David Cox spends his days helping to oversee IBM’sAI models and his nights exploring new ones, using generative AI to brainstorm ideas, write code and test research concepts at a pace he says was impossible just a year ago.
Generative AI arrived with promises of helping workers draft emails, summarize documents and automate routine tasks. For Cox, who helps leadIBM Research’s large language model (LLM) development efforts and directs the MIT-IBM Computing Research Lab, the technology has become something closer to a research collaborator. He uses it to explore new neural network designs, digest scientific papers, generate visualizations and communicate ideas that might otherwise remain trapped in notebooks or conversations.
“I’ve become painfully productive—obsessively, worryingly productive,” Cox told IBM Think in an interview.
The remark points to a broader shift unfolding across the AI industry. While businesses continue to experiment with chatbots and productivity tools, Cox says the latest generation of models is beginning toreshape how scientific and technical work gets done.
Cox said he uses LLMs throughout the day, whether he’s sketching out new neural network designs, mapping research projects, writing code or wrestling with difficult mathematical problems. He now typically has around 10 lines of inquiry underway at any given time, turning to the technology as a sounding board to quickly test ideas and push them toward experiments.
From ideas to work
The observation carries particular weight because Cox sits unusually close to the technology itself. A former Harvard University professor with a doctorate in computational neuroscience from MIT, he now leads the IBM Research organization responsible for developing the company’s language and code models.
The change did not happen overnight. Although Cox has worked with generative AI since the technology emerged, he said something shifted this year.
“We crossed some threshold around January, where it stopped being something I was keeping tabs on to track its progress, and much more something that enabled me to do things that I couldn’t do before,” Cox said.
For most of his career, turning an idea into a working experiment required finding the right people and persuading them to spend time on it. Academic researchers mostly rely on graduate students and postdoctoral fellows to implement concepts and run experiments, while industry researchers work through teams, balancing competing priorities.
Those constraints have loosened dramatically. Rather than waiting for collaborators to become available, Cox can sketch out an idea, brainstorm with an AI model, build a research plan and begin testing it almost immediately.
A faster path to discovery
Much of Cox’s day job involves thinking about the future of AI systems. As the executive responsible for IBM’s foundation model strategy, he oversees teams developing the LLMs that underpin many of the company’s AI offerings.
Modern AI systems rely on neural networks, mathematical structures that learn patterns from data. Most LLMs use a design known as a transformer architecture, the breakthrough approach behind tools such as ChatGPT and Claude.
Researchers constantly search for ways to improve those architectures. Testing a new idea traditionally required substantial engineering work before meaningful results could emerge.
That process now unfolds much faster.
“A researcher early in their career has a lot of reasons to play it safe and chase incremental tweaks on existing methods,” Cox said. “I want to go after things that are radical departures from the status quo.”
Work often begins as a conversation. An idea gets refined through discussion with an AI system, expanded into a research plan and translated into code. Experiments follow. Results arrive quickly. New questions emerge.
Along the way, Cox built AI-driven software that automatically submits jobs to IBM computing clusters whenever spare processing capacity becomes available. Tasks that once would have required dedicated engineering support can now happen almost immediately, on a whim.
The abundance of available help has altered the rhythm of research itself.
“If I can’t sleep for whatever reason, I can just advance one of my pet projects,” he said.
Reading, learning and asking questions
The multidisciplinary background that carried Cox from neuroscience into AI shapes his approach to research. New ideas often emerge at the intersection of fields, making the ability to absorb unfamiliar material particularly valuable.
When a paper catches his attention, Cox still reviews it himself. At the same time, he asks AI systems to analyze the work and explain how it relates to his own projects already under way. Because the systems retain context about ongoing research, they can often immediately connect new information to existing efforts.
“It knows everything I’m working on, so it locks onto why I am interested in something almost immediately,” Cox said.
The same capability has changed how he approaches unfamiliar subjects. Advanced mathematics plays an increasingly important role in AI research, and Cox frequently ventures into sub-fields where he is less fluent.
“I can ask the dumbest questions,” Cox said. Unlike a conversation with a human expert, the exchange carries no fear of embarrassment. The ability to ask basic questions, he said, makes it easier to move into unfamiliar territory and build a working understanding of complex topics.
Some of the most useful applications have emerged in unexpected places.
Visual thinking plays a central role in how Cox develops ideas. Communicating those ideas has often proved more difficult than creating them.
“I sometimes have a picture that is so clear in my head, but which my chicken-scratchings whiteboard struggle to convey,” Cox said.
Instead, he works with AI to explain what he is seeing, and the system can generate elaborate visualizations that make abstract concepts easier to explore and communicate. Those images often expose flaws, assumptions and gaps that were harder to see when ideas existed only in his head.
Human judgment remains essential
For all the productivity gains, Cox doesn’t view AI as a replacement for expertise.
“This isn’t a substitute for knowing what you’re trying to do and being creative,” Cox told IBM Think.
Hallucinations, the industry term for confident but incorrect outputs, remain a persistent problem. Models still generate convincing answers that turn out to be wrong. Researchers must continually verify information, challenge assumptions and correct mistakes. AI systems also make the same kinds of reasoning errors that humans do or miss the point of what the user is trying to do.
An AI system may produce pages of analysis or thousands of lines of code, but users still need enough understanding to recognize when something has gone off track.
“It’s a huge amplifier for people who know their stuff well enough to correct the AI, but I don’t think you’ll get very far if you didn’t understand the content,” Cox said.
People who lack expertise can easily mistake activity for progress, he warned. Researchers who understand a field can identify mistakes and steer the technology back in the right direction. Researchers who don’t invariably end up following flawed ideas into dead ends.
The technology also falls short of generating genuinely original insights, in his view. What it excels at is helping people explore, refine and execute ideas much more quickly than before.
“At least as of today, AI isn’t injecting a lot of genuinely new ideas into the mix,” Cox said. “However, it can take your creativity and unlock it, bringing your ideas to life—or invalidating them—at a breakneck pace. That is already very exciting.”
The latest AI trends, brought to you by experts
Get curated insights on the most important—and intriguing—AI news. Subscribe to our weekly Think newsletter. See the IBM Privacy Statement.
How to choose the right foundation model
Learn how to choose the right approach in preparing datasets and employing foundation models.
