AI Digest

Category

Jensen Huang resigns as Nvidia CEO; CFO Colette Kress named interim CEO

TLDR AI

Nvidia, the company that supplies most of the hardware behind AI, has lost its founder and has no permanent successor yet.

  • Huang resigned as CEO and board member, effective immediately, after reports of extramarital affairs, including one with an alleged Chinese intelligence asset.
  • Kress, CFO since 2013, will hold both roles while the board considers internal and external candidates. Long-time director Harvey C. Jones becomes Nvidia's first board chair.
  • Nvidia will still report third-quarter fiscal 2027 results on November 17 as scheduled.
Read on nvidianews.nvidia.com →

Google releases Gemini 4 Hydrogen, a fast, low-cost model for coding and agents

TLDR AI

Google's new efficient model scores higher than its Gemini 4 Argon model on a standard intelligence index and costs less to run.

  • It scored 51 on the Artificial Analysis Intelligence Index, beating Gemini 4 Argon at its highest thinking level.
  • Pricing is $1.50 per million input tokens and $7.50 per million output tokens.
  • It has a 1M-token context window.
Read on blog.google →
Benchmarks
BenchmarkGemini 4 HydrogenClaude Fable 5.2GPT-6.2 AstraMuse WonderKimi K3.1 ProΔ vs best other
GPQA diamond96.896.096.8—94.2+0.0
FrontierMath-Tiers-1-3-v2-Private88.195.192.682.178.2-7.0
FrontierMath-Tier-4-v2-Private87.8100.0100.070.765.9-12.2
OTIS Mock AIME 2024-2025100.099.7100.099.4100.0+0.0
SimpleQA Verified70.994.078.5—62.2-23.1
Chess Puzzles60.079.078.0—45.0-19.0
EBR-bench70.590.592.431.437.1-21.9
Mystery Game Puzzles65.090.095.051.042.0-30.0
SimpleBench91.794.979.2——-3.2
APEX-Agents78.080.172.266.165.5-2.1
Furniture Assembly57.590.089.254.255.8-32.5
FrontierSWE53.169.9———-16.8
SpireClimb7.715.341.3—0.0-33.6

Δ = Gemini 4 Hydrogen minus the best other model listed. Sources: epoch.ai · Data: Epoch AI (CC BY 4.0)

Forensic analysis finds signs of export-controlled Nvidia Blackwell GPUs in top Chinese models

TLDR AI

The analysis suggests several leading Chinese labs are training on chips that US export controls are meant to keep from them.

  • Six models released since April by DeepSeek, Moonshot AI, MiniMax and Ant Group contain numerical patterns that match Blackwell hardware.
  • Former OpenAI researcher Will DePue made the claim, citing GPU fingerprints found in the published model weights.
  • The analysis does not show where the chips were, how access was obtained, or whether export controls were broken.
Read on unemploymentlabs.com →

Google opens pre-orders for its first Gemini AI glasses, starting at $395

The Rundown AI

Google is taking on Meta, which shipped 81% of AI glasses in the first half of the year, and is selling privacy as the difference.

  • The glasses were built with Samsung and come in frames from Warby Parker and Gentle Monster. They have a 12MP camera, open-ear speakers, up to nine hours of battery and no in-lens display.
  • Google says the glasses blur bystanders' faces before Gemini sees them, show a light while recording, and stop recording if the light is covered.
  • They go on sale December 10 in the US and South Korea and also work with iPhones.
Read on blog.google →

OpenAI pledges $2M for research on AI's impact on children

The Rundown AI

OpenAI is funding outside work to measure the risks AI poses to children.

  • OpenAI joined a child safety initiative that is developing tools to measure risks and build better safeguards and policies.
Read on openai.com →

Anthropic opens a private beta of self-hosted Claude Managed Agents

TLDR AI

Companies can run Anthropic's managed agent services entirely on their own infrastructure, which gives them more security and control.

  • The full Managed Agents service runs in-house, including sandboxes, credential handling and session history.
  • It works with existing cloud providers and supports zero data retention policies.
Read on claude.com →

Claude Meetings lets Claude join Google Meet calls as a participant

TLDR AI

Claude can now work inside live meetings instead of only processing transcripts afterward.

