What 2024 Has in Store for AI -II

"Trough of Disillusionment" Soon to Come

The CEO and co-founder of Humane Intelligence, a nonprofit that tests AI, Rumman Chowdhury, believes that the euphoria surrounding AI may have peaked in 2023. She projects that in 2024, "we will hit the trough of disillusionment." "We're going to come to the realization that this technology isn't as revolutionary as we've been led to believe."

According to her, the most advanced AI models are "immense computational achievements," and the next generation is probably going to be even more advanced. However, because no one has "figured out what large language models are useful for," the hoopla will fizzle out.


Nevertheless, Chowdhury claims that the allure of AI has been advantageous in that it has prompted lawmakers to take action, which has direct implications for technologists. "In a manner never seen before, everyone will pay attention to what's happening in politics." For instance, she notes that even though the E.U. AI Act, the EU's complete AI law, won't take effect until 2026, businesses must begin planning now. "In policy land, two years is going to fly by very, very quickly."

Constructing Larger Models

Google DeepMind unveiled Gemini Ultra, its most recent AI model, in December 2023. The amount of "compute," or processing power, that Google DeepMind needed to train the model was not disclosed, but Epoch, an AI forecasting company, calculates that 90 septillion floating-point operations were used in its training—probably more than any other model that has been created to yet.


With around one-fourth as much compute used to train its epoch estimations, OpenAI's GPT-4 performed marginally worse than Gemini Ultra. Scholars have charted the correlations between the computational resources allocated to train a model and its efficacy in doing specific tasks, such as word prediction. Their findings indicate that exponential growth in training compute translates into linear improvements in model performance.


In 2024, he predicts, companies will introduce larger models that will once again improve little by little. For example, chatbots will make fewer mistakes. However, "it's not going to be like, 'Wow, this is like a totally new game-changing capability,' from the user's perspective."

Sam Altman, the CEO of OpenAI, anticipates something similar. In an October podcast interview, he said that the public's response to new AI models would be comparable to that of new iPhone releases.