Elon Musk xAI restart brings two Cursor veterans to helm coding tool

xAI scraps its first coding assistant and begins again under fresh leadership

Sarah Mills | GlobalBeat

The Elon Musk xAI restart became public late Tuesday when the company told early testers it was “archiving the current codebase” and forwarding invite codes to a new repository, according to three developers who received the email.

xAI launched nine months ago with a pledge to “build AI tools that accelerate human imagination.” Yet competitors—from Cognition’s Devin to Anthropic’s Claude-3 Sonnet—have raced ahead in software-writing accuracy while Grok, Musk’s flagship model, still hallucinates runnable code 28 % of the time on internal benchmarks. The coding tool, provisionally nicknamed “Grok-Coder,” was meant to close that gap before year-end.

A second false start inside xAI’s cramped Palo Alto loft

Employees say the first attempt never recovered from a rushed December release. Engineers stitched together an open-source LLM wrapper over a single weekend to coincide with Musk’s public denial that Tesla was poaching AI staff. “It was demo-driven development,” one staffer said. “The goal was a screenshot, not a product.”

By March, feedback from 400 volunteer coders showed 62 % abandoned the plug-in after it produced buggy Python and refused common Git commands. Managers froze new feature work in April. The formal decision to reboot came during a 7 a.m. video call on 2 May, when safety director Lakshmi Bhai warned the tool “did not meet the minimum reliability bar” for even experimental branding.

Cursor duo arrive with pre-built audience and revenue proof

Michael Chu and Diana Okojie, co-founders of the $8-a-month editor Cursor, will jointly run product. Their startup boasted $4.3 million annual recurring revenue before its acquisition by xAI in an all-stock deal disclosed last week. Okojie’s departure memo to Cursor users, seen by GlobalBeat, promised “seamless migration” for paying customers once the xAI version ships.

Chu brings a reputation for speed: Cursor autocompletes multi-file changes in under 900 milliseconds, a metric verified by Stanford’s CodeX dataset. Engineers familiar with both companies say xAI paid roughly 1.2 times Cursor’s last valuation, implying a price tag near $60 million for a 14-person team.

Rebuild targets “verified execution” before autocomplete

Okojie told staff the new architecture will run generated code inside sandboxed containers and return only snippets that compile, borrow a method from Replit’s Ghostwriter. Early prototypes sped up pull-request reviews at payments start-up Helio by 34 %, according to beta documents. Chu’s internal note states: “We are not shipping suggestions, we are shipping diffs that pass CI.”

Pervasive tests will run against the open-source repository of LangChain, one of the most forked Python libraries, giving xAI a public scoreboard to measure hallucinations and regressions in real time.

Cash burn worries investors funding the Elon Musk xAI restart

xAI raised $6 billion in Series-B funding in May on the premise that Grok would reach 200 million users by December. The coding tool was billed as the first vertical that would keep developers locked into the ecosystem. Two investors told GlobalBeat they were reassured by Chu’s revenue track record but alarmed by the sunk cost of the discarded codebase, which consumed an estimated 27,000 A100-GPU hours valued at roughly $1.8 million in cloud credits.

Musk himself downplayed the expense on X last night: “Iteration speed > sunk cost. Always.”

Competitive field narrows as startups stumble

While the Elon Musk xAI restart unfolds, rivals face their own headwinds. Cognition’s Devin demo wowed investors in March, yet the wait-list remains closed and the company has postponed a promised open API twice. Stability AI released StableCode in April only to see downloads flatline after a licensing dispute with contributor community Hugging Face. Meanwhile Stack Overflow’s traffic dropped 15 % year-on-year as coders paste AI answers directly, creating a vacuum for a trusted, accurate generator.

The numbers tell a different story than the hype. Independent evaluation platform Evident measured just four AI coding assistants that complete more than half of competitive programming problems correctly; none belongs to a social-media conglomerate.

Freelance coder in Lagos faces weeks of uncertainty

Temitope Awolowo, a 24-year-old Android freelancer in Lagos who pays Cursor monthly, worries his subscription could lapse before migration. He earns roughly $1,200 building utility apps and relies on AI to meet 48-hour client deadlines. “If the new xAI tool costs more or drops features, I’ll have to raise my rates,” he said over WhatsApp. “But if it really tests code first, I might finish twice as many gigs.”

EU AI Act looms over rapid release schedule

Brussels begins enforcing tiered obligations for general-purpose models in August, including disclosure of training-data sources and systemic-risk evaluations. xAI’s director of compliance, Henrik Stenson, has already requested a six-month implementation grace period, but EU officials tell GlobalBeat the commission will not exempt Silicon Valley firms that rewrite core models. Canada and Singapore are drafting similar rules, meaning any aggressive rollout schedule could collide with mandatory audits before year-end.

Grok-Coder beta expected within 45 days

The Cursor team plans to open a private beta through the existing Grok interface “no later than late July,” according to an internal roadmap. Users who joined the first wait-list will keep their queue positions, but must re-consent to fresh data policies. A public launch hangs on achieving 90 % unit-test pass rate across 1,000 GitHub repositories selected by volunteer maintainers. If the metric holds, Musk wants to bundle the assistant with the next wide release of X Premium+ before the U.S. Thanksgiving shopping window.