OpenAI Releases GPT-6.1 Terra, Focused on Visual Reasoning and Computer UseOpenAI Shares Model Results on Kakeya Conjecture and Other Math ProblemsOpenAI Maps ChatGPT Work Usage Across 165 Activities in Nearly 100 CountriesGreptile Extends Bug Detection Beyond Code Review with GPT-6 AstraCodeRabbit Uses GPT-6.1 Sol to Investigate Security Risks in Complex CodebasesFanatics Uses OpenAI's Shop the Look, Sees 10% Higher EngagementSPECMEM Speeds Up Agent Tool Calling with Live Memory ManagementWhen Does Muon Precondition? Structured Gradient Noise Shapes Its BenefitsRL Critic Already Knows LLM Confidence, No Extra Training NeededSelf-Directed Repair Turns Agent Mistakes Into Lessons That Persist Across SessionsBenchmarks Are Surprisingly Compaction-Friendly, Study FindsRSI-Opt: LLM Agents Recursively Self-Improve Optimizer Settings During TrainingScale AI's AgentEnv Evals Now Run on Vercel Sandboxwebframe Tested as a Scraper and Sandbox for Running Web Pages Like ProgramsVercel Observability Adds Custom Monitors for Query-Based AlertsVercel Agent Validates Alerts Before They Reach Paging SystemsVercel Sandbox Adds stdin Support for Running CommandsVercel Connect Adds CLI Search Across 1,000+ ServicesReplit on the Self-Driving Company: AI Drives Coordination Costs Toward ZeroUsers Turn to ChatGPT for Tech Support as Big Tech Abandons Customer ServiceOpenAI Releases GPT-6.1 Terra, Focused on Visual Reasoning and Computer Use

OpenAI published model results on the Kakeya conjecture, the Exponential Time Hypothesis, and quantum chromodynamics, along with guidance on how researchers can verify and build on them. The work is part of OpenAI’s effort to engage academic communities and emphasizes reproducibility over headline claims. For mathematicians and physicists, the release offers a way to use models as exploration tools while requiring independent verification of each step.

OpenAI published Signals data covering 165 work activities across nearly 100 countries, showing how people use ChatGPT for work and how that usage changes over time. The data is granular down to specific work activities rather than broad industry categories, making it possible to see which tasks AI enters first. For companies and researchers, it offers first-hand evidence of actual usage distribution rather than survey-based intent.

Greptile is extending bug detection from code review into testing actual user workflows. According to OpenAI, Greptile uses GPT-6 Astra to test user workflows, Daybreak to strengthen security, and Codex to reduce rework. The division of labor is clear: one model simulates and validates real usage paths, one handles security review, and one cuts the cost of fixes. The shift signals code quality tooling moving from static checks toward validating runtime behavior.

CodeRabbit uses GPT-6.1 Sol to investigate security risks across complex codebases, with its security agent helping developers review findings and proposed fixes. The move extends code review tools from style checks toward locating and explaining security vulnerabilities, giving developers not just alerts but candidate fixes to evaluate. Specific false-positive and false-negative rates were not disclosed in the announcement.

Fanatics is using OpenAI to power Shop the Look, a feature that helps fans complete their game-day outfits. Testing across roughly 30,000 MLB products showed 10% higher engagement. The feature turns product browsing into outfit recommendations, so fans can assemble a full look without picking items one by one. For e-commerce platforms, it’s a concrete example of wiring LLMs directly into product recommendation, measured by engagement rather than conversion.

Tool-calling agents are usually optimized for accuracy, but in deployment their real cost is latency: every tool call is decoded autoregressively. The paper introduces SPECMEM, which applies speculative decoding to speed this up without changing outputs, and retrieves past calls to improve draft quality. The approach puts memory inside the decoding step rather than only in context assembly. For teams running high-frequency tool-calling agents, latency is a key metric affecting both user experience and cost.

Muon has become a popular optimizer for large language model training, yet its theoretical foundations remain poorly understood. This work studies how the geometry of gradient noise shapes Muon’s preconditioning benefits, analyzing the question both theoretically and empirically. The results point to noise structure: preconditioning pays off when gradient noise has particular structure. For teams training large models, this offers a basis for judging when Muon helps and when gains are limited. The paper is an arXiv preprint.

