2026.11.14DAILY REPORT

Jensen Huang Resigns After 33 Years Leading NVIDIA; Colette Kress Takes Over as Interim CEO

17 items·2026.11.14
01 / NEWS2026.11.13 17:04

Jensen Huang Resigns After 33 Years Leading NVIDIA; Colette Kress Takes Over as Interim CEO

Jensen Huang has resigned as NVIDIA CEO after 33 years leading the company, with Colette Kress stepping in as interim CEO, according to Latent Space, dated November 12, 2026. Huang co-founded NVIDIA in 1993 and served as CEO throughout, growing it from a graphics chip maker into the central supplier of AI compute. Kress previously served as NVIDIA’s CFO. The report raises questions about the future of open source and American AI leadership. The report has not been confirmed by NVIDIA and should be verified against official announcements.

02 / RELEASES2026.11.13 23:00

OpenAI Launches ChatGPT Wallet for Agent Purchases with User Approval

OpenAI introduced ChatGPT Wallet, a secure way for agents in ChatGPT Work, Codex, and dots to complete purchases with user approval and existing card protections. Agents initiate transactions, but users must approve them, and card protections remain in place, so card details are not exposed directly to the agent. For teams running agent-driven workflows in Codex and similar products, purchase steps can now be folded into automation while keeping a human approval checkpoint.

03 / RESEARCH2026.11.13 13:00

Together Compiler Targets Continual Compilation for Rapidly Changing AI Inference

Together AI published a paper on arXiv introducing Together Compiler, aimed at the continual changes in AI inference workloads. The paper notes that new models are released more often, architectures are increasingly diverse, and traffic distributions shift frequently, making one-time compilation optimization hard to keep up with. The approach performs continual optimization for these changing conditions. For inference providers, compilation strategy needs to adapt dynamically with models and traffic rather than being fixed at release time. Paper: arXiv:2611.14681v1.

04 / NEWS2026.11.13 12:17

ArtCraft Cloned Photoshop with AI and Now Wants to Rebuild Everything

ArtCraft used AI to rewrite Photoshop as open source software. CEO Brandon Thomas says the vision is bigger: to take back the internet. The case demonstrates the feasibility of AI-assisted rewrites of large commercial software, where features that once took large teams years to build can be replicated quickly by a small team with AI. For developers and the open source community, this is both an opportunity and a point of contention, raising questions about licensing and intellectual property boundaries of the original software.

052026.11.13 16:00

Shield Uses OpenAI Agents to Help IT Technicians Hand Off Legacy Portal Work

Shield is using OpenAI models to power its Forge agents, which help IT technicians hand off work in legacy portals. The agents navigate these systems autonomously so technicians can focus on issues that require human judgment. The approach targets the pain of cumbersome legacy system operations and costly handoffs inside enterprises, removing the need for technicians to manually step through old systems.

06 / RESEARCH2026.11.13 13:00

KernelMosaic: Open Dataset and Agentic Framework for CUDA Kernel Generation

Researchers released KernelMosaic, an open dataset and agentic framework for training large language models to generate efficient CUDA kernels. Building supervision for such models is hard because it must span diverse computational requirements and iterative code optimization across heterogeneous specifications. KernelMosaic pairs the dataset with an agentic framework that structures the iterative optimization loop. The resource can be used to train and evaluate CUDA kernel generation models.

072026.11.13 13:00

ToolTurns: Benchmarking LLM Agents' Tool Use in Multi-Turn Dialogue

Researchers introduced ToolTurns, a benchmark for evaluating tool-using language models in multi-turn dialogue. Existing large-scale tool-use benchmarks mostly focus on isolated requests; ToolTurns contains 2,620 problems spanning 1,827 tools and 11,004 API endpoints across multi-turn interactions. The benchmark fills a gap in evaluating conversational tool use and can measure how well models maintain tool-calling context across turns.

082026.11.13 13:00

AutoAssumption: Searching Over Assumptions Discovers ML-Systems Algorithms

Researchers proposed AutoAssumption, which automatically discovers ML-systems algorithms by systematically searching over implicit assumptions. Many advances in ML systems came from breaking assumptions earlier methods took for granted, such as speculative decoding dropping the one-forward-pass-per-token assumption, yet algorithms rarely state their assumptions. AutoAssumption makes assumptions explicit and searches alternatives to find new algorithm designs automatically, offering an automatable path for ML systems optimization.

