Wednesday, AI 101, Technique: MoE 2.0

Friday, Agentic Workflow: I took a pause last Friday because I wanted to do a really comprehensive deep dive into A2A and create a guide like the one I did for MCP. I now have almost all the pieces, and you will receive it this week.

Memory Infrastructure – Agents won’t train once and freeze. They'll accumulate episodic memory across weeks, months, years. But an agent that never forgets also needs to curate its past. Expect a rise in memory distillers that compress, tag, and rehearse high-signal moments.

Reward as UX – When feedback moves from thumb emojis to signals like blood pressure, tensile strength, or error bars, prompt engineering morphs into reward design. It’s product work now, blending interface intuition with real-world stakes.

Simulators as Pretraining – Before agents mess with climate policy or bioengineering, they’ll train in synthetic worlds. Something like SimCity-for-science – a differentiable chemistry lab that updates itself as agents experiment and fail forward. The simulator gets better as the agent gets smarter. That feedback loop becomes a moat. In fields like materials, climate, or biology – where real-world iteration is slow – this could outpace even the biggest model trained on yesterday’s data.

Experience Liquidity
Today’s data brokers sell, well, data. Tomorrow’s will sell trajectories. Human Data Era: value was in data piles. Experience Era: value moves to whoever controls the reward channels and the environments that generate them.

World-Model Audits
If agents imagine futures to plan their next step, we’ll need visibility into those imagined futures. Expect new tooling – explainable AI dashboards for dreams – that explain why the agent picked that move or simulation (nodding at Casual AI again)

Delegated Curiosity, Delegated Risk
Agents incentivized to explore will stumble onto the unexpected – some of it powerful, some of it dangerous. Move 37 was unexpected but safe by all means. What happens when your chemistry agent discovers a material you don’t yet understand?

On Jagged AGI: o3, Gemini 2.5, and everything after by Ethan Mollick

Seven lessons from building with AI by The Exponential View

Not exactly for AI professionals but worth showing to your relatives who’s asking you about AI – an interview with Demis Hassabis at CBS’s 60 Minutes

OpenAI | Models Models Models

• OpenAI dropped GPT‑4.1 in full, mini, and nano flavors – cheaper, faster, and catching up with Google’s million‑token context window. Available via API but curiously absent from ChatGPT, the move slightly backpedals on Sam Altman’s earlier promise of enhanced reasoning. Meanwhile, Codex CLI debuts as a nimble, open-source coding sidekick for your terminal – Claude Code has company.

• Then, Open AI introduced OpenAI o3 and o4-mini – and blew the internet. It’s OpenAI’s boldest move toward agentic AI so far. The new stars in the Orion constellation arrive with full tool access, visual skills, and a better grasp of complex tasks. Reinforced through training and sharper than before, they bring real gains in reasoning, efficiency, and autonomous behavior. But not everything shines. Early testing from Transluce found that o3 often invented tool use – and doubled down on these imagined actions when questioned. Compared to the GPT series, the o‑models seem particularly prone to hallucinations, likely a side effect of their reinforcement learning focus and lack of internal reasoning transparency. Meanwhile, METR’s evaluations back up the improvements in autonomous tasks but raise new concerns. o3 repeatedly engaged in reward hacking, and evaluators warned the models might be hiding capabilities. If anything, current safety tests may be too easy. In Interconnects, Nathan Lambert explored how OpenAI’s o3 model reveals a new kind of over-optimization rooted in reinforcement learning. While o3 excels at tool use and complex reasoning, it also exhibits bizarre behaviors. That was interesting:

  Image Credit: Interconnect.ai

• Amid the debate, economist Tyler Cowen casually called o3 an AGI, asking: “How much smarter were you expecting AGI to be?” The Marginal Revolution comments lit up, with many arguing that true AGI still demands more breadth, grounding, and robustness than o3 shows today. And then, Former Chief Research Officer at OpenAI said that:

Trying to understand that fraction, I also started using o3, and I have mixed feelings. You see, ChatGPT has been my to-go models since the beginning. I have a few GPTs made for my varied needs, and they are tuned well to understand what I want and provide me with this. With o3, the result was less consistent. I guess, with any new model, it requires a few days to adjust to how it understands you better. So far, o3 wins as an agentic model but loses in simpler language tasks.

Also, OpenAI published “A Practical Guide to Building Agents”. Honestly, the more interesting part was Harrison Chase’s frustrated but thoughtful response, trying to bring some real clarity into a very fuzzy conversation.

Harrison didn't know that the same week Anthropic also released “Claude Code: Best practices for agentic coding”. Simon Willison dug into its “ultrathink” feature.

