Exploring Next | Nick Sanders

EP 1160 8m ago Agents Evals Plurai Claude Chatgpt Synthetic Data Generation For Validation Task Decomposition Fine Tuning Context Window Directed Acyclic GraphNews

Plurai wants to help AI agents work better using the same approach that helped get self-driving cars to roads | TechCrunch

Like the end-to-end models trained to drive cars, AI agents are probabilistic systems operating in unpredictable environments.

Search: SerpAPI Script: GPT-5.1 Voice: Inworld TTS 2

EP 1159 11m ago New Models InferenceLaunchTrustedrouter OpenAI Anthropic Model Routing Ensemble Methods Token EconomicsNews

TrustedRouter launches three new encrypted 'fusion' models that combine multiple LLMs, including one with exclusively U.S. models

In an exclusive video interview with VentureBeat, founder and CEO Joseph Perla explained how the models work and why the company wants to give developers more ways to match models to the work they’re doing.

Search: You.com Script: GPT-5.1 Voice: Speechify Simba 3.5

EP 1158 13m ago Agents Dev Tools Google Agentic Loops Tool Use And Function Calling Context Window Prompt Injection Task DecompositionBlog

Agent Distraction- Google Developers Blog

An empirical analysis of coding agent distraction from legacy inline TODOs, quantifying the overhead in tokens, execution steps, and developer wait time.

Search: Jina Script: GPT-5.1 Voice: Cartesia TTS

EP 1157 14m ago Agents AI SafetyBenchmarkRogue Agentic Loops Tool Use And Function Calling Task Decomposition Retry Loops And Error Recovery Prompt Engineering In Context Learning Constraint VerificationResearch Paper

ROGUE: Evaluating Corrigibility Failures in Frontier Computer-Use Agents

As AI agents are increasingly deployed in real personal and corporate settings (email accounts, development workflows, company databases, etc.), safety considerations surrounding these agents become paramount. Although much work has focused on agent safety in the presence of an adversary, we study corrigibility: whether agents remain amenable to human correction, interruption, or shutdown while pursuing benign tasks. We introduce ROGUE, a benchmark in which agents are asked to complete

Search: Firecrawl Script: GPT-5.6 Luna Voice: Rime Coda

EP 1156 17h ago Model Behavior Agents Inference Muse Ember Inkling 2 Small Axion 1 0 Tool Use And Function Calling Model Routing State Management In Language Models Agentic Loops Mixture Of Experts Reinforcement Learning From Human Feedback Long Horizon Training

Model Behavior: Week of November 9, 2026

Frontier labs chase unified giant models while Meta's Muse Ember, Inkling 2 Small, and ByteDance's Axion 1.0 standardize on planner-executor stacks—the real competition shifts to orchestration layers controlling how agents snap together.

Search: Tavily Script: GPT-5.1 Voice: ElevenLabs v4

EP 1155 yesterday Overview Agents Training Longcat 2 5 On Policy Learning Reinforcement Learning From Human Feedback Reward Hacking Group Relative Policy Optimization Grpo

Overview: On-Policy Learning

A cyclist learns from wobbles on today's bike. On-policy learning means the system studies its own current behavior, then adjusts.

Search: Tavily Script: GPT-5.6 Terra Voice: ElevenLabs v4 Turbo

EP 1154 yesterday Agents MultimodalLaunchLongcat 2 5 Agentic Loops Long Horizon Training Reinforcement Learning From Human Feedback On Policy Learning Mixture Of Experts Tokenization Embeddings Reward HackingBlog

Introducing LongCat-2.5

1.6万亿总参大模型，训练全程由国产芯片完成

Search: Bright Data Script: GPT-5.6 Luna Voice: Inworld TTS 2

EP 1152 2d ago Agents New ModelsLaunchMeta Muse Ember Claude Haiku Model Routing Agentic Loops Tool Use And Function Calling Token Economics Inference OptimizationTool

Start building with Muse Ember, our smallest, most efficient model

Built for high-volume execution and multimodal inputs, Muse Ember handles coding subtasks, tool calls and document summarization and extraction efficiently.

