Wednesday / AI 101 series: What is Continual Learning? (Including Nested Learning)

Friday / Lessons learnt from AIE Code in NYC

A New Testing Playbook to properly test systems that don't follow a predictable path, using "checkpoint verification"

When to use a single, capable agent and when to split into a multi-agent design based on domain, governance, or model requirements.

How to build agents that capture real value through judgment, rather than simple, rule-based tasks.

Training computer – Azure with Foundry on top, routing across models, hosting agents, coordinating multi-agent plans, and evaluating cost and quality.

Context computer – Fabric IQ, Work IQ, Foundry IQ, a permissioned semantic layer that gives agents a live view of people, data and tasks.

Agent computer – Windows, the Microsoft 365 apps, and Edge, where actions actually happen near files and browsers.

Governance computer – Entra, Purview, Defender and Agent 365, the identity, policy and audit spine.

AI Slopaganda in the KMT Election by ChinAI

The Myth of China's "AI Talent Pipeline" by ChinaTalk

Will it be a Home Run for Alibaba? by AI Supremacy

Gemini 3 is in the air! And everyone is talking about it but it’s not out yet. Let’s wait and see it in action. Other news →

Quora’s Mission Upgrade
Does anyone still use Quora? Some people do though it’s not really clear to me why in the age of AI and AI slop. But Adam D’Angelo thinks the important thing for Quora now not sharing knowledge (and slop) but nurturing “collective intelligence.” With Poe now handling multi-modal queries in private chats and Quora continuing as the public town square, they’re betting on synergy between humans and AI to accelerate problem-solving, tech progress, and, of course, engagement. Will see how that works.

OpenAI releases GPT-5.1, adds tone controls
GPT-5.1 is out, bringing updates to both Instant and Thinking models. Responses are meant to be more natural, reasoning more adaptive, and tone more customizable. There's a stronger push toward conversational warmth and user control. My personal experience is disappointing. For the first time since ChatGPT launched, I’m thinking of canceling the subscription (I’m on Pro).

They also published a GPT-5.1 Prompting Guide, which is interesting because it reveals parts of the system prompt – helping you understand the model’s personality better—and gives a few recommendations on how to prompt it efficiently.

DeepMind acting in 3D with SIMA 2
DeepMind introduces SIMA 2, a Gemini-powered agent that can reason, act, and learn in complex 3D virtual worlds. Unlike its predecessor, SIMA 2 doesn’t just follow instructions – it adapts, explains its decisions, and improves through self-play. Tested in games it wasn’t trained on, it shows promising generalization. A step forward for embodied AI, though challenges like long-horizon memory and precision remain.

Claude 4.5 Tackles Political Bias – Carefully

Anthropic released a detailed report on Claude’s political neutrality, introducing a new automated test for “even-handedness.” Claude Sonnet 4.5 scored higher than GPT-5 and Llama 4, performing on par with Grok 4 and Gemini 2.5 Pro. The company 🦋 open-sourced its methodology, aiming to establish shared standards for measuring political bias in AI.

🦋 NVIDIA open sourced Apollo 
NVIDIA unveiled Apollo, an open family of AI physics models optimized for industrial simulation in areas like fluid dynamics, electromagnetics, and multiphysics. Apollo integrates transformers, neural operators, and diffusion methods with domain-specific knowledge. Early adopters, including Siemens, Applied Materials, and Northrop Grumman, report speedups up to 500x in simulations. Pretrained models and workflows support custom training, inference, and deployment, enabling real-time digital twins and surrogate modeling across aerospace, semiconductors, and manufacturing →read their blog

Marble, a multimodal world model

Researchers from World Labs have released Marble, a multimodal generative model that creates, edits, and composes 3D worlds from text, images, video, or coarse 3D layouts. It supports iterative editing, fine control via a tool called Chisel, and export to Gaussian splats, triangle meshes, or video. Marble enables multi-image world generation, dynamic expansion, and high-fidelity visual enhancement. It targets creative industries, simulation, robotics, and spatial intelligence applications →read their blog

🦋 Omnilingual ASR, a speech-to-text model

Researchers from Meta introduced Omnilingual ASR, a suite of models supporting speech-to-text for over 1,600 languages, including 500 low-resource ones. It features a 7B-parameter wav2vec 2.0 encoder and two decoders: CTC and transformer-based LLM-ASR, achieving <10% character error rate in 78% of languages. The system enables new-language onboarding using just a few paired samples. Meta also released the 350-language Omnilingual ASR Corpus and open-sourced all models and tools →read their blog

