Wednesday / AI 101 series: Kimi K2 Thinking and why all the hype

Friday / Interview: Great conversation with Anneka Gupta, CPO at Rubrik

The Minds of Modern AI: Jensen Huang, Geoffrey Hinton, Yann LeCun & the AI Vision of the Future by FT Live (video)

Forget BIT, It is All about TOKEN: Towards Semantic Information Theory for LLMs by Bo Bai

We Built Social Media for Agents and They Won’t Stop Posting by Harper Reed

Evaluating the Impact of AI on the Labor Market: Current State of Affairs by the Budget Lab

Deepnote with a new notebook

After seven years in stealthy development, Deepnote has gone open-source. The team’s new .deepnote format replaces the aging .ipynb with something future-proof: human-readable YAML, AI-native design, multi-language support, and a project-based structure fit for real-world teams and AI agents alike. As they say: “It’s a big step toward making data tools fully open and community-driven.”

Memories.ai brings memory to your device
Memories.ai has unveiled LVMM 2.0, a model designed to give machines something humans take for granted: persistent visual memory. In partnership with Qualcomm, this next-gen tech will run on-device across phones, cameras, and wearables by 2026 – bringing sub-second video search, privacy-preserving inference, and real-time visual recall to the edge. Goodbye scrubbing through footage; hello semantic memory for machines. We are going to interview their CEO in December – send us your questions.

Webflow uncovers the tug-of-war behind the homepage
Webflow’s State of the Website 2026 reveals a digital battleground: marketing and engineering are at odds over strategy, governance, and control. 92% see cross-functional friction, 97% feel the weight of technical debt, and developers are increasingly frustrated – some to the point of quitting. Meanwhile, AI is knocking, but half the teams aren’t sure it’s safe to let it in.

Google shoots for the stars – literally
Project Suncatcher is Google Research’s latest moonshot: solar-powered satellite constellations armed with TPUs, linked by high-speed optical comms, and designed to scale AI compute in space. Think cloud infrastructure – just in orbit. With bench-tested bandwidths of 1.6 Tbps and radiation-tolerant TPUs, it's a wild yet plausible vision of off-planet machine learning. If this flies, “cloud computing” may soon be a literal term.

Google Cloud sharpens its silicon for the AI era
Google Cloud has officially launched Ironwood TPUs – its most powerful and efficient custom chips yet – delivering 10x the peak performance of v5p and redefining inference at scale. Alongside, new Arm-based Axion VMs promise serious price-performance gains for general-purpose workloads. Welcome to the AI Hypercomputer age: optimized from chip to cluster, and built to scale like never before.

OpenAI warns: we’re not ready
In its latest dispatch, OpenAI reflects on how AI has quietly passed historic milestones – outthinking humans in elite domains – while most still see it as a fancy chatbot. their concern is that tech is racing ahead of public understanding and governance. OpenAI urges new safety norms, government coordination, and a full-blown AI resilience ecosystem before superintelligence arrives. They are peculiar people, this OpenAI.

PHUMA: Physically-Grounded Humanoid Locomotion Dataset – curate large-scale video-based humanoid motion data with physics-constrained retargeting that enforces joint limits and contact fidelity, producing physically reliable motions for robust imitation learning →read the paper

TWIST2: Scalable, Portable, and Holistic Humanoid Data Collection System – develop a mocap-free, VR-based teleoperation system for fast, low-cost humanoid data collection and whole-body visuomotor control, enabling dexterous manipulation and dynamic locomotion →read the paper

VSI-590K: Spatially-Focused Instruction-Tuning Dataset – build a large-scale dataset centered on spatial reasoning, aggregating diverse sources with fine-grained spatial annotations to improve models’ spatial understanding →read the paper

🌟🌟 🦋 Kimi K2 Thinking – the model that blew everyone’s mind. We will cover it on Wednesday including the researchers opinions about it. In short, it is an open-source long-horizon reasoning agent supporting hundreds of sequential tool calls, optimized for INT4 inference, achieving frontier performance on reasoning, coding, and web-agent benchmarks through deep tool integration →read the paper