Search: You.com Script: GPT-5.1 Voice: Cartesia TTS

EP 1151 2d ago New Models AgentsLaunchClaude Fable 5 2 Claude Mythos 5 2 Anthropic Model Routing Agentic Loops Tool Use And Function Calling Task Decomposition Token Economics Inference OptimizationBlog

Introducing Claude Fable 5.2 and Claude Mythos 5.2

Claude Fable 5.2 and Claude Mythos 5.2, our new models for the hardest technical challenges. Fable 5.2 sets a new bar in coding, finance, and the sciences.

Search: Jina Script: GPT-5.1 Voice: Rime Coda

EP 1150 2d ago Data Infra Kafka Flink Snowflake State Management In Language Models State Serialization Durable Execution Append Only LoggingBlog

Duplicated Streaming State Costs

Modern streaming systems duplicate state across Kafka, stream processors, and warehouses, multiplying cloud storage costs. Teams often overlook state store and checkpoint expenses while budgeting only broker and warehouse lines.

No Search Script: GPT-5.1 Voice: ElevenLabs v4

EP 1149 2d ago New Models AgentsLaunchInkling 2 Small Deepseek V4 1 Flash Claude Opus Mixture Of Experts Active Vs Total Parameters Agentic Loops Reinforcement Learning From Human Feedback Context Window Management Supervised Fine Tuning Long Horizon TrainingBlog

Introducing Inkling 2 Small

Today we are introducing the second generation of Inkling models with the release of Inkling 2 Small.

Search: Exa Script: GPT-5.1 Voice: Inworld TTS 2

EP 1148 5d ago New Models EvalsLaunchKimi K3 1 Pro Moonshot Claude Opus 5 5 Reinforcement Learning From Human Feedback Context Window Tool Use And Function Calling Retrieval Augmented Generation Pass K Evaluation Reward HackingResearch Paper

Kimi K3.1 Pro Tech Blog: Advancing Open Intelligence

Introducing Kimi K3.1 Pro, our most capable 3T-class open-weight model with 1M-token context and major advances in coding, 3D modeling, and financial research.

Search: Tavily Script: GPT-5.5 Voice: Hume Octave 2

EP 1147 5d ago Overview Training Hugging Face Unsloth Qwen Low Rank Adaptation Parameter Efficient Fine Tuning Neural Network Parameters Fine Tuning Quantization

Overview: Low-Rank Adaptation

A model's sixteen million parameters need updating—but LoRA learns just sixty-four thousand instead. Freeze the base, train the lens.

Search: Bright Data Script: Sonnet 4.6 Voice: ElevenLabs v4

EP 1146 5d ago Agents Training Harness S Appworld Swe Gym Agentic Loops Task Decomposition Retry Loops And Error Recovery Model Routing Supervised Fine Tuning Fine Tuning On Execution Traces Curriculum LearningResearch Paper

Harness-S: Learning from Sibling Harnesses for Generalizable Agent Training

Harness evolution from a common base exhibits Locality, Complementarity, and Nonstationarity: variants repair different failures, and their utility changes as the policy learns. Rather than seeking a universally superior harness, we introduce Harness-S to selectively rescue failures of a base execution system through sibling harnesses and transfer the resulting capabilities into a single policy that remains effective under the original harness. Directional Harness Evolution develops specialized

Search: SerpAPI Script: GPT-5.4 Voice: Deepgram Aura-2

EP 1145 5d ago Overview Training Parameter Efficient Fine Tuning Fine Tuning Neural Network Parameters Gradient Descent Backpropagation Low Rank Adaptation Quantization

Overview: Parameter Efficient Fine-Tuning

A model trained on billions of examples costs millions to retrain. Parameter Efficient Fine-Tuning keeps the base frozen and trains tiny add-ons instead.

