so far sSince deep learning hit the mainstream in 2012, the hype surrounding AI research has often outweighed it. this is real. But over the past year, a series of breakthroughs and major milestones suggest the technology may finally be delivering on its promise.
Despite the obvious potential of deep learning, over the past decade, regular warnings about the dangers of runaway superintelligence and possible technological unemployment have led most AI systems to perform tasks such as identifying images of cats and speaking English. tempered by the fact that it was dedicated to providing questionable translations from to English. Chinese.
However, in the last year there has been an undeniable gradual change in the capabilities of AI systems in fields as diverse as creative industries, basic science, and computer programming. Additionally, these AI systems and their outputs are becoming increasingly visible and accessible to the general public.
Nowhere is progress more evident than in the burgeoning field of generative AI. Generative AI is an umbrella term for many models that focus on creative tasks.
This is largely thanks to a model called Transformers, which was actually first announced by Google in 2017. But the results they produced in 2022 blew previous iterations out of the water.
The most famous of these is ChatGPT, an AI chatbot based on the latest version of OpenAI’s GPT-3 large language model. The service, which was opened to the public at the end of November, surprise people You have an amazing ability to hold natural-sounding conversations, answer complex technical questions, and compose compelling prose and poetry.
Earlier this year, another OpenAI model called DALL-E 2 took the internet by storm. surreal images In response to bizarre prompts like “Raccoons playing tennis at 1990s Wimbledon” and “Ancient Roman Spider-Man.” Meta went one step further in her September, short video clip From text prompts, Google researchers were even able to create AI. generate music It will play in the style of an audio clip.
The impact of this explosion of AI creativity and fluency is difficult to measure right now, but it has already fueled predictions that it could. replace traditional search engines, kill college essays, lead to death of art.
This is due to the increased functionality of these models and theMeMake it more accessible with services like ChatGPT, DALL-E 2, and text-to-image generator Midjourney (at least for now), and they’re free for everyone.Going further, an independent AI lab S.table D.iffusion is even Open sourcing text-to-image AIso that anyone with a reasonably powerful computer can run it themselves.
AI has made progress in the last year, even on more mundane tasks. In January, Deepmind announced AlphaCode, his AI-powered code generator. Comparable to the average programmer at a coding competition. in a similar vein, GitHub co-pilot, of AI coding tools developed by GitHub and OpenAI moved from prototype to commercial subscription service.
Another big bright spot in the field is the increasingly important role of AI in basic science. In July, DeepMind announced that the groundbreaking AlphaFold AI Predicted structures of almost all proteins It is scientifically known and has the potential to revolutionize both life sciences and drug discovery. The company also announced that it held training in February. AI controlling swirling plasma It’s inside an experimental fusion reactor.
And while AI seems to be slowly moving away from the toy problem that the field has been obsessed with over the last decade, it’s also making big strides in one of the mainstays of AI research: gaming.
in November, M.eta showed off an AI ranked in the top 10% of players ranked. board game diplomacy, requires a challenging mix of strategy and natural language negotiation with other players. That same month, a team at Nvidia trained his AI to Play the complex 3D video game Minecraft Use only high-level natural language imperatives. And in December he’s going to DeepMind Cracked the devilishly complex game Strategowhich includes long-term planning, bluffing, and a healthy amount of uncertainty.
But it wasn’t all smooth sailing. While the output of generative AI like ChatGPT is superficially impressive, many are quick to point out that it is very compelling. bullshit generatorThey are trained with a huge amount of texts of varying quality from the internet. And ultimately all they do is guess what text is most likely to come after the prompt, and have no ability to judge the veracity of their output. There is concern that the internet will soon be flooded with seemingly nonsense.
This was revealed with the release of Meta’s Galactica AI, which was supposed to summarize academic papers, solve math problems, and create computer code to help scientists speed up their research. The problem was creating compelling-sounding material. completely wrong or very biasedthe service was pulled in just three days.
Bias is a critical issue for this new breed of AI. This AI is trained on vast amounts of material from the internet, rather than the more carefully curated datasets that previous models were fed with.A similar problem has surfaced in ChatGPT, where despite the filters put in place by OpenAI, it has been tricked into that’s all wHait and Asian Men Make Great ScientistsAnd the popular AI image generation app Lensa was called sexualZefemale portraitIn particularrye Asiande’ssst.
Other areas of AI a A substellar year. One of his most touted real-world use cases, self-driving cars, has seen a major setback. Ford and Volkswagen backed Argo closureTesla fend off fraud claims About delivery failureWhathmm sfairy-dA growing chorus of voices claiming to be alive The industry is in a rut.
despite obvious progress that is Like Gary Marcus, deep learning reached the limit, This is because they cannot truly understand the material they are being trained on, and instead are simply learning to make statistical connections that can produce convincing but often flawed results.
But for those behind this year’s most impressive results, 2022 is just a fraction of what’s to come. The next big breakthrough is Multimodal model From text to images to audio, it combines increasingly powerful features. It remains to be seen if the field can maintain its momentum in 2023, but in any case this year could be a turning point for his AI research.
Image Credit: DeepMind / Unsplash