How AI is changing engineering work

IBM researcher Kaoutar El Maghraoui uses a rotating cast of AI assistants to work through research papers, compiler stacks, codebases and debugging sessions.

“There’s an old proverb: ‘If you want to go fast, go alone. If you want to go far, go together,’” El Maghraoui, a Principal Research Scientist at IBM’s AI Hardware Center, told IBM Think in an interview. “AI has collapsed that tradeoff for me. I move faster than I did working alone, and I’m covering more ground than I could without a collaborator.”

El Maghraoui is not alone. Companies are using generative AI across offices, call centers and customer support systems, and engineers are using the technology in a different way: to learn faster. Tools like Claude, ChatGPT, Gemini, and IBM’s Bob AI coding partner help researchers work through unfamiliar subjects, understand technical papers, debug software and explore problems that previously could take months to study on their own.

Her work sits at the center of that ecosystem. At IBM, El Maghraoui focuses on AI hardware and software co-design tied to the company’s Spyre AI accelerator chip, work that requires constant movement between machine learning systems, hardware architecture, runtime infrastructure, compiler technology and software integration.

Much of that work now happens alongside AI systems.

Among the newest additions to her workflow is IBM’s new AI development partner Bob. Unlike consumer AI systems trained broadly on internet data, Bob operates inside IBM’s development environment and understands repository structures, internal systems and enterprise workflows.

“What draws me to Bob, [and] Bob Shell in particular, is that it suits the way I already work,” she said. “Terminal-native, close to the code, without the overhead of switching context to an IDE plugin.”

She said the difference becomes especially noticeable inside large enterprise codebases, where context matters as much as raw code generation.

“Having an AI assistant that understands the internal environment rather than treating every codebase as a blank slate is a meaningfully different proposition,” she said.

At the same time, El Maghraoui hops between many public AI tools, each assigned to different kinds of work. Claude became her preferred system for coding and deep technical reasoning. ChatGPT helps her survey unfamiliar technical areas quickly. Gemini becomes useful when she needs to work with extremely large context windows or load entire document sets into memory.

“The way I think about it is simple,” she said. “When I need to work with code, I go to Claude first.”

Keeping pace with modern AI research, she suggested, has become increasingly difficult as the volume of papers and tooling accelerates. New benchmarks, frameworks, optimization methods and architectural claims now arrive continuously. The larger challenge, she said, involves deciding what deserves attention.

“I use Claude, Gemini and ChatGPT to rapidly go through deep technical papers,” she said. “What might take a full afternoon of careful reading becomes a 30-minute conversation where I can actively probe the ideas.”

“It turns reading from a passive activity into an active one,” she said. 

At the same time, she described the systems less as authoritative experts than as imperfect collaborators, whose weaknesses become obvious to experienced engineers quickly. “ChatGPT can be confidently wrong on complex reasoning,” she said. “It has breadth but that sometimes comes at the cost of depth, and you need domain knowledge to catch it.”

Much of El Maghraoui’s research focuses on AI accelerators, the specialized hardware systems designed to run machine learning workloads more efficiently than traditional CPUs. GPUs, TPUs, ASICs and newer accelerator architectures all make different tradeoffs involving memory, scheduling, computation and software integration.

One major focus involves “dataflow” architectures, including IBM’s Spyre accelerator. Traditional processors execute instructions sequentially. Dataflow systems instead trigger computation when data becomes available, allowing neural network tensors to move dynamically through hardware pathways.

That architecture can reduce inefficiencies associated with instruction scheduling and control overhead, but it also adds more complexity to software systems that map neural network graphs onto hardware resources.

“AI tools have been helpful in reasoning through these design trade-offs and in exploring research on what a well-integrated dataflow backend should look like,” she said.

As she dug deeper into the systems, she kept running into areas she had never formally studied. One of the biggest was compiler infrastructure, the low-level software that translates code into instructions computers can execute.

“The compiler piece was a real gap for me,” she said. “I’m not a compiler engineer by training, and I’m still learning.”

Modern AI frameworks like PyTorch contain increasingly sophisticated compiler and optimization layers. Systems like Dynamo and Inductor help transform machine learning graphs into optimized execution pathways tailored to different hardware backends.

Understanding those systems requires knowledge spanning machine learning, runtime behavior, optimization theory, systems engineering and low-level execution pathways.

AI systems have changed how El Maghraoui approached that learning process. Instead of moving slowly through documentation and research papers alone, she could work interactively through concepts until they became intuitive.

“I can now follow conversations with compiler specialists, navigate the stack and apply the concepts to real problems,” she said. “That shift happened through months of working through things with Claude.”

The same process helped her learn how PyTorch, a widely used AI development framework, works across different types of hardware, including NVIDIA and AMD graphics chips, Intel accelerators and Apple silicon.

“Piecing that together from documentation alone would have taken months,” she said.

Elsewhere in her workflow, AI has become a useful debugging partner. Fixing software problems often means sorting through error logs, crash reports and unfamiliar code to figure out what went wrong.

For El Maghraoui, AI systems can significantly shorten that early search process. “I can paste in a failing test, a confusing stack trace or a section of PyTorch internals and get a structured hypothesis about what’s happening and where to look next,” she said. “That initial triage cuts the time I spend staring at logs dramatically.”

Outside IBM, she teaches High Performance Machine Learning at Columbia University, where she faces another challenge tied directly to AI’s rapid pace: educational materials age almost immediately.

“What was current when I designed a module can be outdated by the time students take the exam,” she said.

AI systems now help her review course materials, identify emerging developments and determine which concepts are worth introducing in class.

“The challenge isn’t just keeping up,” she said. “It’s figuring out what actually matters: which signals are important, which are noise.”

She said durable conceptual foundations matter more than chasing every new tool release. “The goal is to teach in a way that gives students a solid conceptual foundation,” she said. “Not just what’s new today, but the mental models that will still be useful when today’s tools are already obsolete.”

For all the gains AI systems provide, she repeatedly returned to their limitations. In highly specialized or proprietary environments, models can still generate explanations that sound convincing but collapse under scrutiny.

“For Spyre-specific nuances,” she said, “the model doesn’t have that context and can produce plausible-sounding but wrong reasoning.”

That problem becomes even sharper at the frontier of research itself, where systems encounter sparse documentation, evolving architectures and experimental behavior that no model fully understands.

“I treat AI outputs as first drafts and hypotheses, not conclusions,” she said.

In El Maghraoui’s view, AI systems amplify expertise rather than replace it. “AI gets you most of the way,” she said. “The last mile requires hands-on judgment you can’t offload.”