BestBlogs Issue #78: The Infinite Brain

Zhipu AI open-sources GLM-4.7 with SOTA coding performance. Scores 73.8% on SWE-bench and 84.9% on LiveCodeBench V6, surpassing GPT-5.2 in Code Arena blind tests. Three new thinking modes improve complex task stability. API and full source code now available.

MiniMax M2.1 achieves SOTA multi-language programming performance, excelling in code generation, mobile development, and office automation. Benchmark results show it outperforms Claude Sonnet 4.5, with the newly open-sourced VIBE full-stack application benchmark. Ideal for developers focused on AI coding and agent applications.

Fun-Audio-Chat 8B is an end-to-end voice dialogue model that bypasses traditional ASR+LLM+TTS pipelines for lower latency. Key features include emotional awareness and Speech Function Call support for completing complex tasks via natural voice. Weights and code are fully open-sourced—worth checking out for voice interaction developers.

Based on Berkeley's survey and DeepMind experiments, reveals the reality of production Agent systems: 68% limited to 10 steps, multi-Agent systems face coordination tax and error amplification, and increased budget doesn't linearly improve performance. Real breakthroughs require advances in tool management, validation capabilities, and communication protocols.

A 2025 masterclass summary by Prof. Hung-yi Lee on SLM evolution. Key highlights: the trade-offs between cascade and end-to-end models, leveraging LLMs to fix speech semantic gaps, and how TASTE/STITCH architectures enable "think-while-speak" reasoning with zero latency. Ideal for grasping the logic behind GPT-4o's voice mode.

Samir Mody, AI Engineering Lead at The Browser Company, discusses their journey from the Arc browser to the AI-native browser Dia. Key insights include optimizing tools and processes for rapid iteration, viewing model behavior as a craft, and addressing AI safety as an emergent product attribute. Mody elaborates on how their engineering culture, prototyping strategies, and team structure evolved to tackle the challenges of building an interface that reasons, plans, and acts. He highlights the importance of internal tools for rapid prototyping and evaluation, introduces JEPA for automated prompt optimization, and underscores the necessity of embedding AI safety measures like user confirmation for sensitive actions to mitigate prompt injection risks, citing Dia's autofill, scheduling, and email functions as examples. The talk concludes by stressing the imperative for companies to fully embrace technological shifts.

Spring AI's Tool Argument Augmenter enables LLMs to "articulate" their reasoning during tool calls. By dynamically injecting additional fields (like reasoning steps and confidence levels), developers can capture complete decision-making logic without modifying tool code. This significantly enhances AI agent explainability and observability, supporting intelligent systems with long-term memory and self-reflection capabilities.

This article provides a profound retrospective of a real-world enterprise-level frontend AI Agent project, moving from "technical success" to "product failure" and eventually finding a breakthrough through a mindset shift. The author details the technical journey of building a prototype with the Claude Agent SDK, tackling challenges like training on private component libraries, implementing local file systems, and establishing automated quality loops. The core insight lies in the reflection on the "Agent Island" phenomenon: technical feasibility does not guarantee product adoption. The author advocates for shifting focus from building standalone Agents to encapsulating "Skills" that integrate into existing developer workflows (e.g., Cursor or Claude Code).

This article thoroughly discusses two core concepts in the AI Agent field: Model Context Protocol (MCP) and Agent Skills. It highlights that MCP primarily addresses the connectivity between Agents and external tools/resources, while Agent Skills focus on encapsulating domain knowledge and operational procedures, thereby equipping Agents with the knowledge of how to effectively use tools. A key innovation is the “progressive disclosure” mechanism embedded in Agent Skills. Through a three-tier architecture—metadata, skill body, and additional resources—information is loaded on demand, significantly alleviating context explosion and high costs often associated with overly large tool JSON Schemas in traditional MCPs. The article emphasizes that MCP and Skills are complementary, not competitive, and proposes a layered hybrid architecture that integrates both to optimize costs, improve maintainability, and enhance reusability. Finally, it details the SKILL.md specification and principles for writing high-quality Skills, while anticipating future trends such as industry standardization, ecosystem building, and automated capability discovery. It also cautions against potential risks like security vulnerabilities and fragmentation.

This year-end review of RAG technology's development in 2025 highlights its irreplaceable status as a data infrastructure in enterprise-level AI implementations, despite ongoing questions about long context windows and attention shifting to Agents. The article details RAG's technical advancements, including decoupling retrieval into distinct "search" and "retrieve" stages, and integrating TreeRAG and GraphRAG to address the limitations of traditional RAG in complex queries. It emphasizes RAG's trend of evolving from a mere knowledge base to a general-purpose data foundation for Agents, introducing the concept of "context engineering" and analyzing the critical role of domain knowledge, tool data, and conversational state data in Agent context assembly. Ultimately, the article forecasts RAG's future as a "Context Engine" or "Context Platform." Concurrently, it examines the progress of Multimodal RAG and its associated engineering challenges.