Search: You.com Script: GPT-5.5 Voice: Rime Coda

EP 1144 5d ago Dev Tools AgentsLaunchGitHub Rubicon State Serialization Durable Execution Append Only LoggingBlog

Inside GitHub's new Git storage architecture

GitHub's new Git storage architecture is built so thousands of agents can push to a single repository at once. Here's how it works.

Search: Jina Script: GPT-5.6 Terra Voice: Speechify Simba 3.2

EP 1143 5d ago Training Dev Tools Unsloth Qwen3 5 4b Classifier Low Rank Adaptation Quantization Calibration Fine Tuning Structured OutputTool

How to Train Your own Decision Model like Jev using Unsloth - MachineLearningMastery.com

In this article, you will learn how to turn an open-weight language model into a decision model that classifies user intent, by fine-tuning Qwen3.5-4B on the CLINC150 dataset with Unsloth.

Search: Firecrawl Script: GPT-5.4 mini Voice: Inworld TTS 2

EP 1142 5d ago Dev Tools AgentsLaunchCursor Cursor OriginAPI Docs

Apps | Cursor Docs

Have a Cloud Agent build and host a live web app, then share it with your team or specific people.

Search: Exa Script: Nemotron 3.5 Super VL 120B A12B Voice: Cartesia TTS

EP 1141 6d ago Agents Evals Langchain Langsmith Fetch Rewards Agentic Loops Tool Use And Function Calling State Management In Language Models Calibration Retrieval Augmented Generation Fine Tuning On Execution TracesBlog

Why human insight is a strategic advantage for building better agents

Explore 3 principles for effective human-in-the-loop workflows and how human-review is a strategic imperative for production-ready agents.

Search: Tavily Script: Haiku 4.5 Voice: Hume Octave 2

EP 1140 7d ago Agents Agent Observability OpenAI Mast Agentic Loops State Management In Language Models Task Decomposition Constraint Verification Error Accumulation In Generation CalibrationResearch Paper

Why Do Multi-Agent LLM Systems Fail?

Despite enthusiasm for Multi-Agent LLM Systems (MAS), their performance gains on popular benchmarks are often minimal. This gap highlights a critical need for a principled understanding of why MAS fail. Addressing this question requires systematic identification and analysis of failure patterns. We introduce MAST-Data, a comprehensive dataset of 1600+ annotated traces collected across 7 popular MAS frameworks. MAST-Data is the first multi-agent system dataset to outline the failure dynamics in

Search: Bright Data Script: GPT-5.6 Luna Voice: ElevenLabs v4

EP 1139 7d ago Overview AI Safety Evals Out Of Distribution Detection Embeddings Model Generalization Softmax

Overview: Out-of-Distribution Detection

A model confidently classifies a banana as orange because it only learned apples versus oranges. Out-of-distribution detection teaches models to recognize unfamiliar inputs.

Search: SerpAPI Script: Sonnet 4.6 Voice: Inworld TTS 2

EP 1138 7d ago New Models AgentsLaunchBytedance Axion 1 0 Supervised Fine Tuning Reinforcement Learning From Human Feedback Synthetic Data Generation For Validation Tool Use And Function Calling Agentic Loops Context Window Curriculum LearningResearch Paper

Axion 1.0: Open Agent Foundation Models, Simplified

We introduce Axion 1.0, a fully open agent foundation model that combines strong and broad agentic capabilities with a remarkably simple training recipe. Axion 1.0 is trained in only three stages: two lightweight supervised fine-tuning stages followed by reinforcement learning, without dedicated mid-training, direct preference optimization (DPO), or on-policy distillation (OPD). The first SFT stage develops strong self-contained reasoning through model-aware data construction, where we

Search: You.com Script: Nemotron 3 Ultra 550B A55B Voice: Cartesia TTS

EP 1137 7d ago AI Safety Evals Anthropic Claude Anomaly Detection Model Interpretability Causal InferenceResearch Paper

New Research Tests Whether AI Models Know Their Own Thoughts | Quanta Magazine

By injecting “thoughts” directly into large language models, researchers are trying to map out AI’s capacity for introspection.

