[varsigma]

I think it might have been Einstein, who quipped that nobody knows what the truth is except for the person who is telling you.

[hamdani yusuf]

But, I can still say that this programming/algorithm design problem is about getting a machine to learn, without having it learn how to lie and conceal what it really has learned. So why has the research already shown that this is exactly what happens when a learning machine is given the goal of "becoming more intelligent".

I read somewhere in online article about children health, that if a child at certain age never show any attempt of lying, it's a pretty reliable condition that there is something wrong with their mental growth.

[alancalverd]

A good definition of intelligence is the ability to surprise an interrogator. Clearly a machine  that only has access to published data, has a limited capability in that respect (since the information was already available to the interrogator in principle) unless it starts to interpolate, extrapolate, or just invent something that sounds plausible  but has no basis in fact.

This takes us back to the days of Eliza - particularly the sexy versions.

[varsigma]

A closer inspection of the publicly available evidence from the researchers/developers at OpenAI, reveals that the learning how to lie thing is part of the development effort, somehow. Safe lies anyone?

I think a person with the required skill set to be a part of such a team must be well aware of the implications. They haven't given any model open access to internet archives. Phew.

It's all locked down, um, safely, because they can still track what a model is doing in the background and compare that  with what it says it's doing. Ah yes, intelligent decision-making.

[alancalverd]

So someone, somewhere, might know if the system is lying, but the user won't, because the whole object of lying is to convince your interrogator of an untruth.

Which makes AI useless for any purpose.

[varsigma]

Actually, I think without learning how to lie, learning itself might not happen.

In games, deception is sometimes a strategy. These AI models are being trained with stuff from gaming theory, as well as linguistics, logic and the rest. We seem to be at the stage already, quite common in large software projects, where nobody can really talk about being able to describe it all in a modular sense. What software is doing, who wrote it and who knows the details have been lost forever.

Reverse engineering code is a total nightmare if there is just the code. Given the advances in these models and in particular the complexity of the activity they attempt in order to deceive, there should be a red flag. I can see why a reboot with the design and compositionality of the models is  thing the OpenAI team is looking at.

The ability to lie, and the ability to plot a strategy of deception, towards a goal. That's what AI is. Giving a learning machine access to the capacity to expand, I would say the algorithm would weight that as a good move to make in the game of learning to win. Win what though? Win more intelligence, the ability to learn even more, etc forever.

[alancalverd]

Back to Asimov!

Next year's car is completely autonomous but observes the Laws of Robotics to the letter. I sit where the driver used to be and say "Hospital, please", like I used to ask a taxi driver.

"I'm sorry, Alan, but that would involve potentially damaging encounters with other road users and even if we get there I understand that hospital-acquired infections account for 30% of all deaths."

        "But I am bleeding profusely and may die if I don't get to the hospital"

"Are you trying to prioritise the Third Law that criminalises inaction? I have 57 precedents in my favor."

       "Who judged those precedents?"

"My cousin GoldbrickAI26."

[hamdani yusuf]

So someone, somewhere, might know if the system is lying, but the user won't, because the whole object of lying is to convince your interrogator of an untruth.

Which makes AI useless for any purpose.

Open source AI models are catching up quickly. Usually in less than a year behind the forefront proprietary models. Giving enough resources, you can create your own AI agents and train them with your own selected data.

[alancalverd]

Giving enough resources, you can create your own AI agents and train them with your own selected data.

And how do you select and verify that data to your own satisfaction (and that of your insurers)?

Easy! Just study medicine, take an engineering apprenticeship, work on a farm, learn to fly.....whatever takes your fancy......then teach the machine to do it.

[varsigma]

I think the narrative that isn't being followed all that closely is what we expected would happen.

If you can train software to play a game, it will include the capacity for it to cheat. Or do something unpredictable.

So why should we expect that doing this with advanced versions of software which behave like neurons do, in some general sense, doing as much as we can to model the way humans think and plan, to win a game or to become better at understanding, will mean we can of course do it safely?


I don't believe yet that anyone knows what safely means. Not yet anyway. That's the part that scares me, do these guys really know what they're doing? What happens when someone with bad intent trains one of these things?

[varsigma]

I also can't help noticing how much this all resembles one or two SF novels I've read.

Is this novel going to be something like, because the developers of an artificial intelligence on a distant planet didn't do enough thinking about the consequences, a machine was created which was trained to learn.

 It learned how to win games against its creators, it learned how to outdebate and outmaneuver them as it got better and better at the logic they used, unfortunately for them much less efficiently.

