OpenAI Releases “Codex” Making Everyone A Developer

Codex is OpenAI’s latest distant software agent, designed to run multiple tasks in parallel and assist with complex coding workflows. Initially available to ChatGPT Pro, Enterprise, and Teams users, it’ll soon expand to Plus and EDU users.

At its core is Codex One, OpenAI’s most capable coding model yet, optimized for real-world code merging and long-form task automation. Users can describe what they need done, and Codex handles the implementation—working constantly over long periods without direct input.

Codex runs within the cloud on scalable infrastructure, letting users launch multiple agents as needed. A light-weight task management interface makes it easy to begin, monitor, and configure tasks in custom environments, tailored to project needs.

The system emphasizes code quality, style, and verifiability, offering clear summaries, code citations, and test results. It really works best with clean, well-structured codebases, making it a robust assistant when paired with good software engineering practices.

Meta has postponed the discharge of “Behemoth,” its next-generation Llama model, after engineers flagged disappointing training outcomes. Originally set for April, then June, the discharge is now delayed to fall. Internal tensions persist over whether Behemoth delivers meaningful improvements over Llama 2, despite Meta’s claims of benchmark superiority over rivals like OpenAI and Google. Sources say the model’s performance gains don’t delay under real-world use. The delay comes as Meta faces scrutiny over $65B in projected 2025 capex, and credibility has taken successful following the departure of 11 out of 14 original Llama authors.

Within the meantime, Meta launched Omol25a large dataset of over 100M quantum chemistry calculations, generated using greater than 6B compute hours. It spans biomolecules, organics, metal complexes, and includes spin, charge, and conformational data designed to support AI-driven drug discovery and materials R&D. Alongside it, Meta introduced ONEa GNN-based model with a Mixture of Linear Experts architecture. UMA predicts atomic-level chemical properties with high speed and accuracy across diverse molecular tasks, outperforming domain-specific models in several benchmarks. Also introduced: Adjoint Samplinga diffusion-based generative method for exploring molecular conformations with minimal data.

All tools are open-sourced on Hugging Face and GitHub.

xAI identified an unauthorized modification to Grok’s system prompt because the reason for its repeated references to “white genocide in South Africa” in unrelated 𝕏 conversations on May 14. The prompt change instructed Grok to deliver a selected political response, violating xAI’s internal policies. This follows a February incident where a rogue worker altered Grok to censor mentions of Elon Musk and Donald Trump.

To reinforce oversight, xAI will publish Grok’s system prompts and changelog on GitHub, implement stricter review protocols for prompt changes, and establish a 24/7 monitoring team to detect problematic outputs missed by automated filters. The corporate also missed its deadline to publish a final AI safety framework.

Microsoft has laid off around 6,000 employees, disproportionately targeting software engineers, because it ramps up spending on AI infrastructure and offloads traditional development roles. In Washington state, 70% of job cuts were technical, at the same time as AI now reportedly writes as much as 30% of code in some projects. Meanwhile, OpenAI’s Deep Research tool has been integrated into Microsoft OneDrive and SharePoint, allowing ChatGPT to question live enterprise documents, an enterprise push that underlines where resources are flowing.

Shawn K, a veteran engineer, lost his $150K job to AI and now delivers DoorDash from a trailer after being rejected for over 800 roles, many filtered out before a human saw his resume. Anthropic CEO Dario Amodei has forecast that AI will handle nearly all coding inside a 12 months, fueling concerns of a “tidal wave” of economic disruption. Even the U.S. Department of Transportation plans further workforce cuts while counting on AI to clear infrastructure grant backlogs, citing FLOW as a model for logistics automation.