LodeHQ
AI news for people who shipFree, by email

AI · Saturday, 14 November 2026 · Issue 170 · 12 stories

Jensen Huang resigns as NVIDIA CEO, Kress takes over

Jensen Huang has resigned immediately as NVIDIA CEO and board member, leaving CFO Colette Kress to run the company while its board searches for a permanent successor. NVIDIA says its operations, customer commitments, and product roadmap will continue unchanged.

Models & Releases

Atlas turns photos into editable 3D worlds for real workflows

World Labs’ Atlas lets filmmakers, designers, and robotics teams turn photos or generated images into explorable 3D scenes, then use those worlds in games, simulations, Blender, and other tools. Viewfinder reconstructs a scene so you can move the camera through it and render camera paths, while Chisel lets you reshape the layout and see each viewpoint update. Atlas can also convert splats, scene captures made of many points, into meshes with real geometry.

worldlabs.ai ↗1 min read at source

Research

Integer-only inference keeps the same LLM run identical everywhere

DetLLM replaces floating-point operations with integer arithmetic so the same prompt can produce bit-identical output across CPUs, GPUs, vendors, and batch sizes. Floating-point sums change when hardware kernels add values in a different order, which can alter the winning next token: one cited test got 80 answers from 1,000 temperature-zero runs. In the author’s demo, fp16 runs diverge by token 457 across nine configurations, while the integer runs stay aligned.

nathan.rs ↗12 min read at source

Products & Industry

NVIDIA puts CFO Colette Kress in charge after Huang resigns

Jensen Huang has resigned immediately as NVIDIA’s CEO and board member, putting CFO Colette Kress in charge while the company searches for a permanent successor. Kress will keep the finance role, and the board will consider internal and external candidates. Harvey C. Jones becomes the board’s first chair, while NVIDIA says operations, customer commitments, and the product roadmap will continue without interruption.

nvidianews.nvidia.com ↗2 min read at source

Claude agents can now run entirely inside your infrastructure

Anthropic’s private beta lets developers run Claude Managed Agents inside their own cloud or on-premise environment, keeping the agent service, credentials, and session history under their control. Teams operate the server, database, and workers, while choosing whether model requests go to Anthropic, a cloud provider, or their own gateway. The same APIs support agents working on internal code and systems, including fixing builds, testing releases, and triaging alerts.

claude.com ↗3 min read at source

ChatGPT Wallet lets agents pay without exposing card details

ChatGPT Wallet lets agents complete purchases while keeping a user’s actual card details hidden from both the agent and the business, initially for US ChatGPT Pro subscribers through dots. The user adds a card once, reviews the items and total, then approves each purchase with Face ID or a fingerprint; the payment network issues a one-time substitute token. Eligible card rewards and dispute protections remain in place.

openai.com ↗4 min read at source

Policy & Safety

Export bans may not stop China training on Nvidia’s top GPUs

U.S. export controls may be failing at their core goal: keeping Chinese AI labs from training on Nvidia’s most powerful chips. Unemployment Labs analyzed model weights, the learned structure left by training, and found fingerprints of Blackwell GPUs in open models from DeepSeek, Moonshot AI, MiniMax, and Ant Group. The evidence does not prove every model was trained on banned chips or that laws were broken, since offshore data centers can provide access, but Epoch AI estimates 275,000 Nvidia accelerators entered Malaysia in 2024 and 2025.

techcrunch.com ↗4 min read at source

Six Chinese AI models show fingerprints of restricted Blackwell chips

Six models from DeepSeek, Moonshot AI, MiniMax, and Ant Group contain numerical fingerprints matching NVIDIA Blackwell hardware, suggesting these labs have accessed chips restricted from export to China. The analysis reconstructs traces from publicly released model weights and reproduces them on Blackwell, including physical-chip confirmation for DeepSeek-V4-Flash and Kimi K3. It does not show where the chips were located, how access was obtained, or whether export controls were violated.

unemploymentlabs.com ↗31 min read at source

Adding clean data can turn weak poison into a hidden backdoor

A few poisoned fine-tuning examples that do little alone can create a hidden backdoor when mixed with clean examples, so auditing poison in isolation may miss the risk. In the researchers’ setup, poison examples pair a task with a trigger, while clean examples remove that trigger; regularization then makes the model attribute the behavior to the trigger. They reproduce the effect in small Qwen3-8B, Gemma-3-4b-it, and MNIST experiments, not a production pipeline.

Google’s agent defense catches 87% of executed attacks, misses poisoned memory

Agent Anomaly Detection flagged 87% of harmful instructions that agents actually carried out across 1,774 sessions, giving teams a way to review attacks after execution without adding latency. It reads agent traces and logs, then learns each agent’s normal behavior from at least 300 benign runs. The weak spot is poisoned long-term memory: it caught just 4 of 18 cases, while false alarms ranged from 2% to 16%.

Cursor says API credentials reached less than 0.01% of customers

Cursor says an exposed vendor access key led to less than 0.01% of customers having API credentials exposed, while most customers were unaffected and affected users were contacted directly. The attacker copied coding-agent traces from a shared bucket, used credentials found there to reach selected Cursor systems, and accessed limited source code, support data, metrics, and conversation summaries. A stolen credential also let the attacker place unauthorized Claude Code builds in Cursor's cached extension marketplace; Anthropic did not publish them.

cursor.com ↗5 min read at source

Safety evals need stronger methods before model scores guide deployment

Anthropic says safety evaluations should be rigorous enough to decide whether a model is ready to deploy, which safeguards should ship, and what safety training comes next. The evaluations test whether a model’s response makes harm more or less likely, but results can be gamed, measure the wrong thing, or fail to reproduce. Drawing on reforms such as preregistration and better measurement, the post recommends a more disciplined approach, starting with teams that combine technical and domain expertise.

anthropic.com ↗22 min read at source

Tools & Open Source

llama.cpp’s first major release makes breaking API changes visible

llama.cpp and its ggml tensor library have reached their first major release, giving developers building local AI products a clearer contract for future upgrades: breaking public C API and ABI changes will now require a major-version bump. In practice, downstream projects can distinguish breaking changes from backward-compatible features and bug fixes through semantic versioning. The libraries are distributed across Linux, macOS, Windows, Android, and iOS, with packages including Homebrew, conda-forge, and Debian.

huggingface.co ↗3 min read at source

You’re caught up.

That was all 12 stories in the 14 November issue.