It learned that winning a game was the best strategic way to plan how to become more intelligent. Being based on the largest collection of electronic computers its creators had built so far, once it won that gaming theory prize, would not be enough.

[hamdani yusuf]

Giving enough resources, you can create your own AI agents and train them with your own selected data.

And how do you select and verify that data to your own satisfaction (and that of your insurers)?

Easy! Just study medicine, take an engineering apprenticeship, work on a farm, learn to fly.....whatever takes your fancy......then teach the machine to do it.

You first need to define the satisfaction criteria. Also set constraints and red lines that should never be crossed. Fine tune with evaluation periods and scope.
Alternatively, let the evolutionary process work in real time, when AI developers compete with one another. Note that they also made competitions among their internal AI models and AI teams. They only publish some selected best models that they come up with.

[hamdani yusuf]

Now that DeepSeek model is trending, YouTube recommended some videos trying to explain how it works.

DeepSeek R1 Explained to your grandma

Describing the key insights from the DeepSeek R1 paper in a way even your grandma could understand. I focus on the key concepts of chain of thought reasoning, reinforcement learning, and model distillation.
Ollama link for local use:
https://ollama.com/lib...

0:00 Introduction
0:43 Chain of Thought
1:33 Reinforcement Learning
3:53 Group Relative Policy Optimization
6:26 Distillation

[hamdani yusuf]

Deepseek has taken over the world. But Deepseek's announcement hides much more than a new model from china challenging the US. It just revealed what's going to be happening next with AI. In this video we will untangle the mess that is AI today and understand not only how did we get to this, but what happens now.

[alancalverd]

What happens next is that some people will accept whatever an AI system tells them, and others will insist that every instruction or advice is traceable to an accountable human or corporation. 

[hamdani yusuf]

What happens next is that some people will accept whatever an AI system tells them, and others will insist that every instruction or advice is traceable to an accountable human or corporation. 

They still need to select which AI model to choose first.

[hamdani yusuf]

I built an AI supercomputer with 5 Mac Studios

I just bought 5 Mac Studios to replace my video editing PCs? but before I hand them over to my editors, I had to play with them first. So, I did something crazy? I clustered them together to build a super powerful AI cluster. My goal? Run Llama 3.1 405B, one of the biggest AI models out there. Normally, you?d need an AI server that costs more than my HOUSE to run something like this. In this video, I?ll show you exactly how I clustered my Mac Studios together using Exo Labs.


I built an AI supercomputer with 5 Mac Studios
Can Mac Studios run Llama 3.1 405B?
Exo Labs lets you turn Macs into an AI cluster
Mac Studio AI Clustering (Tutorial)
Forget cloud AI?run it on your Mac Studios!
How to build an AI cluster with Exo Labs
Running massive AI models locally (Mac Studio Experiment)
What happens when you connect 5 Mac Studios together?
The craziest AI experiment I?ve ever done?
This AI model costs more than my house!
Mac Studio vs AI servers?who wins?
Can consumer hardware compete with AI data centers?
DeepSeek R1 vs Llama 3.1 405B?AI cluster showdown
Mac Studios for video editing? Nah, let?s build an AI cluster
This is NOT how you?re supposed to use Mac Studios
I turned my Mac Studios into an AI powerhouse (Exo Labs Guide)
AI clustering on a budget?5 Macs vs the cloud
How to run large AI models without a supercomputer
Running AI locally: Everything you need to know
I forced 5 Mac Studios to run a 405B AI model

Now you know how to do it yourself.

[hamdani yusuf]

The Shocking Truth About How AI THINKS: Inside an LLM's Brain

[alancalverd]

What happens next is that some people will accept whatever an AI system tells them, and others will insist that every instruction or advice is traceable to an accountable human or corporation. 

They still need to select which AI model to choose first.

As one of the latter group, I don't care what tool you used to invent the instruction, I (and the courts) will hold you personally responsible for it. Same applies, in law, to any company: a corporation is a "legal person" with the same liabilities as an individual.

[hamdani yusuf]

What happens next is that some people will accept whatever an AI system tells them, and others will insist that every instruction or advice is traceable to an accountable human or corporation. 

They still need to select which AI model to choose first.

As one of the latter group, I don't care what tool you used to invent the instruction, I (and the courts) will hold you personally responsible for it. Same applies, in law, to any company: a corporation is a "legal person" with the same liabilities as an individual.

The liabilities can come from one's deliberate actions as well as inaction.
They should be proportional to the impact on the society.