  • Claude joins as a visible participant and takes notes, answers questions and builds artifacts during the call.
Read on support.claude.com →

llama.cpp and ggml reach v1.0.0 with semantic versioning

Hugging Face Blog

Many apps depend on llama.cpp for local AI, and it now commits to a stable API that marks breaking changes clearly.

  • From now on, breaking changes to the public C API will trigger major-version bumps for both llama.cpp and ggml.
  • Distribution covers Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and XCFramework.
  • llama.cpp supports more than 150 model architecture variants and all mainstream speculative decoding methods.
Read on huggingface.co →

Replit Agent can now buy supplies and run Google Ads for businesses

TLDR AI

Coding agents are moving into commerce and marketing tasks for the businesses they build for.

  • Purchases go through Link with Visa and Stripe for secure, controlled transactions.
  • The agent analyzes content and suggests strategies for ad campaigns.
Read on replit.com →

Perplexity releases Decider v1.3 and Decider Flash v1.3 classification models

TLDR AI

These fast models return probabilities that applications can use to classify, route, rank and evaluate inputs.

  • They support binary and multi-class classification with ordered or unordered labels.
  • Both versions are more accurate and robust across text and vision tasks.
  • A new feature highlights selected content for named entity recognition, personal information detection and extractive question answering.
Read on perplexity.ai →

Hark launches Handoff², a faster computer-use agent

The Rundown AI

Brett Adcock's startup says its web agent is both faster and more reliable than its previous version.

  • Hark says it completes web tasks 2x faster for users.
  • Its success rate in internal tests rose to 82%, from 73%.
Read on hark.com →

Meta's Muse assistant can now phone US businesses for users

TLDR AI

Meta is matching a calling feature that Grok Bot launched a week earlier.

  • AI calling is now available to all paid Muse subscribers.
  • The assistant contacts US businesses on users' behalf.
Read on techcrunch.com →

LiquidCodec tokenizes audio fast enough for local voice apps

TLDR AI

A fast, small audio codec makes it easier to build conversational apps that run on-device.

  • LiquidCodec generates speech quickly with a small memory footprint.
  • It is designed for applications that hold a conversation while running locally.
Read on liquid.ai →

Atlas is a single model that works across text, images, video and 3D

TLDR AI

Putting every input into one shared spatial context lets a single model generate, reconstruct and simulate worlds.

  • All inputs are combined into a shared spatial context.
  • Tasks include world generation, reconstruction and simulation.
Read on worldlabs.ai →

Meta finds a pretrained vision encoder isn't needed, but dedicated vision capacity is

Hugging Face Daily Papers

The controlled study shows which parts of vision encoding actually matter when building multimodal models.

  • An encoder trained from scratch together with the model matches a pretrained one, with no separate training stage.
  • Encoder-free models trail on visual understanding even when the backbone gets the encoder's parameter count.
  • Vision-only modules placed inside early backbone layers perform comparably to a separate encoder and can lower time to first token.
Read on huggingface.co →

Ai2's XYBench tests whether assistants spot the real problem behind a misguided question

Hugging Face Blog

Frontier models answer the literal question more often than they address what the user actually needs.

  • The benchmark has 8,115 questions from Stack Overflow, Stack Exchange, curated examples and WikiHow. It won the Best Resource Paper Award at EMNLP 2026.
  • On Stack Overflow, Claude Sonnet 4.6, GPT-5, Gemini 3.1 Pro, Qwen3.5-9B and DeepSeek-V4 Pro focused mainly on the user's underlying goal in only 35–41% of responses.
  • When asked to choose between two answers, the same models picked the goal-focused one 71–84% of the time. They can recognize good advice more often than they give it.
Read on huggingface.co →

LongCat's G-STAR compresses images 32–64x for faster autoregressive generation

Hugging Face Daily Papers

Heavy compression makes autoregressive image generation fast enough for practical use without hurting quality.

