Antonio Scandurra

Antonio Scandurra is a co-founder of Zed whose AI work includes polishing its early Assistant Panel and, with Nathan Sobo, explaining the engineering decisions behind Zeta, Zed’s open edit-prediction model. His earlier work includes Atom, Teletype’s collaborative editing architecture, and distributed database development in Rust. He uses the GitHub identity as-cii, which links to the social handle as__cii.[2][3][4][5][6]

In June 2023, Scandurra was polishing Zed’s forthcoming Assistant Panel, an OpenAI integration that brought a ChatGPT-like conversation into the editor’s multi-buffer interface. Each message could be edited or deleted, making the conversation itself editable rather than a fixed sequence of exchanges. In his June 9 development report, he said the team was using it internally and anticipated opening it to users the following week.[5]

Scandurra and Sobo’s February 2025 Zeta account addressed a different problem: predicting changes at arbitrary locations, rather than only completing text at the cursor. Zeta, derived from Qwen2.5-Coder-7B, took recent edits and the cursor position as context and rewrote a surrounding snippet to incorporate predicted changes. The approach allowed edits in multiple locations, but required generating more tokens than ordinary fill-in-the-middle completion.[6]

Their team evaluated predictions by asking Claude to check plain-English assertions instead of requiring an exact token match. For a quicksort example, the assertion asked whether the function recursed on both sides of the pivot. This accommodated different valid implementations. Initial experiments prompting Qwen2.5-Coder-32B became inconsistent as the evaluation suite grew: prompt changes that fixed new cases broke earlier ones, and the larger model did not fit the team’s latency requirements.[6]

The team moved to supervised fine-tuning with Unsloth and LoRA. About 50 Claude-generated synthetic examples seeded an initial fine-tune; internal use of a version behind a feature flag then helped build a dataset of around 400 examples. Unrelated deletions and insertions still occurred when the model worked on small excerpts from larger files. They followed with direct preference optimization, using positive and negative examples to teach which edits to avoid. Scandurra and Sobo reported that roughly 150 carefully selected examples improved these difficult cases.[6]

To accelerate snippet rewriting, the team exploited the resemblance between input and output through speculative decoding, using n-gram search in the input to identify starting points for parallel token generation. They set prediction targets below 200 milliseconds at the median and below 500 milliseconds at the 90th percentile. Their account credits Baseten’s performance engineers with optimizing serving to achieve those targets; these were reported project results, rather than independent benchmarks.[6]

Scandurra and Sobo presented Zeta as an open-source model with an open dataset and endorsed extending Zed’s open approach to AI: “we see no reason to change that approach when it comes to working with AI.” They also said submitted training data would be reviewed before publication for safety and privacy.[6]

The editor infrastructure behind this AI work predates it. Scandurra’s collaboration with Sobo began on Atom and included architecting Teletype and researching foundations that later became Zed.[2][7]

Scandurra joined the Atom team in 2014 while still at university, after his open-source contributions attracted its attention. Zed’s later hiring account specifically credits his consistent, high-quality pull requests. He subsequently worked with Sobo on Teletype’s architecture. Zed’s biography also credits him with developing a real-time, distributed, conflict-free database in Rust for Ditto.[2][4]

In a February 2024 edited interview, Scandurra recalled trying iframes and Canvas APIs to speed Atom’s line layout. Measuring text and locating the cursor between characters proved awkward with the available browser tools. He described the experience as “bending over backwards to try to achieve something that in principle should have been simple.” He also recalled a substantial period of reading CRDT papers with Sobo, followed by two or three rewrites of their implementation.[7]

Scandurra described Rust as a “sweet spot” for building on abstractions while retaining performance, though he criticized slow compilation. Thorsten Ball reported a concrete pairing session in which Scandurra changed a find-all-matches operation from finding results individually to finding them together; that particular release-build test fell from about one second to four milliseconds. Scandurra also acknowledged the onboarding cost of owning the editor’s stack: newcomers needed to learn a custom codebase, although its authors could explain it and the team retained control over its behavior.[7]

His view of engineering quality was selective: “The more core something is, the more it deserves thought and quality.” He argued that GPUI and SumTree warranted particularly careful work because other code depended on them, including randomized testing of SumTree’s edge cases. Peripheral improvements could receive less polishing. When randomized tests exposed failures, he favored treating them as information—even if they required a rewrite—rather than a reason to panic.[7]

A February 2024 graphics investigation, coauthored with Sobo, shows Scandurra working through performance problems on users’ hardware. Theo Browne’s M2 MacBook exposed scrolling hitches the founders had not seen on their M1 machines. Relaxing a Metal command-buffer wait reduced blocking, and Scandurra built and AirDropped a test binary at Browne’s dining table. The change then exposed a race: the CPU could overwrite memory while the GPU was still reading it. The team introduced pooled instance buffers, returning each buffer only after its frame completed.[8]

Further work on ProMotion and CADisplayLink initially appeared successful, but users reported 60 FPS caps and occasional freezes. In the article’s corrective addendum, Ball credited Scandurra with proposing CVDisplayLink. That switch alone did not stabilize rendering: they also changed command-buffer presentation and removed synchronization during ordinary drawing while retaining triple buffering. Scandurra then built special handling for startup and window resizing, where synchronization was still needed. Ball reported stable 120 FPS on his and Scandurra’s machines after the complete correction, which shipped in Zed 0.121.7 and Preview 0.122.2.[8]