Wednesday / AI 101 series: Guardian Models

Friday / Interview: Ulrik Stig Hansen, co-founder of Encord

“70% of usage is non-work” is true for Free/Plus/Pro users, but you can’t generalize that to all usage when a giant slice of the pie – enterprise accounts where work dominates – is off the table.

The “work vs non-work” trend is real within consumer accounts, but doesn’t tell us what’s happening in offices, classrooms, or enterprise workflows. Because they also use ir for work and non-work.

If the researchers had just titled it How Consumers Use ChatGPT, no problem.

Because they didn’t, the report risks being quoted as “proof” that ChatGPT is mostly non-work everywhere, which isn’t supported by their own sampling choices.

OpenAI report “How people use ChatGPT”

Anthropic’s MCP goes public
The MCP Registry has landed – an open catalog and API for discovering publicly available MCP servers. It’s designed as a single source of truth, enabling both public and private sub-registries to thrive without stepping on toes. With a community-moderated model and open-source foundation, it’s a foundational step toward scaling context-aware AI. A quiet launch, but one with deep roots and broad ambitions.

Oracle’s Loud Pivot
After a decade of quiet infrastructure work, Oracle just shouted its way into the AI big leagues. With a record-setting compute deal in the works and AI demand visibly swelling its backlog, Oracle looks less like a dusty database vendor and more like the connective tissue of enterprise AI. It skipped the model arms race and built the rails – data, governance, and distribution – for others to ride.

Devin Goes to Eleven
Cognition AI, the team behind coding agent Devin, just raised $400M at a $10.2B valuation – up from $4B earlier this year. With ARR jumping from $1M to $73M in under a year and net burn under $20M, the numbers are as aggressive as the company culture. Long hours, layoffs, and buyouts haven’t scared investors – or slowed growth. It’s a hyperloop ride in both valuation and velocity. We just published a super detail deep dive about them →read it here

Absolutely brilliant Melanie Mitchell Magical Thinking on AI

One of our favorite visionaries Demis Hassabis on AI, Creativity, and a Golden Age of Science | All-In Summit

Fully autonomous robots are much closer than you think – Sergey Levine with Dwarkesh Patel

VaultGemma – train a 1B decoder-only Gemma variant fully under differential privacy, demonstrate practical DP scaling laws, and release open weights for privacy-preserving applications →read the paper (pdf)

Hunyuan-MT / Hunyuan-MT-Chimera – build multilingual translation models across 33 languages and aggregate multi-setting outputs at test time to boost robustness, achieving state-of-the-art WMT2025 performance →read the paper

mmBERT – pretrain a modern multilingual encoder on 3T tokens with annealed language learning to lift classification and retrieval in both high- and low-resource languages →read the paper

Qwen3-Next – combine gated DeltaNet and gated attention with an ultra-sparse MoE and native multi-token prediction to deliver long-context efficiency while activating ~3B of 80B parameters →read the paper

A Survey of Reinforcement Learning for Large Reasoning Models →read the paper

Reinforcement learning foundations for deep research systems: A survey
Researchers from Huawei Technologies surveyed RL approaches for training deep research systems with hierarchical agents. They examined data synthesis methods like cross-document and obfuscated queries, RL techniques for long-horizon credit assignment, reward design, and multimodal reasoning, and frameworks such as GRPO and DUPO. The survey highlights system bottlenecks, coordination strategies, and benchmarks, offering a roadmap for building scalable, tool-using, and evaluation-ready agentic research systems →read the paper

🌟 Tool-space interference in the MCP era: Designing for agent compatibility at scale (Microsoft) – analyze how tool catalogs interact in the Model Context Protocol ecosystem and propose ways to prevent cross-agent inefficiencies →read the paper

🌟 Paper2Agent: Reimagining Research Papers As Interactive and Reliable AI Agents (Stanford) – convert research papers into interactive MCP-based agents that can execute the original workflows and extend them →read the paper

🌟 Virtual Agent Economies (Google DeepMind) – conceptualize agent-to-agent markets and explore auction mechanisms, mission economies, and governance for steerable AI economies →read the paper

WebExplorer: Explore and Evolve for Training Long-Horizon Web Agents – generate challenging web navigation data and train agents with long contexts and tool calls for state-of-the-art browsing →read the paper

EnvX: Agentize Everything with Agentic AI – transform GitHub repositories into autonomous agents capable of natural interaction and cross-repository collaboration →read the paper

