Artificial intelligence (AI) refers to technologies that exhibit some aspects of human intelligence and has been a prominent field of computer science for decades. AI tasks include everything from selecting objects in a visual scene to knowing how to structure sentences to even predicting stock price movements.
Scientists have been trying to build AI since time immemorial. The dawn of the computing era.of cutting edge approach For much of the past century, it has been necessary to create large databases of facts and rules and have logic-based computer programs make decisions based on these. But this century has seen a change, with new approaches that allow computers to learn their own facts and rules by analyzing data. This has led to significant advances in this field.
Over the past decade, machines have demonstrated seemingly “superhuman” abilities in all sorts of fields. Detecting breast cancer from medical imagesto play Devilishly tricky board games Chess and Go — and even more Predict protein structure.
Since the large-scale language model (LLM) chatbot ChatGPT launched in late 2022; growing consensus We may be on the verge of replicating a more general intelligence similar to that found in humans, known as artificial general intelligence (AGI). “I can't overstate how important a change this is for the field,” said Sarah Hooker, director of Cohere For AI, a nonprofit research institute founded by the AI company Cohere.
How does AI work?
Scientists can take a variety of approaches to building AI systems, but machine learning is currently the most widely used. This involves having a computer analyze data to identify patterns that can be used to make predictions.
The learning process is managed by an algorithm (a human-written set of instructions that tells the computer how to analyze the data), and the output of this process is a statistical model that encodes all the patterns discovered. You can enter new data into it to generate predictions.
Many types of machine learning algorithms exist, but one of the most widely used today is neural networks. These are a collection of machine learning algorithms loosely modeled after the human brain that learn by adjusting the strength of connections between networks of “artificial neurons” as they search for training data. This is the architecture used by many of today's most popular AI services, such as text and image generators.
Most of today's cutting-edge research is related to deep learning, which refers to the use of very large neural networks with many layers of artificial neurons. The idea has been around since his 1980s, but its use has been limited due to the large amount of data and computational requirements. And in 2012, researchers discovered that special computer chips known as graphics processing units (GPUs) can speed up deep learning. Since then, deep learning has become the gold standard in research.
“Deep neural networks are like machine learning on steroids,” Hooker said. “These are both the most computationally expensive models, but they are usually large, powerful, and expressive.”
However, not all neural networks are the same. Different configurations, or what are known as “architectures”, are suitable for different tasks. Convolutional neural networks excel at visual tasks, with connectivity patterns inspired by animal visual cortex. Featuring a form of internal memory, recurrent neural networks are specialized for processing sequential data.
Algorithms can also be trained differently depending on the application. The most common approach is called “supervised learning,'' where a human assigns a label to each piece of data to guide the pattern learning process. For example, add the label “cat” to an image of a cat.
In “unsupervised learning,” the training data is not labeled and the machine must solve the problem on its own. This requires more data and can be difficult to work with. However, because the learning process is not constrained by human bias, richer and more powerful models can be created. Many of the recent breakthroughs in LLM use this approach.
The last major training approach is “reinforcement learning,” which forces the AI to learn through trial and error. It is most commonly used to train AI systems and robots that play games (such as humanoid robots like Figure 01 or miniature robots that play soccer) to repeatedly try a task and respond positively or It involves updating a set of internal rules in response to negative feedback. . This approach powered Google Deepmind's groundbreaking AlphaGo model.
What is generative AI?
Although deep learning has had a series of great successes over the past decade, few have captured the public's imagination quite like ChatGPT's eerily human-like conversational abilities. It is one of several generative AI systems that use deep learning and neural networks to generate output based on user input, such as text, images, audio, and even video.
Text generators like ChatGPT work using a subset of AI known as “natural language processing” (NLP). The origins of this breakthrough date back to a new deep learning architecture called “Transformer” introduced by Google scientists in 2017.
Transformer algorithms are specialized for performing unsupervised learning on large collections of continuous data, especially large chunks of written text. It's good at this because it can track relationships between far-flung data points much better than previous approaches, and this allows you to better understand the context of what you're looking at.
“What I say next depends on what I said before. Our language is connected over time,” Hooker said. “That was one of the pivotal advances, the ability to actually see words as a whole.”