Reliable confidence estimation is essential for trustworthy LLM deployment, but dedicated confidence training requires extra data and compute. This paper asks what prediction problem confidence actually corresponds to, and finds that the critic network from reinforcement learning already carries a usable confidence signal, eliminating the need for a separate confidence training stage. If it holds up, teams can reuse the critic already produced during RL training, skipping an extra round of data and compute while still attaching usable confidence scores to outputs.

Today’s agents recover from mistakes only temporarily: in-context reflections are easily forgotten across sessions and compete for limited context space. This work proposes self-directed repair, which converts agent mistakes into experience that persists, so agents avoid repeating the same errors in later tasks. For long-running agent systems, it addresses the problem of falling into the same trap repeatedly. The paper is an arXiv preprint.

Long-horizon agentic tasks are pushing LLM agents past their context windows, forcing agent scaffolds to free up space through context compaction. Compaction is generally lossy by design, and agents can silently drop critical information. This paper examines how benchmarks behave under compaction and finds they are surprisingly tolerant of it. The result suggests existing evaluations may understate compaction’s real risks, and that benchmark scores leave a blind spot when judging long-task agent reliability.

Conventional neural network training follows a predefined optimization recipe. RSI-Opt proposes an LLM agent that monitors training dynamics and adjusts optimizer settings as training progresses. The authors note that a fixed agent policy cannot accumulate experience, so they add a recursive self-improvement mechanism that lets the agent learn from its own adjustment history. The goal is to reduce manual tuning and make training adapt to observed dynamics. The work is an arXiv preprint and has not been peer-reviewed.

AgentEnv, Scale AI’s framework for testing agents against tasks and scoring results, now runs on Vercel Sandbox. The Vercel provider ships in AgentEnv 0.9.1334 and later, running the agent and its task environment in separate Sandboxes. That isolation brings evaluations closer to real deployments and keeps tasks from interfering with each other. Developers can run an evaluation on Vercel using the official example.

webframe is a small project by Jeff Lindsay that lets you “run a web page like a program”: point it at a URL or file, and it executes in a hidden headless web view, piping console.log() to standard output along with any exceptions. Simon Willison evaluated it both as a scraper and as a sandbox. For developers extracting structured data from dynamic pages, executing the page and capturing output is more direct than parsing HTML, though the sandbox isolation needs careful assessment.

Vercel Observability now lets users create custom monitors that trigger alerts based on query results. Trackable signals include AI Gateway costs, Core Web Vitals and other performance metrics, Workflow and external API failure rates, and Vercel Sandbox CPU time. Each monitor tracks a single signal, giving developers a way to get notified about cost or performance anomalies before users report them.

Vercel Agent can now investigate alerts from Vercel monitors before they reach a paging system. With validation enabled for a webhook destination, only alerts the agent classifies as Critical are sent through. Users enable it by creating or linking a webhook destination in a monitor’s notification settings. For on-call engineers, this filters out noisy alerts and reduces being paged at night for non-critical issues.

Vercel Sandbox now supports sending input to running commands via standard input (stdin). Developers can use it to exchange messages with stdio-based MCP servers such as agent-browser, or pipe data into command-line programs while they run. In the TypeScript SDK, enable it by setting stdin: true and detached: true in runCommand(). This lets interactive AI agent tooling run directly inside Sandbox without workarounds or modifications to existing programs.

Vercel has added a discovery feature to Vercel Connect, letting developers and their agents browse and search more than 1,000 services without leaving the terminal. The vercel connect discover command searches by name, alias, keyword and description, with results filterable by category and connection type. The change means developers configuring integrations no longer need to jump to a browser for docs, and agents can locate available services directly from the CLI.

A Replit blog post starts from Ronald Coase’s 1937 question of why firms exist: organizing activity inside a company is often cheaper than coordinating through markets, but firms carry their own coordination costs in management, meetings and more. The post introduces the idea of the self-driving company, in which AI takes on much of that coordination, compressing those costs and changing where company boundaries fall. For founders and teams, the practical question is how much business a small team can run with AI doing the coordination.

Big tech platforms have effectively eliminated human customer support, making it impossible for users to reach a real person. As a result, many people with ordinary technical problems are turning to ChatGPT for help. But the chatbot often gives unreliable or misleading answers to routine tech support questions. Jason Koebler highlights how AI chatbots have become the de facto first — and often only — line of support, despite lacking the ability to resolve actual account or device issues.