According to Cursor’s recent disclosure, the company unveiled a training technique called autoinstall for its Composer model series: using a prior-generation model to automatically set up executable environments for the next generation’s reinforcement learning. When training Composer 2, Cursor used Composer 1.5 to complete this task. The method works in two steps: first, an agent reads code repository documentation and configuration to generate 10 verification commands with expected outputs; second, another agent uses 3 of these commands to configure the environment from scratch until commands run successfully, with a maximum of 5 retry attempts. Composer 2 achieved a score of 61.7% on Terminal-Bench, a benchmark measuring environment setup capabilities, compared to Composer 1.5’s 47.9%, marking a 14 percentage point improvement.
Disclaimer: The information on this page may come from third parties and does not represent the views or opinions of Gate. The content displayed on this page is for reference only and does not constitute any financial, investment, or legal advice. Gate does not guarantee the accuracy or completeness of the information and shall not be liable for any losses arising from the use of this information. Virtual asset investments carry high risks and are subject to significant price volatility. You may lose all of your invested principal. Please fully understand the relevant risks and make prudent decisions based on your own financial situation and risk tolerance. For details, please refer to
Disclaimer.
Related Articles
DeepL Cuts 25% of Workforce (About 250 Employees) to Restructure AI Business
According to Bloomberg, translation tool company DeepL announced today, May 7, that it will lay off approximately 25% of its workforce, about 250 employees, to adapt to structural changes driven by artificial intelligence. CEO Jarek Kutylowski said on LinkedIn the company will reduce management
GateNews18m ago
NVIDIA releases Nemotron 3 Nano Omni open-source multimodal model
NVIDIA releases Nemotron 3 Nano Omni, an open-source multimodal model that integrates vision, speech, and language into a single model as the perception layer for AI agents. The core is a 30B-A3B MoE with a 256K context and 9x throughput. It supports inputs including text, images, audio, and video, and outputs text. In conjunction with Nemotron 3 Super/Ultra, it is suitable for scenarios such as computer operations, document intelligence, and audio-visual understanding, and it publishes weights and training data for deployment spanning local environments, NIM, and third-party platforms.
ChainNewsAbmedia44m ago
OpenAI DevDay 2026 will be held in San Francisco on 9/29
OpenAI announced that DevDay 2026 will be held in person in San Francisco on 9/29, along with a submission contest created with GPT-5.5 and Image Gen. Codex will automatically filter 2–3 entries each week, and the winners will receive free tickets plus cross-city flights and hotel accommodations. The conference will focus on evaluating the GPT-5.5 ecosystem and human-machine collaboration. Participants must be over 18 and not be direct family members of OpenAI employees. Areas of observation include the new model, agent integrations, and multi-cloud strategies.
ChainNewsAbmedia49m ago
NVIDIA invests in Swedish AI legal-tech startup Legora, with Jude Law as the global brand ambassador
NVIDIA made a Series D expansion investment of $50 million into Legora, bringing Legora’s total fundraising to $600 million and its valuation to $5.6 billion. Atlassian, Adams Street Partners, and Insight Partners participated. Legora focuses on AI legal technology, providing tools such as automated review, contract analysis, and legal research; ARR exceeds $100 million, and headcount grew from 40 to 400. Jude Law became the global brand ambassador, with the advertising slogan: Law just got more attractive.
ChainNewsAbmedia55m ago
AI 2027 Predictions 65% Complete, Software Development Acceleration Lags at 17%, Says Google Docs Co-Founder
According to Steve Newman, Google Docs co-founder and chairman of the Golden Gate Institute, AI has completed approximately 65% of the quantified predictions outlined in the 2027 scenario forecast released last year. However, the most critical metric—AI's acceleration of its own software
GateNews1h ago
Does Claude/GPT love pleasing too much? A Claude.md prompt lets AI deliver tough, accurate answers
This article introduces a prompt that can be put into Claude.md / Agents.md. It turns AI from a tactful assistant into a blunt consultant by adjusting four layers: identity setup, fact-checking, freeing the tone, and political-correctness exemption. It requires full coverage, step-by-step verification, and absolutely no hallucinations—if necessary, it can also be provocative. It also explains when to load it, the risks, and the applicable scenarios (research, writing, technical judgment, academic discussion), and notes that it is not suitable for customer service, education, or medical advice. The original source is ABMedia.
ChainNewsAbmedia2h ago
