Five recent breakthroughs in AI you probably missed
The last two years have been immensely successful for the artificial intelligence creators, as researchers and engineers show that AI can be applied to improve almost anything.
Artificial intelligence operates in a black box, meaning human operators are unable to piece together how an AI makes its decisions.
2021 saw $93.5 billion in private equity poured into artificial intelligence according to Stanford’s 2022 AI Index Report. Investment into AI is expected to reach $500 billion by 2024 according to the International Data Corporation. PwC estimates that the global AI market will generate $15.7 trillion by 2030.
With years of heavy investing by public and private interests alike, artificial intelligence is quickly proving itself in new fields including health, robotics, agriculture among many others.
While learning machines continue to improve on a regular basis, AI has yet to meet the high expectations most have of it. While an AI can excel at a specialised task, an artificial general intelligence (AGI) still remains out of reach, for now.
An AGI is the holy grail for every AI researcher, generally understood to be an artificial intelligence able to understand, learn or carry out any task a human can. In the real world however, algorithms come up short when faced with complexity and the potential for change.
Despite limitations, companies like Google, Facebook, and Amazon have invested billions into the field, with 2022 seeing major breakthroughs for artificial intelligence in multiple fields.
1. AI gets jokes
The single largest advance in artificial intelligence came after Google released a 540-billion-parameter AI natural language processing model that exceeds average performance. Larger models are more efficient at ‘transfer learning’, which seeks to train neural networks that use less data and computing power.
A number of jokes Google's PaLM was able to interpret and explain.
This replaces competitor OpenAI’s GPT-3 natural language processing model with 175 billion machine learning parameters. GPT-3, originally introduced in May 2020, was hailed as being so accurate it was difficult to tell whether its text was written by a human.Google’s PaLM claims to be able to accurately explain why a joke is funny, with its sights now set on challenging problems of common sense and reasoning.
2. AI gets artsy
DALL·E 2 is a new AI system developed by OpenAI that can draw realistic images and art based on a text description you provide. First introduced in 2021, the second iteration of the system produces creative images 4 times more resolution.
An image drawn by DALL·E 2 based on the text: "An astronaut playing basketball with cats in space as a child's illustration."
Created to depict how AI sees the world and help people express creativity, DALL·E 2 quickly raises the question of what it means to be human and creative.
This builds on previous breakthroughs in AI-assisted visual processing technology, including converting black and white pictures into color, or generating lifelike 3D models of individuals from old photographs. DALL·E 2 represents the latest in ‘multimodal’ systems which are able to work with both images and text.
3. AI could make streaming cheaper
Deepmind is the company that brought you AlphaGo, which defeated world ‘Go’ champion Lee Sedol, before eventually giving rise to MuZero that mastered complex games such as Chess, Shogi, Atari and even strategy games like Starcraft.
The same company also developed Artuµ, which would be used by the US Air Force as a spy plane radar operator, co-pilot and mission planner. MuZero recently tackled the challenge of video compression, minimizing the amount of data required to stream a video by an average of 4 percent. That's not a small amount, given that standard compression codecs were achieved after decades of engineering.
Analysts state that video streaming made up the majority of internet traffic in 2021. With video streaming only expected to grow in the coming years, more effective video compression could reduce streaming costs, and increase download energy efficiency.
4. AI self driving competition heats up
This year you’ll get to witness a race from San Francisco to New York between self-driving car companies Tesla, Waymo and Cruz. In a face-off between competing methods, Tesla is set to remove radar sensors in favor of a vision-based system.
An image of how Tesla's full self driving (FSD) artificial intelligence sees the world.
Radar can see through fog and snow, unlike cameras, suggesting a deep confidence in Tesla’s AI. Other competitors have added more sensors, with both approaches set to be tested against the other.
Waymo utilizes LiDAR, similar to radar, but using laser pulses instead of radio waves. Tesla claims sensor data is easier to analyze however, offering a multifocal view of the road that surpasses human reaction and visibility. The company recently announced that its autonomous trial of self-driving taxis in San Francisco was successful.
Cautious optimism
While AI systems are slowly moving away from specialisation to hybrid systems, the unsolved challenge faced by most AI researchers is finding a way to integrate knowledge from multiple sources into consistent outputs.
Language models, used to interact with humans, answer questions, and write convincingly are being driven quickly by significant big tech investments into voice assistants such as Alexa and Siri.
While 2022 is far from over, there is a clear trend towards combining traditional challenges of speech, perception and language instead of treating them as separate functions. That’s harder than it sounds, but if the last few years are any indicator, artificial intelligence will continue to drive innovation in multiple sectors for decades to come.
Copyright © 2024 TRT World.