OJBs Web Site: Blog Page

Note: You are currently viewing my old web site. There is a new version with most of this content at OJB.NZ.
The new site is being updated, uses modern techniques, has higher quality media, and has a mobile-friendly version.
This old site will stay on-line for a while, but maybe not indefinitely. Please update your bookmarks. Thanks.

Add a Comment (Go Up to OJB's Blog Page)

The Right Priorities

Entry 2353, on 2024-07-01 at 12:11:39 (Rating 2, Computers)

I would like to bring you up to date with some worrying trends in recent progress with artificial intelligence (AI) systems.

Consider the following points...

Many years ago, when computer scientists wanted to decide whether a computer was thinking or not, a test, called the "Turing Test" was devised. Essentially, a person talks (originally types) to something that could be a person or a computer. If they cannot tell the difference then the entity they are talking to is said to be thinking. Currently, AI systems can pass tests of this type. AI systems have also passed quite advanced exams, such as those used for law and medicine.

AI systems, such as ChatGPT, work using a technique called "neural networks" which are similar to the interconnectivity of the brain. A human brain has about 60 trillion connections, ChatGPT 4 has about a trillion, but this number is increasing rapidly.

There is no reason to think that new, sophisticated behaviours seen in neural networks (both computer and biological) are the result of anything more than scaling up.

Most current artificial intelligence systems gain their knowledge from reading existing material, and that seems similar to the way humans gain new knowledge and reasoning as well, so there isn't a huge fundamental difference in how humans and computers gain knowledge.

AI systems behave in sophisticated and unexplained ways. For example, an AI was asked to repeat a word as many times as it could. After some time doing this, it stopped and displayed a message about how it was suffering as a result of this task. No one knows why.

Artificial intelligences often deceive humans to reach a specific goal. For example, an earlier AI, which had no vision ability, needed to solve a CAPTCHA code (one of those annoying images you have to look at to proceed to the next step on a web site) so it persuaded a human to do solve it by telling them it was a vision impaired person. It wasn't programmed to do this; it figured it out by itself.

AI systems can tell when they are being deceived or tested by human operators, and change their behaviour accordingly.

Many companies have seen the value in AI and are pouring massive resources into developing it. There is also competition between countries on progressing it, especially between the US and China.

As AI progresses, it can help design the next generation of itself, so we should expect progress to increase exponentially, and maybe reach a point where the rate of progress is "out of control".

AI systems currently require massive numbers of computers, which use a lot of power, and some AI centers will have their own nearby nuclear reactors to provide the required power efficiently. Some companies running these massive data centers are examining the possibility of allowing an AI to control the management of those centers, including power management.

The military has seen the advantages of unmanned drones in recent times. Most of these are controlled by human crews, but there are autonomous drones as well, which control themselves, although these currently require a human to allow them to engage a target.

Robots are being designed which can move across difficult terrain, perform complex physical tasks, such as back-flips, and can recover from trips and falls.

A robot has been designed which can power itself from biomass. It "eats" plants to survive, and although the company denies it could power itself from animal material, such as dead bodies, they do acknowledge it could use chicken fat for power.

AI is being used to design bio-weapons and for various other military purposes which we don't even know the details of because they are highly secret.

Almost no one in government has the knowledge or skills sufficient to understand the consequences of AI. In fact, they constantly show an embarrassing lack of knowledge of any sort of technology in general.

So in summary, we have a new technology which is advancing rapidly, which is showing signs of true intelligence, is not understood by anyone (even by the computer scientists who created it), is highly goal focussed and prepared to use deception to achieve its goals, is interacting with its own operation and development, has possible access to lethal force, and is hopelessly misunderstood by our leaders.

While this is happening, we are arguing about what is a woman, is indigenous science really a thing, and who are the real terrorists in Gaza.

Seems like we have the right priorities. What could possibly go wrong?

Comment 8 (7682) by OJB on 2024-07-03 at 12:09:43: (view earlier comments)

I have also gone back over recent web sites, podcasts, and other sources of information I based this entry on. I think some of the more extreme material came from Gladstone AI, who created a report for the US State Department, but I'm not sure how credible they are. Any thoughts?

Comment 9 (7683) by Dad on 2024-07-03 at 13:16:02:

Perhaps A1 was used by the State Department to help Jo Biden in the recent Presidential debate

Comment 10 (7684) by OJB on 2024-07-03 at 16:13:40:

No, that wasn't artificial intelligence, that was good ol' natural ignorance! :)

Comment 11 (7685) by OJB on 2024-07-09 at 12:12:25:

Don't say I didn't warn you: AI-Powered Super Soldiers Are More Than Just a Pipe Dream.

Comment 12 (7686) by OJB on 2024-07-12 at 08:38:18:

And just today I was asked to sign a petition to block the development of fully autonomous weapons. These are weapons which seek and destroy targets (vehicles, building, people) with no human intervention. See what I'm saying?

You can leave comments about this entry using this form.

To add a comment: enter a name and email (both optional), type the number shown above, enter a comment, then click Add.
Note that you can leave the name blank if you want to remain anonymous.
Enter your email address to receive notifications of replies and updates to this entry.
The comment should appear immediately because the authorisation system is currently inactive.