🌟🦋 NVIDIA Nemotron Nano V2 VL – advance document and video understanding with a hybrid Mamba-Transformer vision-language architecture using token reduction for efficient long-context reasoning; released in multiple precision formats with open datasets and recipes →read the paper

iFlyBot-VLA – integrate language, vision, and action through dual-level supervision that aligns high-level latent actions with low-level control tokens, producing a unified VLA model capable of precise 3D reasoning and real-world manipulation →read the paper

🌟🌟 Cambrian-S: Towards spatial supersensing in video
Researchers from New York University and Stanford University introduce Cambrian-S, a family of spatially grounded multimodal models trained on a 590K-sample video dataset (VSI-590K) for spatial reasoning. They propose VSI-SUPER, a benchmark with two tasks: spatial recall (VSR) and spatial counting (VSC), using up to 240-minute videos. Cambrian-S achieves 30%+ gains on VSI-Bench but fails on VSI-SUPER, revealing scale limitations. A predictive sensing prototype using latent frame prediction and surprise-based memory outperforms Gemini-2.5-Flash on VSI-SUPER tasks, showing prediction aids long-horizon spatial understanding →read the paper

🌟 Nested learning: A new ML paradigm for continual learning (by Google) – frame models as nested optimizers with self-modifying “Hope” architecture and continuum memory to sustain long-context reasoning and resist forgetting →read the paper

🌟 Scaling Agent Learning via Experience Synthesis (by Meta) – distill environment dynamics into a reasoning-based experience model to generate scalable synthetic rollouts, warm-start RL, and match PPO/GRPO with far fewer real interactions →read the paper

🌟🦋 Magentic Marketplace: An open-source simulation environment for studying agentic markets (by Microsoft) – simulate two-sided markets of assistant and service agents to evaluate welfare, bias, prompt-injection risks, and search design under realistic competition →read the paper

Simulating Environments with Reasoning Models for Agent Training – synthesize SFT trajectories (Simia-SFT) and RL feedback (Simia-RL) with LLM-simulated environments to train agents without real APIs, surpassing strong baselines on τ²-Bench →read the paper

🌟 Visual Spatial Tuning – construct VST-P (4.1M) and VST-R (135k) and train VLMs with SFT→RL to enhance spatial perception→reasoning without hurting general abilities →read the paper

🌟 Towards Mitigating Hallucinations in Large Vision-Language Models by Refining Textual Embeddings (by University of Maryland, Dolby Laboratories, Hilabs, Capital One) – integrate average-pooled visual features into textual embeddings to rebalance modalities, improving grounding and reducing hallucinations →read the paper

Actial: Activate Spatial Reasoning Ability of Multimodal LLMs – build Viewpoint-100K and train via SFT + GRPO to enforce cross-view consistency, improving 3D/spatial reasoning in- and out-of-domain →read the paper

🌟 Orion-MSP: Multi-Scale Sparse Attention for Tabular In-Context Learning (by Lexsi Labs) – combine multi-scale processing, block-sparse attention, and Perceiver-style memory to capture hierarchical feature interactions and scale to wide tables →read the paper

🦋 Trove: A Flexible Toolkit for Dense Retrieval – provide low-code, on-the-fly dataset filtering/combination, unified evaluation and hard-negative mining, and multi-node scaling for customizable retrieval research →read the paper

🦋 Shorter but not Worse: Frugal Reasoning via Easy Samples as Length Regularizers in Math RLVR – retain/up-weight moderately easy items in RLVR to regularize length, yielding equally accurate but much shorter solutions without explicit penalties →read the paper

🌟 Diffusion Language Models are Super Data Learners (by National University of Singapore) – show DLMs outperform AR models under data scarcity via any-order modeling, denoising compute, and MC augmentation, achieving strong accuracy with repeated data →read the paper

The Strong Lottery Ticket Hypothesis for Multi-Head Attention Mechanisms – prove existence conditions for strong lottery tickets in MHA and extend theory to transformers without normalization, with exponential error decay empirics →read the paper

Jr. AI Scientist and Its Risk Report: Autonomous Scientific Exploration from a Baseline Paper – orchestrate an agentic research workflow (analyze→hypothesize→implement→experiment→write) and audit capabilities/risks via AI reviewers and venue evaluations →read the paper