For all of AI’s promise, most companies using it are not yet delivering true value—to their customers or themselves. With investors keen to finally see some ROI on their AI investments, it’s time to stop generalizing and start thinking smaller.
Instead of building epic models that aim to accomplish all feats, businesses looking to cash in on the AI gold rush should consider pivoting towards focused models that are designed for specific tasks. By attacking a singular problem with a fresh solution, innovators can create powerful, novel models that require fewer parameters, less data, and less compute power.
With billions upon billions of dollars being spent on AI engineering, chips, training, and data centers, a smaller form of AI can also allow the industry to progress more safely, sustainably, and efficiently. Furthermore, it is possible to deliver this potential in various manners— through services atop commodity generalist models, retrieval-augmented systems, low-rank adaptation, fine-tuning, and more.
Some tech enthusiasts may cringe at the word “small,” but when it comes to AI, small does not mean insignificant, and bigger is not necessarily better. Models like OpenAI’s GPT-4, Google’s Gemini, Mistral AI’s Mistral, Meta’s Llama 3, or Anthropic’s Claude cost a fortune to build, and when we look at how they perform, it’s not clear why most businesses would want to get into that game to begin with.
Even as big players monopolize the field, their sexy, headline-making generalized foundational models seem to perform well enough on certain benchmarks, but whether this performance generalizes to actual value in terms of increased productivity or similar remains unclear.
This is not an argument for green AI but for bringing some realism back into the AI hype cycle. Even if the model itself is a large proprietary one, the tighter the focus, the smaller and more manageable the number of possible outputs to consider becomes. With less token length, models optimized for a specific task can run faster and be highly robust and more performant, all while using less data.
Even Big Tech players are working to focus their AI offerings with smaller, more powerful models.
With OpenAI projecting big returns when it releases PhD-level ChatGPT agents, the ideal is that one day, we will all have our own agents—or AI assistants—that use our personal data to act on our behalf without prompts. It’s an ambitious future, notwithstanding the privacy and security concerns.
While the jump from where we are now to where we could be going seems to be a huge one, building it piece by piece is a clear, lower-risk approach than assuming a massive monolith is the answer.
AI innovators who home in on specificity can build a growing, nimble team of expert models that increasingly augment our work instead of one costly, mediocre assistant who is fat with parameters, eats massive data sets, and still doesn’t get it right.
By creating lighter computing infrastructures that focus on the right data, businesses can fully maximize AI’s potential for breakthrough results even as they cut down the immense financial and environmental costs of the technology.
Amid all the hype around AI and the behemoth Big Tech models fighting for headlines, the long arc of innovation has always relied on incremental, practical progress. With data at the heart of the models that are indeed changing our world, small, focused AI promises faster, more sustainable, and cost-effective solutions—and in turn, offers both investors and users some much-needed ROI from AI.
The opinions expressed in Fortune.com commentary pieces are solely the views of their authors and do not necessarily reflect the opinions and beliefs of Fortune.
