CMU

Carnegie Mellon University (CMU) is a Pittsburgh-rooted university whose AI work spans robotics, model deployment and agent engineering. Its Robotics Institute launched in 1979; in 2018, its School of Computer Science announced a dedicated undergraduate AI degree. More recent CMU-associated research includes machine learning compilation for local inference and coordination methods for software engineering and computer-use agents.[6][8][7][1][4][3]

In August 2023, Tianqi Chen was a CMU assistant professor teaching a machine learning compilation course and running the MLC group. The group’s projects included MLC Chat, an iPhone app for on-device language models; Web LLM, for browser inference; and MLC LLM, a framework for deployment across hardware and software stacks. LLVM-backed targets extended deployment to ARM and x86 CPUs. Latent Space reported MLC Chat speeds of up to 30 tokens per second at the time.[1]

These projects applied compilation technology from TVM. TVM takes model programs represented as computational graphs or loops, combines operators through fusion, plans memory use and generates low-level code for GPU and non-GPU backends. Fusion can combine operations into one GPU kernel rather than launching separate kernels. Memory planning allocates storage ahead of execution, while symbolic shape relationships let the compiler reason about tensors whose dimensions vary. TVM can use existing optimized libraries or generate code for platforms where such libraries are unavailable.[1]

Chen described TVM Unity in 2023 as a second iteration intended to make new models and their optimizations easier to incorporate. MLC separated model development and optimization in Python from deployment into application runtimes such as JavaScript, Python, C++ and Java. His stated goal was “making it deploy universally”. He credited collaborators and earlier compiler work, including Halide.[1]

A browser-deployment experiment followed a native MacBook implementation driven by engineer Jin Lu. After that implementation worked with Metal acceleration, the team tried its existing WebGPU backend. WebGPU gave browser code access to native GPUs, allowing the compiled model to invoke GPU kernels through the browser. Chen described application integration through an npm package and an OpenAI-compatible REST API, while noting that users still had to download model weights before local use.[1]

CMU’s faculty connections also extend into open-source agent development. December 2024 coverage identified Graham Neubig with both CMU and AllHands and discussed OpenHands, the open-source coding agent framework. Those affiliations connect academic and startup work; AllHands is a separate startup.[2]

Centralized Asynchronous Isolated Delegation (CAID), by Jiayi Geng and Graham Neubig, addresses interference between agents working on interdependent software tasks. A central manager constructs dependency-aware plans and delegates subtasks to engineers working asynchronously in isolated git worktrees. Structured integration and executable tests bring their work together. The implementation acknowledges OpenHands’ agent framework, PaperBench and Commit0’s benchmarks, Fujitsu grants, and feedback from Apurva Gandhi, Lintang Sutawika, Emmy Liu and Howard Chen.[4][10][11]

In the July 8, 2026 revision, the CAID authors reported accuracy improvements over single-agent baselines of 25.6% absolute on PaperBench paper-reproduction tasks and 14.7% on Commit0 Python-library development tasks. These are results for the specified benchmark tasks.[10]

Multi-Agent Computer Use (MACU), submitted June 1, 2026 by Jing Yu Koh, Ruslan Salakhutdinov and Daniel Fried, extends coordination to desktop and web tasks. The work, attributed to CMU/LTI in Latent Space’s June coverage, uses a manager to represent dependencies and goals as a directed acyclic graph. It dispatches parallel subagents on ready nodes, then adds, cancels or rewrites nodes as findings arrive. The manager also carries information forward when downstream agents may be unable to observe it again. The authors reported approximately 1.5× faster average task completion on the Odysseys long-horizon web-navigation benchmark.[3][12]

These recent deployment and agent projects sit within a university whose technical education and robotics institutions developed much earlier. CMU’s origins predate its present name by more than six decades.[6][8][1][10][12]

Andrew Carnegie donated $1 million to Pittsburgh in 1900 to establish practical education for working-class men and women. The Carnegie Technical Schools admitted their first class in 1906. In 1912, the institution became the Carnegie Institute of Technology, or Carnegie Tech, and began offering bachelor’s degrees. Carnegie Tech merged in 1967 with the Mellon Institute, a science research center founded by Pittsburgh’s Mellon family, to become Carnegie Mellon University.[6]

The Robotics Institute grew out of discussions in fall 1978 and spring 1979 involving President Dick Cyert, deans Dan Berg and Angel Jordan, Allen Newell and Raj Reddy. Participants estimated that critical mass would require 5–10 faculty and $2–5 million in annual funding. Cyert and Reddy’s Pentagon visit led to an Office of Naval Research commitment. A five-year, $5 million commitment from Westinghouse president Tom Murrin, together with Jordan’s identification of engineering faculty, preceded the institute’s October 1979 launch.[8]

On May 10, 2018, CMU announced an undergraduate AI degree beginning that fall, calling it the first offered by a U.S. university. Directed by robotics and computer science research professor Reid Simmons, the program focused on making decisions from complex inputs such as vision, language and large databases. It combined computer science and mathematics foundations with statistics, probability, computational modeling, machine learning and symbolic computation, plus ethics, social responsibility and opportunities to apply AI to transportation, health care or education.[7]

Students would enter the AI major in their second year after shared first-year foundations. Initial planned capacity was no more than 100 students across the second, third and fourth years, or approximately 30–35 entrants annually. Instruction drew on six School of Computer Science units, including the Machine Learning Department, Language Technologies Institute and Robotics Institute, with participation from Dietrich, Heinz and Engineering. The same announcement described a CMU AI research and education initiative begun the previous year involving almost 200 faculty members.[7]

CMU’s 2025 Annual Financial Report reported an institution-wide endowment market value of $3.5 billion as of June 30, 2025. The report also described a CMU–NVIDIA joint research center for robotics, autonomy and AI within a Pittsburgh AI technology community involving the University of Pittsburgh. CMU researchers were expected to work with an embedded NVIDIA solutions architect and gain early access to technologies supporting their research.[9]