  • The paper shows that with a fixed information budget, group size should grow with the square of the per-axis compression factor.
  • At 1024×1024 resolution it is 20x faster than prior discrete autoregressive models, taking 2.5 seconds per image on one H200.
  • It performs competitively on five text-to-image benchmarks.
Read on huggingface.co →

SimpleHC matches mHC hyper-connections with only 2.4% training overhead

Hugging Face Daily Papers

Frontier models increasingly use hyper-connections, and this version is simpler and cheaper to train.

  • The authors drop residual mixing and share memory access, which they identify as a main source of overhead.
  • SimpleHC performs comparably to mHC when scaling from 15B to 30B MoE models.
  • Its gradients are smoother during training and it stays stable under stress.
Read on huggingface.co →

Ant Group's UniSpark-9B uses hybrid attention to handle both multimodal understanding and generation

Hugging Face Daily Papers

Mixing full attention with linear attention gives one model that can both understand and generate images.

  • It combines full attention with GatedDeltaNet linear attention on a Mixture-of-Transformer-Experts backbone.
  • It scores 2470.3 on MME-S and 0.83 on GenEval.
  • It was trained on 159.7M generation samples, 213.4M understanding samples and 624.8K interleaved text-image sequences.
Read on huggingface.co →

NVIDIA's BirdDreams driving world model generates a camera view and a bird's-eye view together

Hugging Face Daily Papers

Predicting the top-down layout alongside the camera video gives more realistic traffic and tighter control of the self-driving car's path.

  • Two video diffusion transformers attend to each other, one producing perspective video and one producing bird's-eye-view (BEV) video.
  • It needs only a single camera image, with no HD map or vehicle-state inputs.
  • On PhysicalAI-AV and nuScenes, it follows input trajectories for the self-driving car more faithfully than prior driving world models.
Read on huggingface.co →

SO-Arena finds the best System One model ranks only seventh on reliability

Hugging Face Daily Papers

Strong accuracy does not mean these fast decision models make reliable decisions.

  • The benchmark covers 11 domains and five capability families, including interactive sandbox tasks.
  • It measures reliability by whether decisions stay stable when the answer options change, and whether models notice when no valid option is offered.
  • It evaluated 21 models.
Read on huggingface.co →

Decision models that act poorly can still teach agents well

Hugging Face Daily Papers

Judging actions after an episode ends exposes consequences that acting in real time misses, which makes it useful for RL.

  • As actors, all six models tested on ALFWorld were pulled away from a required first subgoal toward an easier one.
  • Used as a teacher, JEV raised a Qwen2.5-1.5B policy from 72.8% to 82.9% success on ALFWorld, compared with GRPO alone.
  • On WebShop, success rose from 56.8% to 71.4%.
Read on huggingface.co →

Open Vision Banana is a fully open recipe for turning image generators into perception models

Hugging Face Daily Papers

It provides an open, reproducible version of Vision Banana, whose data, weights and code were never released.

  • It reaches 73.9 mIoU zero-shot on Cityscapes, compared with Vision Banana's reported 69.9.
  • A test-time method called legend ensembling improves mIoU by up to 8.7 points on Cityscapes and 15.0 points on COCO-Object without retraining.
  • The data pipeline, code and all checkpoints are released.
Read on huggingface.co →
BenchmarkGemini 4 HydrogenClaude Fable 5.2GPT-6.2 AstraMuse WonderKimi K3.1 Pro
GPQA diamond96.896.096.8—94.2
FrontierMath-Tiers-1-3-v2-Private88.195.192.682.178.2
FrontierMath-Tier-4-v2-Private87.8100.0100.070.765.9
OTIS Mock AIME 2024-2025100.099.7100.099.4100.0
SimpleQA Verified70.994.078.5—62.2
Chess Puzzles60.079.078.0—45.0
EBR-bench70.590.592.431.437.1
Mystery Game Puzzles65.090.095.051.042.0
SimpleBench91.794.979.2——
APEX-Agents78.080.172.266.165.5
Furniture Assembly57.590.089.254.255.8
FrontierSWE53.169.9———
SpireClimb7.715.341.3—0.0

Highlighted: released in this period. Bold: best score. Independent scores from Epoch AI (CC BY 4.0).