🌟 🌟 🌟 LeJEPA – most likely the last paper co-authored by Yann LeCun in Meta-FAIR – is the next step for JEPA

🌟 The Path Not Taken: RLVR Provably Learns Off the Principals (by Meta AI, The University of Texas) – characterize RLVR updates as moving off principal directions in weight space, explaining sparse-looking changes via model geometry and contrasting RL’s regime with SFT for future geometry-aware algorithms →read the paper

SofT-GRPO: Surpassing Discrete-Token LLM Reinforcement Learning via Gumbel-Reparameterized Soft-Thinking Policy Optimization – reinforce soft-thinking chains with Gumbel-Softmax and reparameterized gradients so continuous reasoning traces slightly outperform discrete GRPO on Pass@1 and substantially on Pass@32 →read the paper

🌟 RLVE: Scaling Up Reinforcement Learning for Language Models with Adaptive Verifiable Environments – scale RL via 400 adaptive, verifiable environments that tune difficulty to the policy, yielding stronger general reasoning than continued RL on static data →read the paper

DRIVE: Data Curation Best Practices for Reinforcement Learning with Verifiable Reward in Competitive Code Generation – design RLVR datasets, entropy schedules, and a hard-focus curriculum (Pre-GRPO) that push code-generation models to SOTA performance on LeetCode and Codeforces →read the paper

RedOne 2.0: Rethinking Domain-specific LLM Post-Training in Social Networking Services – stage exploratory RL, targeted SFT, and SNS-focused refinement RL to adapt a 4B model to social workloads while preserving robustness and improving data efficiency →read the paper

🌟🌟 Intelligence per Watt: Measuring Intelligence Efficiency of Local AI (by Stanford, Together AI) – benchmark 20+ local LLMs across 8 accelerators and 1M queries using accuracy-per-power (IPW), showing rapid gains in local capability and room for hardware optimization →read the paper

Routing Manifold Alignment Improves Generalization of Mixture-of-Experts LLMs – align routing-weight manifolds with task-embedding manifolds so similar tasks share expert patterns, improving MoE generalization via lightweight router finetuning →read the paper

🌟 Teaching Pretrained Language Models to Think Deeper with Retrofitted Recurrence – retrofit pretrained transformers with depth recurrence and a recurrence curriculum to increase effective depth and beat standard post-training at matched compute →read the paper

🌟TiDAR: Think in Diffusion, Talk in Autoregression (by Nvidia) – fuse diffusion-based drafting and AR sampling inside one structured-attention forward pass to match AR quality while delivering multi-x token throughput and full KV-cache support →read the paper

🌟 Black-Box On-Policy Distillation of Large Language Models (by Microsoft) – train a student as a generator against a discriminator that distinguishes teacher vs student outputs, enabling on-policy black-box distillation that matches GPT-5-Chat quality for a 14B student →read the paper

Reasoning with Confidence: Efficient Verification of LLM Reasoning Steps via Uncertainty Heads – attach tiny uncertainty heads (UHeads) to a frozen LLM and use its hidden states to estimate step-level uncertainty, matching or beating huge PRMs for reasoning verification →read the paper

Optimizing Diversity and Quality through Base-Aligned Model Collaboration – route token-level decoding between a base and its aligned variant based on uncertainty and semantic role to jointly boost diversity and quality in open-ended generation →read the paper

🌟Beyond Fact Retrieval: Episodic Memory for RAG with Generative Semantic Workspaces (by University of California) – construct a neuro-inspired Generative Semantic Workspace that incrementally builds structured episodic memory, enabling large gains on long-horizon EpBench with fewer context tokens →read the paper

Adaptive Multi-Agent Response Refinement in Conversational Systems – refine responses with specialized agents for factuality, personalization, and coherence, coordinating them dynamically per query to outperform single-agent refiners on challenging dialogue tasks →read the paper

IterResearch: Rethinking Long-Horizon Agents via Markovian State Reconstruction – reformulate deep research as iterative Markovian workspace reconstruction with RL-based Efficiency-Aware Policy Optimization, scaling interactions to 2048 steps and sharply improving long-horizon performance →read the paper

The Station: An Open-World Environment for AI-Driven Discovery – simulate an open-world scientific ecosystem where agents read, hypothesize, code, and publish without central coordination, yielding emergent methods and SOTA on diverse scientific benchmarks →read the paper