Hello. Welcome to Eye on AI.
In the 19th century, major shipping companies competed to be the first to get seasonal tea from China to London. The first teas to hit the market had a premium, so they were expensive. So investors built clipper ships that were faster than ever before, carried more sails, and had sleek copper-bottomed hulls. To encourage captains to push their ships to their limits, tea merchants offered prize money to the first crew to arrive at London's docks. The first ship also earned the right to display a special “Blue Ribbon” pennant.
The race was expensive, but for the captain and his crew, it was just as much about ego and honor. Risk was also taken into consideration, as the Clipper was built for speed rather than stability. They required great skill to navigate.of Tepin and Ariel, The Tea Race of 1866 split the prize after an epic battle around the world, with ships arriving at the mouth of the Thames within an hour of each other after 99 days of voyage, but both ships later sank. did. In fact, all five of his clippers entered in the 1866 race were eventually wrecked or lost at sea.
What does this have to do with AI? Well, today it feels like we're watching his 21st century version of the Great Tea Race with AI. Leading AI companies are leapfrogging each other in various dimensions of capability and performance in this contest. This contest seems to be a bit about money, but the ego and prestige of getting credit for being first to market with a particular feature is very important. The seasonality of this rapid release of new models is also reminiscent of the arrival of new tea chests in London every September.
Over the past two weeks, OpenAI and Google have both announced new AI models and product features at a breakneck pace, each pushing the boundaries of what the technology can do. First, OpenAI gave ChatGPT the ability to remember past conversations with users as well as personal details and preferences. Google then widely released its most powerful model, Gemini 1.0 Ultra. This was followed by a limited launch of a new Gemini 1.5 Pro model with similar features to the Ultra, but in a smaller and cheaper package. However, what makes 1.5 Pro special is its extremely large “context window.” This is the amount of content you can type in the prompt. 1.5 Pro can analyze 1 hour of video, 11 hours of audio, or approximately 7 books' worth of text. And on Thursday, OpenAI showed off Sora, a new text-to-video generation model that can generate one-minute videos of him in stunning quality.
The pace shows no signs of slowing down, with indications that more announcements will be made in the coming weeks. Moreover, these developments will no doubt force other AI companies to respond more quickly. Cristóbal Valenzuela, CEO of Runway, which was arguably the leader in the field of text-to-video generation, responded to the release of OpenAI's Sora by simply tweeting, “The game is on.” Google DeepMind released a model called Lumiere in January that competes with Runway's Gen 2.0 model, but it's definitely working on releasing a more capable version for Sora. I wouldn't be surprised if Anthropic, as well as tech giant Meta and well-funded startup Inflection, debut models that fit into Google's long context window of his 1.5 Pro in the coming weeks.
For those of us watching from the shore, so to speak, this is all as thrilling as it was for 19th century newspaper readers who followed the Great Tee Race. However, it seems a little dangerous. And unlike the Great Tee Race, the risks are not just to those participating in the race, but to all of us.
Giving ChatGPT more memory makes it more convenient for users, but it also increases the risk that this model will leak users' personal information, as has already happened once with previous versions of chatbots. Sola's surreal videos could lead to even more convincing deepfakes. (Currently, Sora is available only to “red teams,” select individuals and companies specifically hired by OpenAI to test models for safety and security vulnerabilities.) (It is not clear whether the video was generated by an AI.) Ethicists note that OpenAI did not add any kind of visual watermark to the video it used for the demonstration, clearly indicating that it was generated by an AI. criticized. They also accused the company of revealing little about how Sora was trained, leading many to wonder if copyrighted material was perhaps used without the owner's consent. was doubtful. In the future, systems like Sora could also put many people in Hollywood out of work.
There are even bigger risks. This set of model enhancements will increasingly accelerate our transition to super-powerful AI software that could be dangerous in the wrong hands or pose a danger to humanity on its own. It means that there is. To be sure, there's no evidence that tech companies are paying much attention to safety as they rush to roll out model after model.
OpenAI claimed that by learning through video footage, Sora gained an intuitive understanding of physics and common sense reasoning that models trained in other ways lacked. In making this claim, the company describes the model as a key step toward its official goal of creating artificial general intelligence, a single AI system that can perform all the economically valuable cognitive tasks that humans can. I tried to position it. Except that many, including Elon Musk, were quick to point out that Sora's understanding of physics was questionable. (Even the OpenAI researchers didn't seem to fully understand that Sora can't flap or fly like a bird, as seen in one of the videos they published.) (I highlighted the example.) They also seemed to have a hard time depicting certain aspects of the chair. He correctly understands the natural world, such as the number of legs an ant has.
So the AGI framing is just hype and is actually a way for OpenAI researchers to justify working on a project whose sole purpose is to show the world that they can beat Google and Runway at the video generation game. maybe. But while building Sora as a step toward his AGI, OpenAI researchers spent a lot of time detailing the tests they've done and the precautions they've taken so far to make their new model safe. It's disturbing to see you're not spending. They said they are red-teaming this model at the moment and have not disclosed anything about when they plan to release it. But simply revealing its existence will likely prompt other companies like Runway and Google to accelerate efforts on competing products. And in this race, like the captain of a clipper ship, caution may prevail over speed.
In the Great Tea Race, the captains knew their destination and had a well-established route to get there. The AI Great Race is different. In a sense, it combines elements of his 19th century competition with elements of an even earlier era of sailing. From the 15th century to his 18th century was the Age of Exploration, when captains sailed beyond the horizon to unknown worlds. They were competing with each other even then. The wealth and fame of the entire kingdom rode with them. What they discover on this journey will change the world. But they will also bring disease, death, conflict, and conquest to those they encounter.
We have to hope that the AI Great Race ends like the Great Tea Race, where only the ships themselves are at risk. Because of the tea race, the core technology, the clipper ship itself, was short-lived. Even in 1866, a ship equipped with a steam engine in addition to sails transported all the clippers back to London, which took her 15 days. Three years later, the Suez Canal opened, further reducing navigation times. Within a decade, steamships had largely supplanted clippers in world trade. Today's transformer-based neural network AI models also know, perhaps without seeing millions of examples during training, that chairs don't normally fly like birds and that ants have six legs. It may be similarly overtaken by other AI technologies that can be grasped.
There is also another lesson learned from the tea people. In the years after 1866, China had a bumper crop of tea. Prices have fallen dramatically and there is little premium for being first to market. This can also happen with AI. Because the freely available open source model will quickly match the capabilities of today's top proprietary software, and businesses no longer feel the need to pay big bucks to access generative AI capabilities.
In the meantime, watching the races can be both exciting and a little scary. But we should all be a little skeptical about the motives and whether the risks are ultimately worth the reward.
More AI news below. But before that, if you want to learn about the latest developments in AI and how it impacts your business, join leading figures from business, government, and academia at . luck'appointment of London edition our AI brainstorming meeting. In London it will be from April 15th to 16th. You can apply to participate here.
I'd also like to highlight a great recent interview with Fortune CEO Alan Murray. luck's next leadership Podcast with Wasem Khaled, co-founder and CEO of Blackbird AI. You can check it here.
jeremy kern
jeremy.kahn@fortune.com
@jeremyakahn
This article originally appeared on Fortune.com
