Aardvark Daily

Please visit the sponsor!

The new best game online is crashing AI

I am no longer a gamer.

Back in the early days of computing, and we're talking the late 1970s through 1990s, I was an avid gamer and spent far too many hours behind a keyboard engaged in such activities.

The original text "Adventure" game was an obvious milestone game but I also wasted far too much of my life plaing PacMan and Missile Command on my Atari 400 -- or tediously crafting code to create my own rather lacklustre games.

Over time however, the alure of computer gaming waned and by the late 1990s the attraction was completely gone.

I just never warmed to the likes of Tetris and such, with my last remaining memories being those of the 2D platform games like Captain Commic or Duke Nukem. Castle Wolfenstein was the only 3D first-person-shooter I've ever played and even when I bought my daughter a PS1 I just never spent any time using it myself.

However, now I have found a new game and it's fun.

There are no flashy Real-Engine powered photorealistic graphics with raytracing and other cool technologies involved.

You don't need the latest multi-cored processor with 3D caching tech or a GPU that costs more than your Jap-import car is worth either.

My favourite game right now is Google Gemini.

"But that's not a gain, it's an AI chatbot" I hear you saying.

Perhaps so, "but we can use it as a game" (to quote one of my favourite lines from that timeless TV series The Biederbeke Connection, the third instalment of the trilogy).

I figure that being skilled in AI query engineering is a valuable skill to have so I've been spending lots of time playing around, using Gemini as my target because it's free and seems to be perfectly adequate for the job.

My goal has been to gain a better understanding of how to gain better control over the system and sidestep the built-in failsafes and protections that Google has added to prevent what it considers unsafe use or mis-use of the service. Not that I want to mis-use it but I do want to know just how difficult it is to circumvent the protective measures that are in place.

I'll probably write a whole column on the results of my experimentation but suffice to say that at this stage I can bring Gemini to its knees pretty quickly and easily. By "to its knees" I mean that it'll either complain that it has "encountered an error" or simply becomes totally unresponsive -- spinning its little wheel forever (or for at least an hour or more).

The key to these circumvention tactics is indirection.

If the AI won't answer a question directly then you need to come up with a method of adding a layer of indirection to your query. Such indirection often works around the protections that are built in and produces the required results -- or just a total meltdown producing a zombie-state.

After many hours of probing and analysing results I think I have a pretty good handle on how these AI systems work (or Gemini at least) and I have to say that the more you learn, the less impressive these AIs become. Sure, on the face of it they seem super-clever and intelligent but that's a pretty thin veneer to a rather simple pattern-matching engine that seems to rely on raw computing power to sift through millions of possibilities to come up with a "best overall match" output, wrapped in fluffly human-friendly verbage.

It reminds me of those early days of computing, back in the 1960s, when you'd see computers on TV documentaries answering questions posed by a reporter or journalist. A the time it seemed awesome but in hindsight it was clearly just some simple programming. That's the state AI seems to be at today. AI appears to be much smarter than it really is.

Sure, AI is super-good at some things we're not. For example, I see that AI has apparently designed new anitbiotics for use against resistant superbugs. Sounds hyper-intelligent but in reality, what AI is probably doing here is throwing all the possibilities against a wall and seeing what sticks. This is just like the early days of computing when we marveled that a box of blinky lights could, in a single minute, perform calculations that would take a room of human beings years to perform.

There's no magic, there's nothing to see here except the incredible speed with which modern computers can do mathematics, logical operations and move data around.

AI is great -- but it's not your friend and it's not all that mysterious.

Carpe Diem folks!

Please visit the sponsor!

Here is a PERMANENT link to this column

Rank This Aardvark Page

Change Font