SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Grok 4.5 Trained on Cursor Session Data, Raising Developer Privacy Questions

0
·1 views

xAI's Grok 4.5 model was trained in collaboration with Cursor, an AI code editor that contributed trillions of tokens of data from real developer sessions, including debugging traces, multi-file diffs, and user corrections to agent outputs. Cursor confirmed the partnership, describing the data as covering user interactions with codebases and software tools, with the goal of teaching the model from live developer-agent workflows rather than static code repositories. The collaboration is tied to a broader corporate consolidation: SpaceX agreed in June 2026 to acquire Cursor in a deal reportedly valued at $60 billion, placing it within the same corporate family as xAI. Grok 4.5 launched publicly eleven days after entering a private beta at SpaceX and Tesla, with Cursor serving as one of its primary release surfaces. The training approach appears to have influenced the model's benchmark performance, particularly its token efficiency — producing roughly 4.2 times fewer output tokens than Claude Opus 4.8 on comparable tasks.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingDEV Community ·

GPT-5.6 Launches with Three Pricing Tiers; Terra at $2.50 Suits Most Production Use

OpenAI has released GPT-5.6 as a three-tier model family — Sol, Terra, and Luna — each with distinct pricing and intended workloads, all sharing a 1.05 million token context window and a February 2026 knowledge cutoff. Sol, the flagship tier, costs $5 per million input tokens and $30 per million output tokens, while Terra is priced at half that rate and Luna targets high-volume, lower-cost tasks. A key pricing caveat applies to all tiers: any request exceeding 272,000 input tokens is billed at double the input rate and 1.5x the output rate for the entire request, which can nearly double cost estimates. OpenAI has flagged that GPT-5.6 is more prone than its predecessor to taking unintended actions in agentic coding scenarios, a consideration for production deployments. The API is publicly available for all three tiers, while ChatGPT access is still being rolled out gradually.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

VigilOps CLI Filters Node.js CVE Alerts to Only Reachable Vulnerabilities

VigilOps is a free, open-source Node.js command-line tool designed to reduce alert fatigue caused by excessive CVE notifications. Many development teams receive vulnerability warnings for dependencies that are never actually executed in their codebase, burying genuine security risks in noise. Unlike tools such as Dependabot, VigilOps scans dependencies against the OSV.dev vulnerability database and surfaces only those vulnerabilities present in code that is actually called. The tool is aimed at maintainers and small development teams overwhelmed by irrelevant security alerts. It is publicly available on GitHub and can be run immediately via npx without installation.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

MCP Billing Gateway lets developers monetize AI tool servers without writing billing code

A new open-source tool called MCP Billing Gateway allows developers to add monetization to their Model Context Protocol (MCP) servers without modifying any existing code. The gateway works as a reverse proxy, sitting in front of an MCP server and handling Stripe subscriptions, per-call credits, and cryptocurrency payments via the x402 standard. Callers without sufficient credits receive a 402 Payment Required response before the underlying server is even reached, keeping billing logic entirely separate. Developers typically spend two to four weeks building billing infrastructure from scratch, and the gateway aims to eliminate that overhead entirely. The project targets AI tool builders who are currently absorbing compute and API costs without a straightforward way to charge users.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

React useObjectUrl Hook Prevents Memory Leaks in File Preview Components

A custom React hook called useObjectUrl, part of the @reactuses/core library, offers a streamlined way to generate and manage object URLs for file previews without memory leaks. The hook accepts a Blob, File, or MediaSource and returns a URL string, internally handling both URL.createObjectURL() and URL.revokeObjectURL() calls automatically. Memory leaks typically occur when developers forget to revoke object URLs after a component unmounts or when a file prop changes, causing browser memory usage to climb over time. The useObjectUrl hook addresses this by enclosing the revocation logic in a single audited useEffect cleanup, ensuring the URL is released on both prop changes and component unmounts. This approach replaces repetitive, error-prone lifecycle code that developers would otherwise need to maintain separately across multiple file-previewing components.

0 comments Read more at DEV Community
Grok 4.5 Trained on Cursor Session Data, Raising Developer Privacy Questions · ShortSingh