So many people worry about artificial intelligence. the central worry, for many people is if the AI can reach general levels of consciousness. I think the worry about this is the general lack of a good metaphor for understanding what we're looking at.
Generally people don't have a good way to understand what AI is and how software production is being part of this process. In fact, there is a very simple way.
This paper
or this you-tube video
Introduce the work of a group that has trained pigeons to look for and identify cancerous breast tissue.
Briefly, they place a pigeon in front of a screen. The pigeon is shown a scan of breast tissue that may be cancerous. The pigeon pecks on two buttons: one labeled "cancerous" and the other labeled "not cancerous."
If the pigeon is correct, it receives food; if not, it receives none. The pigeons quickly learn to be about 80% accurate in identifying cancerous material, which isn't quite as accurate as a human but not bad from 48 images. However, when their scores are combined, the final system is not more accurate than a clinician.
This is quite amusing, except if we substitute AI for pigeons, it becomes quite miraculous, and we should expect AI to replace human clinicians soon.
The difference between artificial intelligence and pigeons becomes apparent when working with pigeons. You realize that much of the real magic lies in how you present images to the pigeons. It is this data preparation that allows the pigeon brain or AI to process the information. By manipulating digital information correctly, we can make the job of the pigeon or AI easier or harder.
Pigeons are tetrachromats, meaning they have four types of cone cells in their retinas that allow them to see a range of colors. However, their color perception is thought to be most sensitive in the short-wavelength (blue) and medium-wavelength (green) regions of the spectrum. They may have difficulty distinguishing between certain colors, particularly those in the red-orange range. Therefore, by applying some Photoshop filters to the original images from the paper, we could make the pigeon's job easier, quicker, and more accurate. More example
If you were a a doctor whose job it was to look at pictures of biopsies of cancerous cells all day you might wonder if your job is at risk. I think most people at this point would wonder who is looking after the pigeons, since they aren't doing the actual work.
Well, this is similar for AI. We have data centers burning electricity and people looking after the machines in the data centers. Would a group of dedicated pigeon fanciers be any more different or expensive? Sure, we could put them into remote data centers, hide them away from the public eye, but you would still have to charge for the energy (seed) and staff time. Given how good we are at growing broiler chickens in factory farms, you do wonder which would have the economic edge Datacenter/pigeon centre.
Most of what AI work does involves figuring out how to present the world to pigeons or AI and getting them to press the right button for the right data.
For example, we could train pigeons to reviewcollege resumes for graduate applicants. the applications would be converted to images and then pigeons peck on a button labeled 'make offer' or 'don't make offer' based on previous data. So saving time and money.
Could we build Pigeon ChatGPT?
There's nothing in principle stopping someone from building a pigeon version of ChatGPT. However, it might require more than one pigeon. The key area is how to present the pigeons with the text. Using something like a Word Embedding (Word2Vec) would be necessary to convert text into picture elements along with the pigeon equivalent for an attention mechanism. The key part is presenting the neural network or pigeon with clear information.
So, could our PigeonGPT achieve general consciousness? Perhaps if we had a thousand pigeons sitting together, each with different screens performing different sub-elements, would it create a 'mind'? If you're not worried about a PigeonGPT then why should you be about a neural network?