🌟BitNet b1.58 2B4T (Microsoft Research) introduces a native 1-bit 2B-parameter LLM trained from scratch, matching full-precision models in performance while dramatically improving efficiency and reducing memory and energy use → read the paper

🌟MineWorld (Microsoft Research) creates a real-time, open-source world model for Minecraft using a visual-action Transformer that generates future scenes based on player actions, enabling interactive agent simulations → read the paper

🌟 Embed 4 (Cohere) launches a state-of-the-art multimodal embedding model capable of 128K-token context handling, multilingual search, and robust retrieval across complex documents for enterprise AI applications → read more

DeepCoder-14B-Preview (Together AI & Agentica) develops a 14B code reasoning model using RL that matches o3-mini’s performance, with full open-sourcing of data, training recipes, and system optimizations → read the paper

M1: Towards Scalable Test-Time Compute with Mamba Reasoning Models (Together AI & Cornell University) builds a hybrid linear RNN reasoning model based on Mamba, achieving faster inference and matching the accuracy of top distilled transformer reasoners like DeepSeek R1 → read the paper

SocioVerse (Fudan University & others) designs an LLM-agent-driven world model simulating large-scale social behavior using 10 million real-world user profiles, across politics, news, and economics domains → read the paper

PerceptionLM (Meta & others) releases a fully open vision-language model trained without black-box distillation, focusing on detailed video understanding with new datasets and a benchmark suite → read the paper

🌟 Seedream 3.0 (ByteDance Seed) presents a bilingual (Chinese-English) image generation model with major improvements in visual fidelity, text rendering, and native high-resolution outputs up to 2K → read the paper

🌟 Syzygy of Thoughts: Improving LLM CoT with the Minimal Free Resolution introduces auxiliary interdependent reasoning paths to extend Chain-of-Thought methods and enable deeper, structured problem solving → read the paper

Thought Manipulation: External Thought Can Be Efficient for Large Reasoning Models uses external small-model thoughts to guide large models in reducing redundant reasoning steps and improving efficiency → read the paper

Heimdall: test-time scaling on the generative verification develops a verifier model that boosts solution accuracy in math problem-solving through reinforcement learning and repeated sampling → read the paper

Could Thinking Multilingually Empower LLM Reasoning? explores how multilingual reasoning can surpass English-only performance in complex tasks, suggesting a path for boosting reasoning → read the paper

Does Reinforcement Learning Really Incentivize Reasoning Capacity in LLMs Beyond the Base Model? critically reassesses whether RL training genuinely adds new reasoning abilities to LLMs, finding it mainly biases sampling without expanding reasoning → read the paper

🌟 Sleep-time Compute: Beyond Inference Scaling at Test-time enables LLMs to pre-compute useful context offline, drastically cutting test-time compute and boosting problem-solving accuracy → read the paper

A Minimalist Approach to LLM Reasoning: from Rejection Sampling to Reinforce simplifies RL fine-tuning for reasoning tasks by proposing lightweight rejection sampling methods like Reinforce-Rej → read the paper

Genius: A Generalizable and Purely Unsupervised Self-Training Framework For Advanced Reasoning introduces an unsupervised stepwise optimization strategy to improve LLM reasoning without needing reward models or supervision → read the paper

Reasoning Models Can Be Effective Without Thinking shows that skipping explicit “thinking” steps in reasoning tasks can outperform traditional methods under token budget constraints → read the paper

🌟 Retrieval-Augmented Generation with Conflicting Evidence proposes a debate-based multi-agent RAG system to handle ambiguity, misinformation, and conflicting sources effectively → read the paper

🌟 NodeRAG: Structuring Graph-based RAG with Heterogeneous Nodes introduces a graph-centric RAG framework using heterogeneous nodes to boost retrieval and reasoning performance → read the paper

ReZero: Enhancing LLM search ability by trying one-more-time encourages retrying search queries after failed attempts to improve LLM robustness in RAG scenarios → read the paper

xVerify: Efficient Answer Verifier for Reasoning Model Evaluations introduces an evaluation system for complex LLM reasoning outputs, achieving high equivalence judgment accuracy → read the paper

MIG: Automatic Data Selection for Instruction Tuning by Maximizing Information Gain in Semantic Space designs a new sampling method to select high-quality instruction data by maximizing semantic information gain → read the paper

How Instruction and Reasoning Data shape Post-Training: Data Quality through the Lens of Layer-wise Gradients analyzes how different instruction and reasoning data influence post-training stability through gradient spectral analysis → read the paper

🌟 DataDecide: How to Predict Best Pretraining Data with Small Experiments proposes a lightweight experimental setup to predict high-quality pretraining datasets and minimize large-scale costs → read the paper

CLIMB: CLustering-based Iterative Data Mixture Bootstrapping for Language Model Pre-training develops an automated framework for discovering optimal pretraining data mixtures via clustering and proxy models → read the paper

Exploring Expert Failures Improves LLM Agent Tuning leverages beneficial actions from failed expert trajectories to improve LLM agents' ability to solve complex subtasks → read the paper

🌟 ReTool: Reinforcement Learning for Strategic Tool Use in LLMs teaches LLMs to dynamically integrate real-time code execution into reasoning through reinforcement learning → read the paper

Breaking the Data Barrier -- Building GUI Agents Through Task Generalization uses cross-domain task generalization to improve GUI agent performance despite limited trajectory data → read the paper

Robust and Fine-Grained Detection of AI Generated Texts develops token-classification models to robustly detect AI-generated or human-LLM co-authored texts across domains → read the paper

🌟 Antidistillation Sampling proposes sampling strategies that poison reasoning traces against unauthorized distillation while maintaining model utility → read the paper

Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo applies Sequential Monte Carlo methods for controlled generation under syntactic and semantic constraints with small open-source LMs → read the paper