🌟Bootstrapping Task Spaces for Self-Improvement (Meta) – train models with exploratory iteration to grow task spaces and enable inference-time self-improvement across math, tool-use, and ML tasks →read the paper

Parallel-R1: Towards Parallel Thinking via Reinforcement Learning – instill parallel reasoning through curriculum and RL, using multi-path exploration as a scaffold for stronger problem solving →read the paper

AgentGym-RL: Training LLM Agents for Long-Horizon Decision Making through Multi-Turn Reinforcement Learning – provide a unified framework and scaling strategy to train LLM agents for multi-turn decision making across realistic environments →read the paper

Harnessing Uncertainty: Entropy-Modulated Policy Gradients for Long-Horizon LLM Agents – stabilize learning with uncertainty-aware gradient modulation, amplifying confident correct updates while dampening unstable ones →read the paper

Staying in the Sweet Spot: Responsive Reasoning Evolution via Capability-Adaptive Hint Scaffolding – dynamically adjust problem difficulty with adaptive hints to keep training efficient and aligned with model capacity →read the paper

SFR-DeepResearch: Towards Effective Reinforcement Learning for Autonomously Reasoning Single Agents – reinforce reasoning-oriented agents with synthetic data to strengthen autonomous deep research skills →read the paper

ΔL Normalization: Rethink Loss Aggregation in RLVR – minimize gradient variance in verifiable reward training by normalizing losses for variable-length outputs →read the paper

Sharing is Caring: Efficient LM Post-Training with Collective RL Experience Sharing – decentralize RL post-training with asynchronous rollout sharing to scale efficiently across heterogeneous hardware →read the paper

🌟 The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs – show how compounding per-step accuracy yields exponential gains in long tasks, and why execution errors dominate over reasoning gaps →read the paper

🌟 Why Language Models Hallucinate (OpenAI) – explain hallucinations as statistical pressures from training and evaluation incentives that reward guessing, not calibrated uncertainty →read the paper

The Majority is not always right: RL training for solution aggregation – train aggregators that reconcile multiple candidate solutions into a correct answer, outperforming majority voting →read the paper

Test-Time Scaling in Reasoning Models Is Not Effective for Knowledge-Intensive Tasks Yet – reveal that longer reasoning often increases hallucinations in fact-heavy settings, limiting test-time scaling benefits →read the paper

🌟 Reasoning Introduces New Poisoning Attacks Yet Makes Them More Complicated (Google DeepMind) – demonstrate decomposed reasoning poison attacks that target chain-of-thought while also revealing emergent robustness →read the paper

🌟 All You Need Is A Fuzzing Brain: An LLM-Powered System for Automated Vulnerability Detection and Patching (Texas A&M University) – build an LLM-driven system that discovers and patches software vulnerabilities, validated in DARPA’s AIxCC →read the paper

🌟 R2AI: Towards Resistant and Resilient AI in an Evolving World (Tsinghua) – propose a safe-by-coevolution paradigm where AI develops immunity-like resistance and resilience through adversarial feedback loops →read the paper

🌟 Statistical Methods in Generative AI – survey how statistical tools can improve reliability, fairness, and safety in generative AI pipelines →read the paper

Guided Decoding and Its Critical Role in Retrieval-Augmented Generation – compare decoding frameworks that constrain RAG outputs to structured formats, balancing hallucination control and usability →read the paper

MachineLearningLM: Scaling Many-shot In-context Learning via Continued Pretraining – equip models with strong in-context ML capabilities via causal model–based pretraining and efficient serialization →read the paper

Revolutionizing Reinforcement Learning Framework for Diffusion Large Language Models – introduce trajectory-aware RL for diffusion LMs, yielding smaller yet stronger reasoning models →read the paper

🌟 Language Self-Play For Data-Free Training (Meta) – use game-theoretic self-play to let models improve without external data, showing stronger task performance than data-driven baselines →read the paper

🌟 Causal Attention with Lookahead Keys – extend causal attention with lookahead keys to blend forward-looking context without breaking autoregressive constraints →read the paper

🌟 Visual Representation Alignment for Multimodal Large Language Models (KAIST) – align multimodal LLMs’ vision pathways with pretrained VFMs to improve fine-grained visual reasoning →read the paper

Can Understanding and Generation Truly Benefit Together – or Just Coexist? – unify image understanding and generation through reconstruction-based RL, showing mutual improvements →read the paper

Visual Programmability: A Guide for Code-as-Thought in Chart Understanding – enable adaptive chart reasoning via code-as-thought pathways and RL-based strategy selection →read the paper