The year 2024 has been huge for the artificial intelligence (AI) industry, witnessing numerous smaller, cheaper models coming out to challenge the domination of the established AI behemoths – and with surprising success, as a new study shows.
Indeed, the AI arena seems to be pretty competitive and evened out, with no particular firms standing much above the others in terms of performance, according to the Artificial Intelligence Index Report 2025 released by the Institute for Human Centered AI at Stanford University in California on April 9.
Specifically, the smaller generative AI models not only seem to be holding out just fine against the big-shots (themselves getting bigger) but have risen to become their equals, coming up rather close to each other on the study’s Chatbot Arena leaderboard.
Why are minimalist models outperforming?
This is because the development of better algorithms has allowed sleeker, more compact models such as DeepSeek to match the performance achievable by a model 100 times larger two years ago, with 2024 being a breakthrough year for these smaller AI models.
Furthermore, China is now developing models that can stand neck and neck with their US competition when it comes to performance, an improvement from 2023 when the leading Chinese models lagged behind the top American model by close to 20% on the Massive Multitask Language Understanding (MMLU) test. By the end of 2024, this difference had dropped to 0.3%.
And despite developers continuing to stick to the ‘the bigger the better’ maxim, companies are also releasing highly capable small models. For instance, the smallest one registered a score above 60% on the MMLU in 2022 using 540 billion parameters, whereas by 2024, a model achieved the same score with only 3.8 billion parameters.
It should also be noted that minimalist models train faster, provide quicker answers, and use less energy than their larger counterparts. At the same time, some of them can emulate the behavior of the larger ones, or utilize better algorithms and hardware than those in older systems, all of which is giving them an edge nowadays.
Meanwhile, one of the best proofs of AI’s growing capabilities is Google DeepMind’s Dreamer AI system, which figured out on its own how to play Minecraft. Elsewhere, the ongoing question of whether AI can be creative seems to be a subjective one, as observing AI models at work seems to influence how we perceive its creativity – and for the better.
ReadNext
