The Age of A (quar) I US
Along with the Internet, Email, and Cell Phones, and Software, the age of computers and chips is enhancing Artificial Intelligence to Large Language Models, Chat, Creative Writing and even Programming. AI has actually been advancing since WWII, when Alan Turing taught the first computers to compose music and to play chess.
A good and comprehensive source for technical AI is the website unite.ai, and that is the main source of this short talk. The Breakthrough of 2022 in Science Magazine notes the creative writing of ChatGPT and artistic creativity of Dall-E2. At UCI Olli we have had an Open Forum session where ChatGPT was well illustrated. There are many articles in common media illustrating this, and you can try ChatGPT for a free trial. The subscription with guaranteed access during overloaded periods cost $20 a month. Microsoft has its creation AI in the New Bing. Apple has lagged in creation AI because of its concern with privacy, waiting until it can put it on the user’s devices.
All of these AI are computer structures based on a Deep Neural Network, with an array of mathematical “neuron” vertexes in layers, fed by inputs from the preceding row of “neurons”, similar to how many neurons in the brain stimulate a later neuron’s firing by emitting chemicals at the synapse. Here is the Deep layering and mathematical connectivity picture. In the network figure, the input is red, and the output is blue.
The Network starts out dumb, and is exposed to connected words in documents, or pixels in a picture. The vertexes are rewarded or the triggering potentials numerically enhanced when the output resembles the successful wording or pictures. Thus the network is trained to artificially mock up intelligence. The recent progress in AI is due to larger datasets for training, more experimentation in network rules, faster processing chips, the use of more neurons, and a lot of feedback from volunteer or paid evaluators. Once a network is trained in a specific area, it can be applied for many uses by many companies or users.
These advances and projected futures are cited on the unite.ai website. The predicted world expenditures on AI are $110 billion, mostly in the tech sector. Businesses have already been using AI in chats for aiding communications, and businesses point out that it can take a month to train a personal advisor, whereas the once trained chatbox can be multiplied freely.
NVIDIA which makes visualization chips for games has expanded into a leading AI creator and marketeer. It’s stock valuation has quadrupled in a year, and it is now a trillion dollar valued company.
Unite.ai lists categories where AI is being developed:
Marketing, with conversational AI, ad targeting, and personalized content;
Legal Services;
Sales;
Healthcare, with drug discovery, robotic surgery, and virtual therapists; and
Finance.
AI targeting was used by Obama in the 2012 election, and by Trump in 2016. It will be heavily used in 2024.
Google Assistant (“Hey Google”) in 2020 already had one billion users in 90 countries, and 500 million users a month. It gives short verbal answers, but refers you to longer sources in the Assistant app. Amazon’s Alexa had 65 million Echos sold last year. 100 million have already tried ChatGPT, which had 1.8 billion uses in the past month.
To show the rapid increase in computer capabilities, GPT-3 used 175 billion parameters in its neural net. In 2022, the speed of computation was up to 35 billion FLOP/second per US dollar (FLOP is Floating Point Operations such as mathematical ones). The cost of training a GPT-3 network went down a factor of 10 between 2020 to 2022, where it became $450,000. By 2030, the cost is expected to be only $30. The cost of a billion informs from an AI in 2022 was $10 million, or a penny for their thoughts.
Warnings about the use of AI are that ChatGPT was only trained up to 2021. The newer GPT-4 is more up to date. Your use of the apps may not be private, and your complete online data will undoubtedly be used in targeting ads. The apps may contain bias, since the internet contains many views. Nothing new there, but leading suggestions on Google search tend to be of good quality, after the ads. The creative output can deliver misinformation, and is subject to “hallucinations” in its creativity.
Concerned about the problems in AI, the non-profit Center for AI Safety (safe.ai), advises on companies installing safeguards. These are to prevent: the Weaponization of AI; Destabilization of society from AI misinformation; Monopolistic control of AI technology; Surveilance or oppressive censorship; Overreliance on AI; Bias; or Transference of wealth.