Trello's Head of Product shares 10 lessons from a decade of AI development. The article highlights overcoming the initial "negative value" phase by building trust through explainable UI. Key takeaways: PMs must master AI debugging and "Evals," adjust routing logic to prevent user gaming, and leverage GenAI to move beyond "one-size-fits-all" features toward hyper-personalized user experiences.

LangChain's latest survey reveals that 57% of organizations have moved AI Agents into production as of 2026. Customer service and data analysis dominate the landscape. The primary bottleneck has shifted from cost to output quality. Key trends include universal adoption of observability, a shift toward multi-model architectures, and the continued dominance of coding agents in daily workflows.

A curated list of 11 indispensable AI tools for 2025, ranging from foundational models (Claude, Perplexity) to automation workflows (Zapier, n8n) and specialized apps (Notion, Gamma, HeyGen). The core focus is on leveraging AI Agent browsers and integrations to eliminate repetitive tasks. Bottom line: Don't waste money on random subscriptions; use these 11 tools to systematize your workflow first.

96-generation entrepreneur Zhai Xingji explains how to restructure enterprise sales processes with AI Agents, focusing on charging by incremental value rather than labor costs. Through automated pre-sales solution generation, they helped clients achieve a 20% conversion rate increase. Emphasizes choosing core business pain points, ensuring technical feasibility, and the need for rapid iteration and decisive decision-making in entrepreneurship.

This report from Lenny and Figma’s AI Insights Manager surveys 1,750 pros to reveal AI’s real ROI: 50%+ save at least half a day weekly. Key takeaways: Founders benefit most, while Designers see the least gain. Engineers are migrating from GitHub Copilot to Cursor and Claude Code, while PMs are using AI to cross into prototyping. The findings suggest the next frontier for AI lies in moving from "output generation" to "strategic exploration" and "thought partnership."

Notion founder Ivan Zhao offers a historical lens on AI transformation, comparing it to steam and steel as the new "infinite brain." The piece analyzes AI's impact across three levels: individual, organizational, and economic. At the individual level, AI agents transform programmers from "10x engineers" to "30-40x engineers," though broader adoption requires solving context fragmentation and verifiability challenges. Organizationally, AI acts like steel reinforcing structure and steam powering production, enabling companies to scale beyond traditional communication bottlenecks. Economically, knowledge work will evolve from "human-scale" Florence to "megacity-scale" Tokyo, creating cross-timezone, high-density operations. Core insight: we must stop treating AI as merely a "copilot" and reimagine how work fundamentally operates.

A rare dialogue featuring Gemini's three co-technical leads (Jeff Dean, Oriol Vinyals, Noam Shazeer) reveals the technical philosophy behind Gemini 3. Key insights include: Flash models now match or exceed previous Pro performance, with Pro's primary role becoming Flash distillation; post-training is identified as the largest breakthrough opportunity; latency and speed are severely undervalued, often more important than absolute intelligence in practice.

This profile examines how Josh Woodward doubled Gemini's user base in eight months. Key takeaways: leveraging small (5-7 person) teams for rapid iteration, using the "block" system to bypass bureaucracy, and shifting AI's role from a search engine to a "content container." It also highlights his vision for "Dynamic Views," where AI moves beyond chat boxes to generate real-time interactive interfaces.

This year-end conversation is essential listening for understanding the 2025 global AI competitive landscape. Guangmi presents a critical thesis: this isn't an AI bubble, but an AI war—an arms race that tech giants and nations cannot afford to lose. The podcast deeply analyzes the competition between NVIDIA's GPU and Google's TPU ecosystems, revealing why OpenAI, Anthropic, and Google alternate in leadership positions. Most importantly, they predict Online Learning will become the third paradigm-shifting breakthrough after pre-training and reinforcement learning, proposing an investment portfolio shift from concentration to diversification. For practitioners tracking AI investments and technical trends, this podcast delivers rare frontline insights.

Y Combinator partners reflect on unexpected AI shifts in 2025. Data reveals YC startups now prefer Anthropic over OpenAI, particularly for code-related tasks. More significantly, startups are building orchestration layers to "arbitrage" models, dynamically selecting the best performer for each task. The discussion covers stabilization signals in the AI economy, long-term value of infrastructure investments, and why "one-person unicorns" remain unrealized.