We’re all talking about AI. How could we not?
In coffee shops, boardrooms and group chats, the subject serves as conversation-starter, ice-breaker and idle chatter.
But are we doing it right? We’re using terms interchangeably, when they mean very different things. We’re mixing up our tech. We’re even using the phrase “AI” wrong. What are the most common mistakes? Take a look.
* What is and isn’t AI?
Any technology that is at least as smart as the average human,
We’re all talking about AI. How could we not?
In coffee shops, boardrooms and group chats, the subject serves as conversation-starter, ice-breaker and idle chatter.
But are we doing it right? We’re using terms interchangeably, when they mean very different things. We’re mixing up our tech. We’re even using the phrase “AI” wrong. What are the most common mistakes? Take a look.
* What is and isn’t AI?
Any technology that is at least as smart as the average human, and can therefore perform tasks a human would generally be required for — such as data-analysis, script-writing or creating new protein structures — is considered artificial intelligence.
There is a fair amount of confusion over what isn’t AI, and much of this stems from the high levels of “personalisation” we encounter online today.
When a platform responds to our usage and preferences, we tend to assume some amount of AI is involved, but it’s usually just an algorithm at work.
There is no need for artificial intelligence at this level of functionality.
Even when Netflix responds to “your viewing history” or a bot player “learns your moves” in a videogame, it is just computer programs matching input with output — the way an ATM or a currency converter might do.
* Do chatbots count as AI?
This is an interesting one. A lot of AI programs function like chatbots. (The user types in a question or a prompt, and the program responds.)
So, how does one tell the difference between an AI and non-AI bot?
It all boils down to a simple question: Can the chatbot generate something new, or can one only ask it to reflect the information it has been fed?
Generative AI programs can consult their vast datasets and the internet, for instance, to answer a query with a response that they have framed, based on their interpretation of the question, their levels of access, and their understanding of the material accessed.
Chatbots such as Amazon’s Alexa (at least the version still in use in India; the US now has a true-AI version available) are not generative and, as a result, can only perform a limited range of functions: answer certain queries, operate certain smart-home devices, or find certain kinds of songs within a streaming service.
* Is it evolving?
A key element of AI is that it learns as it goes. This means that the same question asked two months apart can generate a different, far-more-evolved answer the second time around.
Can you ask an Alexa a question today, and expect it to know better a week from now? Of course not. Put like that, it becomes far easier to tell the difference.
AI’s evolution can involve new capabilities too. Multimodal AI models can process text, images and video; generate code; then learn to create art in the style of Studio Ghibli. Such programs are now being viewed as the software equivalent of the Swiss Army knife. It’s anyone’s guess what tool will be added next.
* Where’s the debate?
A key reason AI is able to learn as it does involves something called the adversarial approach (a term that really should have gone mainstream by now).
An adversarial program is designed to challenge and check on itself all the time.
Whether the program is asked to draw a hat, write a poem or create a new protein structure, there is one half that starts to generate the output — and a crucial other half that immediately begins to critique it.
Based on this latter half’s understanding of what a hat looks like, what kind of hat the prompt specified, and the images of hats available to it, it can tell its creator half: “No, that doesn’t look right.” “No, it still doesn’t look right.” And on and on, until both halves arrive at what they deem to be the best possible output.
As an AI algorithm learns, vitally, both halves improve. The creator is able to do better from draft one, and the adversarial half is able to challenge it more effectively.
The results are plain to see. Midjourney alone has gone from being a laughable but intriguing toy in 2022 to something that could potentially start a war, because it is impossible to tell its fakes from our real world any more.
* Prompt attention
As the language of AI evolves, here are three terms you can expect to encounter.
- Tokens: This is the term for the chunks of text (words, parts of words or even individual characters) that AI models break prompts down into. Large Language Models or LLMs — AI programs where text prompts serve as the things that turn the knobs, to generate the best response possible — work by segregating bits of text so they can be analysed and cross-referenced. They do this by assigning tokens to each bit of text. The more tokens a model can process at once, the better the results often are.
- Hallucinations: When the dataset an AI program was trained on isn’t good enough to address a given query, a user may be met with not so much a weak response as a “hallucination”. This may involve absurd leaps in logic, fabricated details or, particularly in the early years, sheer nonsense. (Click here to see how the first version of ChatGPT responded, when asked about the meaning of life, and why it crossed the road.)
- Frontier models: This is the term for unreleased models now being developed that are expected to far exceed the capabilities of the AI programs we use today. Promises range from programs that can “give us” the solution to world hunger to others that can “solve” the climate crisis, or at least solve the problem of their own resource-guzzling. How many of these are hallucinations? Well, humans have always been susceptible to those.
One Subscription.
Get 360° coverage—from daily headlines
to 100 year archives.
Archives
HT App & Website