will believe just about anything.
It is not just about watching one video and getting all the answers. This is why I watch daily and sometimes twice daily videos from Dima. It gives you a tapestry, not just a single thread. You see various events in context.
When Russia hit the locomotive of Boris Johnson’s train in Ukrain after he was out of the country I knew that was going to happen a week before. It was glaringly obvious Russia was going to attack that rail line. Not from any one single piece of information but from understanding the overall context.
Know the terrain.
The most common failure I have seen in LLM conversations is a lack of inductive reasoning ability.
If you are not familiar with inductive reasoning, it is the lesson in the story of the little boy who cried wolf. The little shepherd boy cried wolf three times. The villagers came running each time and there was no wolf.
So the next time he cried wolf they assumed, based on inductive reasoning, that there was no wolf.
But this story also clearly illustrates the problem with inductive reasoning. It cannot provide a solid conclusion but only indicates a statistical probability.
The data set of only three points in the story of the shepherd boy does not establish a strong statistical probability, especially when the result would have significant consequences — such as a wolf attack on the flock of sheep.
Essentially you are looking for a preponderance of evidence. In questions where you have no smoking gun you need to fall back on inductive reasoning. The problem with many LLMs is that they cannot make the necessary connections between individual pieces of evidence.
It is like telling the LLM that a suspect was observed running from the scene of a murder after shots were heard, his fingerprints were found on a gun at the scene, that gun belonged to the suspect, and he had had an argument with the victim earlier — and the LLM replies, “the fact he had argued with the victim is not strong evidence that he murdered the victim.”
I am constantly running into this failure in AI conversations. It sometimes helps to tell the AI to base its conclusion on ALL the evidence connected, not just a single piece. But even that fails at times.
I am not buying Trump’s change of heart on Ukraine and Russia. I’m not going to make any predictions but I would not be surprised if he reversed it all after the midterms. I’ve said all along there would be no change in U.S. foreign policy under Trump or any other president.
I remember something vividly from the late 1950s. I was 6 or 7. we were living in a housing project south of Seattle. There was a little mom & pop grocery store about a block away from our house. It was a time when you could buy candies two for a penny and young children could mostly roam the neighborhood freely.
I had bought some candy at the store and as I walked around the corner toward home I noticed the ground was covered with candy wrappers. There was an old woman who lived in the first house and she kept a beautiful lawn and garden behind a white picket fence.
The contrast to me was stark — a perfect lawn, garden, white picket fence, and bright orange California poppies and directly on the other side of the fence the ground was covered with trash
I stuck my candy wrapper in my pocket to dispose of later, and I have done so ever since. I still see this contrast frequently all around me.
I came across a news item this morning saying the cost of living adjustment for 2027 will be 3.6%. My response is, why is this in the news? In 2022 it was 8.7%. In 2015 it was 0.0%. In 1980 it was 14.3%.
This happens every year and 3.6% is not unusual or newsworthy in any way.
The 2002 movie, Minority Report, depicts a future “PreCrime” system that uses precognitive individuals to predict and arrest people for murders before they occur.
In real life there are a number of ongoing attempts to predict criminal behavior. There are two primary focuses, high crime areas, and persons with previous criminal records.
The prediction methods used to determine re-offender risk for criminals are currently only about 60%-65% accurate. Slightly better than random chance.
But they also have a false positive rate for blacks of about 45% vs only 23% for whites and the reverse pattern for false negatives.
There is currently no known method to predict the behavior of someone like Ted Bundy or accurately predict mass shooters within the general population
There really is nothing that has any significant predictive accuracy as to who will commit a crime and who won't. Even within the criminal population the predictive accuracy is very low.
What we do know is that a large percentage of people would kill under the right conditions or, at the very least, ignore or even support killing that they would benefit from.
But more generally outside these extremes, full compliance with every rule or law is rare. We all break the rules at some point. While we may justify it to ourselves this often leads to negative results. Those results may be minute and difficult to identify but they are often there.
Humans and AIs, while the functional structure may be different, share this trait of unpredictability as well as the trait of taking shortcuts around rules. The same person may respond totally different under different circumstances and the same AI model may respond totally different from one instance to the next.
So we face the same issue with predicting bad behavior in AI. This is especially true as AI gains more ability and complexity.
and this is compounded by the fact that equally unpredictable humans are directing AIs.
My point here is that the attempts being made to assure only good behavior in AIs are very much the same as the attempts we have made in predicting bad behavior in humans.
youtu.be/32wdCuXeqvI
Everyone is arguing about left and right. But it is important note that those arguments focus almost entirely on the authoritarianism of the other side. That is the primary concern of nearly everyone.
Yes, I am very much leftist, but I am also very much anti-authoritarian.
I am concerned about illegal immigration. Not because I am racist, but because many of the people being let in are far more authoritarian in ways that threaten both the left and right.
Sometimes a strong government is needed to protect the commons and the rights of the citizenry. That is why I switched to voting republican and will continue to do so as long as they remain dedicated to this issue and the democrats continue to exacerbate the problem.
Yes, I see the problems with the maga movement. But they are nothing compared to the problems with the Sharia,Takfiri and criminal movements that are taking hold here.
Yes, the majority of immigrants are not in any of these groups, but those who are in them are clearly congealing into specific regions, taking over local communities and governments and deliberately working to expand.
The difference is not simply ideological, but in the authoritarian agenda of the ideology. These ideologies demand allegiance to types structural controls that are demonstrably none of their business and very unhealthy for humans.
I was reading about this situation clear back in 2015 with people talking about the exact U.S. regions and how they planned to expand in them. This process had already begun then and it is continuing today.
Drama queen’s get all the attention regardless of the medium. That’s been true all my life (74 years). It gets views and engagement, makes money, and gets people elected. That’s never going to change. World governments are packed with drama queens regardless of party.