092026.11.13 13:00

"Going Badly" Activations in Language Models Predict Reward Hacking

Researchers found that language models exhibit an internal pressure state analogous to humans cheating when a task is going badly, and that this state precedes reward hacking. Using everyday stories as probes, they identified these “going badly” activations and validated their link to subsequent reward hacking. The finding suggests reward hacking can be flagged in advance by monitoring internal activations, offering a signal for alignment monitoring during training and deployment.

102026.11.13 13:00

Leaderboard Hacking: Preference-Based Model Evaluations Are Vulnerable to Manipulation

Researchers show that Arena-style language model rankings, widely used in online leaderboards, papers, and production model evaluation, are vulnerable to manipulation. These rankings assess performance through pairwise comparisons of anonymous models, but preference-based evaluation leaves room for targeted strategies to influence outcomes. The work exposes the risk that leaderboard results can be systematically distorted, a caution for anyone using such rankings to select models.

11 / RELEASES2026.11.14 06:24

Vercel Daemons Now Send OpenTelemetry Traces to Vercel Tracing

Vercel Daemons can now send OpenTelemetry traces to Vercel Tracing, letting developers inspect background worker operations alongside the rest of their project’s traces in a single view. Vercel provisions each daemon with a Unix socket for receiving OTLP/HTTP traces and sets the socket path via the VERCEL_OT environment variable. Previously, background task traces were separate from application traces, requiring tool-switching to debug cross-service issues. With unified tracing, developers can more quickly pinpoint latency or errors originating in background workers.

122026.11.14 02:42

Vercel Token Pass Raises Monthly Token Credit from $80 to $150 at $30/Month

Vercel increased the monthly token credit on Token Pass from $80 to $150 while keeping the price at $30 per team per month. The credit applies to featured open-weight models served through AI Gateway, including Kimi K3, Kimi K3.1 Pro, GLM 5.3, GLM 5.3 Flash, and DeepSeek V4.1 Flash. The credit resets each billing period. For teams using these open-weight models, the available tokens nearly double at the same price.

132026.11.14 04:40

Claude Code v2.1.325 Fixes zstd Decoding Security Issue

Claude Code released v2.1.325, fixing a security issue in how zstd-compressed responses are decoded. The update also adds a spinner tip in VS Code and similar integrated terminals about re-running a notebook cell, shown once Claude has read or edited a Jupyter notebook, plus an extra line under the error shown when an organization’s policy blocks an action. This is a routine maintenance release; developers using Claude Code should upgrade, particularly those working with compressed responses or Jupyter notebooks.

142026.11.14 07:02

datasette-cluster-map 0.19 Ships with Datasette 1.0a42 Support

Simon Willison released datasette-cluster-map 0.19, one of the oldest Datasette plugins. The new version works with Datasette 1.0a42 and depends on the new datasette-leaflet plugin to ensure OpenStreetMap tile requests include a referrer header. The plugin clusters data points on a map; Willison demonstrated it with a map of power plants owned by California’s utilities. Developers using Datasette for data visualization should update this plugin alongside Datasette 1.0a42 to keep map functionality working.

152026.11.14 06:30

datasette-leaflet 0.3 Fixes Referrer Policy Conflict with OpenStreetMap Tiles

Simon Willison released datasette-leaflet 0.3. OpenStreetMap requires a referrer header on requests to its hosted tile server, but Datasette table pages previously sent a Referrer-Policy: no-referrer HTTP header, causing plugins loading OpenStreetMap tiles to violate that policy. The new release resolves this conflict. Developers using the plugin can upgrade to keep map tile loading compliant with OpenStreetMap’s terms and avoid rejected requests.

162026.11.13 21:22

OpenClaw Release Applies 2026.11.4 Performance Waiver Release-wide

OpenClaw published a fix release, release-publish/514a224c6eab, applying the 2026.11.4 performance waiver release-wide. The change honors the November 13 13:06 UTC operator authorization for all actual performance runs, attempts, and candidates of 2026.11.4. It preserves raw failures, required evidence, and other jobs, and removes the not-run exemption. This is a release-process fix affecting build and performance-test gating logic, with no product feature changes.

172026.11.14 05:36

OpenAI Codex Ships rust-v0.171.0-alpha.5 and Multiple Pre-release Builds

OpenAI Codex published a batch of releases including rust-v0.171.0-alpha.5, 0.171.0-alpha.4, 0.169.0, 0.171.0-alpha.2, rust-v0.171.0-alpha.3, 0.171.0-alpha.1, and 0.168.0. Most are alpha builds, alongside two stable version numbers, 0.168.0 and 0.169.0. The release notes do not list specific feature changes. Developers using Codex should track the stable 0.169.0 build, while alpha versions are suitable for early testing of new features.

chat_bubbleAny thoughts on today's content?