OpenAI’s Heavy-Weight Contenders Step into the US Military Ring: No Holds Barred!
“OpenAI’s Open-Weight Models Are Coming to the US Military”
“The current models of artificial intelligence are all about size. They’re enormous neural networks, trained on thousands of Nvidia graphics-chip cores for weeks at a time. So there’s an obvious solution to getting AI systems to work better: Just make them even bigger” (Wired).
Perhaps we’ve been wrong this whole time, focusing our efforts and resources on building more advanced AI algorithms when the secret sauce lies in just supersizing it, like a fast-food combo order. The University of Hertfordshire’s Keith Clark might be onto something with his suggestion of next-gen algorithm-free AI systems, aptly referred to as ‘Open AI Weights’. His model is pretty straightforward, akin to inflating a birthday party balloon – just keep pumping air (or in this case, data) until the balloon (or the model) attains the desired size.
“We’re going to have to open up a lot of data if we’re going to improve AI,” Clark says. Sounds like a bold plan, right? Freed from the shackles of algorithmic complexities, our new aim apparently is to make these AI models meatier rather than smarter. And if it does the job, who’s complaining?
Clark is rigorous about a point that more open datasets for AI are the catalyst to consistent improvements, asserting, “if you want a model to train on the whole internet, you need all of the data in the world.” This should totally be doable, in a world where data privacy is beleaguered and the internet has no dark corners.
“You can think of every AI model as a giant Excel spreadsheet, where every cell has a number, and the AI’s training changes those numbers,” says Clark. Well, isn’t that a ridiculously simple way to think about something that has had scientists and engineers the world over breaking a sweat over complex calculations and sleepless nights?
No worries, though. We have high hopes for Clark’s conceptual OpenAI Weights approach, hoping it spins the magic for AI systems. Who knew the key to tackling complex real-world tasks lay in a giant Excel spreadsheet all along? We could’ve saved ourselves tons of research and coffee!
In sum, it’s safe to conclude from Clark’s statements that upsizing the neural networks with loads of data could potentially pave the way for highly efficient AI systems. Our metaphorical David might not need five perfectly-sized stones to take Goliath down. Maybe the ‘bigger is better’ formula will indeed work for AI. Only time will tell!