LLM learns by masking the next word in a sentence before guessing what it is based on the previous word. Because the training data already contains the answers, this approach eliminates the need for human labeling and allows you to simply collect large amounts of data from the internet and feed it into the algorithm. Transformer can also run multiple instances of this training game in parallel, allowing it to process data faster.
By training on such vast amounts of data, Transformers can produce highly sophisticated models of human language, hence the nickname “large-scale language models.” It can also analyze and generate complex long-form text much like text that humans can generate. Transformers didn't just revolutionize language. The same architecture can also be trained on text and image data in parallel, resulting in models like Stable Diffusion and his DALL-E that generate high-resolution images from simple descriptions.
Transformers also played a central role in Google Deepmind's AlphaFold 2 model, which can generate protein structures from amino acid sequences. This ability to generate original data, rather than simply analyzing existing data, is why these models are known as “generative AI.''
Narrow AI vs. Artificial General Intelligence (AGI): What’s the difference?
People are excited about LLMs because of the wide range of tasks they can perform. Most machine learning systems are trained to solve specific problems, such as detecting faces in a video feed or translating from one language to another. These models are known as “narrow AI” because they can only tackle the specific tasks for which they were trained.
Most machine learning systems are trained to solve specific problems at a superhuman level, such as detecting faces in a video feed or translating from one language to another, and are faster than humans. It's also much faster and performs better. However, LLMs like ChatGPT can revolutionize AI capabilities by allowing a single model to perform a wide range of tasks. They can answer questions on a variety of topics, summarize documents, translate between languages, and write code.
This ability to generalize what you learn to solve many different problems, such as DeepMind scientists being included in a paper published last year, makes us speculate that the LLM could be a step toward AGI. Some people. AGI refers to a hypothetical future AI that can master any cognitive task that humans can, reason abstractly about problems, and adapt to new situations without special training.
AI enthusiasts predict that once AGI is achieved, technological progress will accelerate rapidly, reaching an inflection point known as the “singularity,” after which breakthroughs will be realized. . There are also perceived existential risks, from large-scale disruptions to economic and labor markets to the potential for AI to discover new pathogens and weapons.
However, there is still debate as to whether LLM is a precursor to AGI or just one architecture within a broader network or ecosystem of AI architectures required for AGI. Some say LLMs are far from replicating human reasoning and cognitive abilities. According to critics, these models simply memorized a huge amount of information, recombining them in a way that gives the false impression of deeper understanding. This means that it is limited by its training data and is essentially no different from other limited AI tools.
Nevertheless, Hooker said he is certain that the LLM is making a big difference in how scientists approach AI development. Currently, state-of-the-art research uses these pre-trained, commonly working models and adapts them to specific use cases, rather than training models based on specific tasks. For this reason, these are now called “foundation models.”
“People are moving from a very specialized model that does one thing to a foundational model that does everything,” Hooker added. “They are the model on which everything is built.”
How is AI used in the real world?
Technologies like machine learning are everywhere. AI-powered recommendation algorithms decide what to watch on Netflix and YouTube. Translation models, on the other hand, allow you to instantly convert web pages from a foreign language to your own language. Banks will likely also use AI models to detect unusual activity on accounts that could suggest fraud. Surveillance cameras and self-driving cars also use computer vision models to identify people and objects from video feeds.
But generative AI tools and services are starting to penetrate the real world beyond novelty chatbots like ChatGPT. Most major AI developers are now introducing chatbots that can answer user questions on a variety of topics, analyze and summarize documents, and translate between languages. These models are also integrated into search engines, like Gemini was built into his Google search, and companies are also creating AI-powered digital assistants that help programmers write code, like Github Copilot. Building. It can also be a productivity tool for people who use word processors or email clients.
Chatbot-style AI tools are the most commonly found generative AI services, but despite their impressive performance, LLMs are still far from perfect. They make statistical guesses about which words should follow a particular prompt. They often produce results that demonstrate understanding, but they also confidently produce plausible but incorrect answers, known as “.”hallucination. ”
Generative AI is becoming increasingly common, but it is unclear where and how these tools will be most useful. And given how new the technology is, there's reason to be cautious about how quickly it's being rolled out, Hooker said. “It's very unusual for something to be at the forefront of technological possibility and be widely deployed at the same time,” she added. “That brings its own risks and challenges.”