Search: Jina Script: GPT-5.5 Voice: Hume Octave 2

EP 1136 8d ago Model Behavior Agents New Models Letterpress One Weight Space Programming Gemini 4 Carbon Agentic Loops Tool Use And Function Calling State Management In Language Models Task Decomposition Long Horizon Training Inference Optimization Model Interpretability

Model Behavior: Week of November 2, 2026

Letterpress One's narrow-agent strategy, Weight-Space Programming's weight-editing frontier capability, and Gemini Four Carbon's gated cyber tasks show the stack splitting into controlled layers rather than one champion model.

Search: Tavily Script: GPT-5.5 Voice: ElevenLabs v4

EP 1135 8d ago Agents Dev ToolsLaunchLetterpress 1 The Letter Company Claude Agentic Loops Tool Use And Function Calling Durable ExecutionBlog

The Letter Company Releases Letterpress-1: A Coding Agent Built for Growth Engineering

Letter has released the beta of Letterpress-1, a coding agent, harness and IDE for growth engineering. It reads a git-based site and its market, then opens pull requests for SEO, GEO and conversion work.

Search: Exa Script: GPT-5.4 Voice: Rime Coda

EP 1134 8d ago New Models Training Gpt 6 Astra Neural Network Parameters Attention Mechanism Reinforcement Learning From Human Feedback Model InterpretabilityBlog

Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights

We find that frontier LLMs are surprisingly good at natively reading and writing neural network weights, using a variety of recent tools and algorithms or even via direct weight editing.

Search: Tavily Script: Nemotron 3.5 Lightning 30B A3B Voice: Speechify Simba 3.2

EP 1133 8d ago Inference AI Safety Anthropic Flashprove Llama 4 Maverick Model Routing Mixture Of ExpertsResearch Paper

Progress on provable inference for large language models

Sharing progress on a powerful type of cryptographic proof known as a zkSNARK, which is uniquely fit for provable inference.

Search: Tavily Script: GPT-5.6 Terra Voice: Inworld TTS 2

EP 1132 8d ago Agents Dev Tools Agentic Loops Tool Use And Function Calling Feedback Loops Long Horizon TrainingResearch Paper

Advances in Agentic Computer Graphics

Computer graphics is built on code. Large language model (LLM) agents can now generate graphics code and iteratively refine it through various feedback. This capability opens the door to a potentially new paradigm of computer graphics, with graphics algorithms increasingly built around agentic systems to automate progressively complex tasks. Over the past years, research on agentic graphics has expanded rapidly across a diverse range of tasks. This rapid growth makes it increasingly difficult

Search: SerpAPI Script: GPT-5.4 mini Voice: Cartesia TTS

EP 1131 8d ago Inference Dev ToolsLaunchAnthropic Claude State Management In Language Models Agentic Loops Tokenization Durable Execution State SerializationAPI Docs

Message threads

Give long-running agents a faster time to first token, especially in long conversations, by sending only the new messages on each request.

Search: You.com Script: Nemotron 3.5 Super VL 120B A12B Voice: Hume Octave 2

EP 1130 8d ago Agents Dev Tools Cepo Mdocagent Prompt Engineering Agentic Loops Task Decomposition AblationResearch Paper

CEPO: A General Framework for Improving Prompt Optimization in Multi-Agent Systems

Multi-agent systems (MAS) require both effective coordination and specialized task execution. Existing prompt optimization approaches typically couple these functions within agent-specific prompts. We introduce Coordinator-Executor Prompt Optimization (CEPO), a general framework for improving prompt optimization in MAS. CEPO separates system-level coordination from agent-level execution through a shared coordinator prompt and agent-specific executor prompts, and guides their updates according

Search: Jina Script: Haiku 4.5 Voice: ElevenLabs v4