OpenAI has increased the default context window for GPT-6.1 Sol in Codex and Work from 272,000 to 500,000, with no long-context charges when used within subscription allowances; Codex's GitHub code review is also now generally available. NVIDIA has opened up Vera Rubin rack and compute tray designs, with MGX 1.2 supporting racks up to 250 kW, and released four open-source DSX components. Anthropic is reportedly seeking a $2 trillion IPO valuation, while New Constructs values it at just $150 billion, with the dispute centered on operating losses and open-source competition. QGIS is discussing a blanket ban on AI-assisted contributions because the volume of contributions exceeds maintainers' review capacity, but concerns that the ban could also restrict AI-assisted reviews and security audits have drawn opposition.
Codex is expanding its review capabilities, while QGIS considers tighter restrictions on contributions.
OpenAI has increased the default context window for GPT-6.1 Sol in Codex and Work from 272,000 to 500,000, an increase of about 80%. The change is intended to reduce context compaction and help tasks retain substantial context and detail over longer periods. There are no long-context charges when using the model within subscription allowances in these two products.
Read original →On October 12, Codex demonstrated its code review feature, with GitHub code review now generally available. It brings pull requests, review feedback, and private conversations about code into one place, allowing users to find reviews across repositories, understand changes and risks, apply team review guidelines, and draft comments without switching tools. GitLab support remains in preview.
Read original →QGIS's QEP428 proposal (PR #387) would replace its previous human-in-the-loop AI policy with a strict ban and is currently under discussion. Supporters argue that the volume of AI-assisted contributions exceeds maintainers' review capacity, creating PR backlogs or insufficient reviews and diverting effort from accepting human contributions. Some participants object to the ban's overly broad wording, concerned that it could also prohibit using LLMs to assist with code reviews, security audits, and understanding code.
Read original →🔍 Analysis: A 500,000-token context window gives Codex room to retain more complete task details, while GitHub code review brings those capabilities into teams' review workflows. The debate over QGIS's QEP428 reminds developers that tools being able to handle more code does not mean maintainers can accept more contributions. The tension is that AI can both add to the review workload and help work through it. For open-source projects, a ban that also covers AI-assisted reviews and security audits would restrict both incoming contributions and the tools that could ease the review burden.
The open Vera Rubin designs cover power and cooling interfaces and extend to cluster management software.
At the OCP Global Summit, NVIDIA contributed Vera Rubin rack and compute tray designs to OCP through its MGX open architecture. MGX 1.2 supports rack deployments up to 250 kW, adding blind-mate liquid cooling, 5,000 A liquid-cooled busbars, and standardized power and cooling interfaces so partners can reuse designs and build interoperable platforms. NVIDIA also released four open-source DSX components covering network configuration, cluster validation, rack management, and bare-metal deployment.
Read original →The gap between a $2 trillion target and a $150 billion valuation reflects disagreement over whether the business model is viable.
Anthropic plans to list on Nasdaq and is reportedly seeking a $2 trillion valuation; research firm New Constructs values it at just $150 billion. The firm argues that widening operating losses and competition from open-source models leave Anthropic without a viable business model, and says the IPO will test investors' gullibility. However, its assessment is based on reported financial data, and it has previously misjudged DoorDash's IPO performance.
Read original →Muse is experimenting with layered task-status displays, while Karpathy and Aaron Levie discuss the gap in users' understanding.
Meta design lead Alex Cornell described Muse's three layers of status display: expressions for everyday users, task lists for details, and raw logs for engineers to inspect. The same activity expands on demand, avoiding the need to confront everyone with complex information; the main chat retains context, while side chats handle separate tasks. Its generative interface keeps a fixed structure, dynamically updating only status expressions, personalized suggestions, and small animations rather than rebuilding the interface each time.
Read original →Karpathy argues that the gap in people's understanding of large language model capabilities is widening; roughly one to two billion people are only occasional users of free ChatGPT-style products. Box CEO Aaron Levie says this shows that AI adoption is still in its early stages, with most people using chatbots only for simple interactions resembling next-generation search. He argues that while these uses are valuable, they fall far short of demonstrating AI's full capabilities.
Read original →🔍 Analysis: The roughly one to two billion occasional users Karpathy describes still encounter AI mainly through chat. Muse's expressions, task lists, and raw logs offer a path from a simple entry point to progressively deeper insight into task execution. The point of this design is not to show everyone more information, but to give everyday users and engineers explanations of the same activity at different levels of detail. For product teams, making task status understandable is an interface hurdle to overcome in introducing users to capabilities beyond search-like question answering.
MakerMaker distinguishes AI-assisted R&D from systems taking responsibility for improving themselves.
On October 12, MakerMaker.AI co-founders Dhaval Adjodah and Owen He discussed recursive self-improvement (RSI) in an a16z interview. They argue that using AI to automate parts of AI development is not the same as true RSI. Their goal is for the system being improved to also be responsible for improving itself, creating a closed loop that covers models and runtime frameworks and eventually extends to more components of the AI technology stack.
Read original →