Deep Dives
In-depth AI analysis, industry observations, and technical insights — slow down and read.
QQ Bot Burning Through Tokens Too Fast? Engineering Layering and Cost-Saving Tips
QQ bot burning tokens too fast? Bilibili creator Da Fei Yu shares an engineering layering solution: separate deep and shallow memory, switch to coding plans, and add basic RAG—estimated to save 10x to 100x tokens.
AstrBot Plugin in Practice: Advanced Optimization Strategies to Drastically Cut Token Consumption
A Bilibili creator shares advanced AstrBot plugin optimizations: context window compression, shallow/deep memory separation, on-demand story packs, group chat read frequency control, and non-reasoning judgment mode to drastically cut token consumption.
LangChain 1.3 Systematic Tutorial: Hands-On RAG/Agent/MCP/LangGraph in Practice
Bilibili creator Loulan's LangChain 1.3 systematic tutorial covers architecture fundamentals, RAG, Agent, MCP, and LangGraph multi-agent projects, emphasizing systems and practice over tools and theory to build transferable AI development skills.
Cult.fit Uses Codex to Build Operation…
Cult.fit Uses Codex to Build Operational Tools, Accelerating Cross-Functional Teams
Cult.fit is using OpenAI Codex to help cross-functional teams build internal tools, speed up work pace, and change how they work. This article analyzes implementation scenarios, value, and governance boundaries of cross-functional AI programming.
Google Releases Gemini Omni 1.2 Flash: 15-Second AI Video at $0.07 per Second
Google releases Gemini Omni 1.2 Flash video generation model with native 15-second clips and multi-turn conversational editing, cutting 720p pricing 30% to $0.07 per second.
Ollama v0.40.15 Released: Adds Kimi Code CLI Support
Ollama v0.40.15 released, adding Kimi Code CLI support to cmd/launch, commit #19506 GPG-verified. This article covers the update's impact on AI coding workflows and version verification details.
OpenAI Recruits Dwarkesh to Design Evals? Benchmark Methodology Sparks Discussion
OpenAI staff publicly praised a benchmark's methodology and invited Dwarkesh Patel to design evals. This article explains why eval design has become a key link in AI competition and a new path for talent mobility.
Anthropic IPO at $1.5 Trillion? Why Major Investors Are Skeptical
Anthropic's IPO is being discussed at a $1.5 trillion valuation, but major investors are skeptical. This article analyzes the gap between valuation and fundamentals, AI lab revenue-cost structures, and what public market pricing logic signals for AI companies.
Nasdaq Erases $2.5 Trillion: The 'Blue Wave' Triggers AI Panic
Nasdaq market cap evaporates $2.5 trillion as the 'blue wave' political shift sparks market panic over AI regulation and policy outlook. This article analyzes the policy expectation changes behind the sell-off, the fragility of the AI narrative, and implications for investors and practitioners.
Franklin Templeton Rebuilds Investment…
Franklin Templeton Rebuilds Investment Research Workflow with OpenAI
Franklin Templeton is developing AI workflows based on OpenAI technology, starting with equity research to help portfolio managers compare research views and examine investment risks. This article analyzes its workflow approach, risk review boundaries, and signal significance for the asset management industry.
EnvOpt: A New Direction for Agent Environment Optimization, Cutting Costs by 72.7%
EnvOpt proposes a third direction for agent optimization: automated optimization of static environments. On AutomationBench, it cuts costs by 72.7% and improves pass rate by 14.34%, transfers across models and harnesses, and requires no graders or reference answers.
Sparse Autoencoders Reveal How Large Language Models Represent Grammatical Number
Research uses sparse autoencoders to analyze models like Pythia-70M, finding that plural noun and verb features causally control grammatical number output and revealing asymmetric singular-plural representations, offering a new method for explaining LLM grammatical mechanisms.
HyFlash: A New Prefix-Memory Paradigm for On-Device Speculative Decoding
HyFlash introduces a prefix-memory prior separating long-range context modeling from fine-grained draft-position interaction, achieving higher acceptance rates for on-device speculative decoding under limited memory with dramatically reduced KV storage and notable gains in shallow-draft and long-context scenarios.
D²-OPD: A New Method for Online Distillation that Decouples Direction and Distance
D²-OPD proposes a new method that decouples direction and distance in extrapolative distillation, using a per-prefix hard KL constraint to control target distance and adaptively combining teacher shift and initial gap to determine direction, achieving the best results across all three settings on math reasoning benchmarks, surpassing the strongest baseline by up to 9.45 points.
How Do AI Agents Waste Your Money? The Systemic Structure of Token Waste
A study proposes a unified framework to measure AI agent token waste via LLM-as-a-Judge and intervention validation, revealing system-level structure.
Archetypal Contexts in Language Model Neurons: An Architecture for Intelligence?
New research finds separable archetype, topic, and language neurons in the Qwen3.8-27B language model, with a spontaneously emerging sandwiched structure; archetypes may be the core architecture of intelligence and analogical reasoning. Code and data are open-sourced and reproducible.
Recursive Sparse Attention RecSA: Million-Token Context Extrapolation at Logarithmic Cost
RecSA recursive sparse attention uses a tree-structured KV cache for O(log L) selection cost, extrapolating from 32K training to 2M, scoring 60.5 on RULER at 1M and 3.2× faster prefill on a single H200.
SCF: A New Method for Bounded Evidence Aggregation in Language Model User Inference
SCF (Source-balanced Capped Fisher) is a bounded evidence aggregation method for language model user inference that achieves the highest AUROC across multiple datasets through capped document contributions and calibrated decision rules, with finite-sample Type-I error control.
SAFE-Teen Study: Chatbot Safety Failure Rate Reaches 43% in Teen Mental Health Conversations
SAFE-Teen study evaluated 18 chatbots in teen mental health conversations, finding a 43.3% standard failure rate without teen-aware prompts, 24.9% with prompts, and a 10–12 percentage point increase in domain-specific failures when teens push back.
Safety Vulnerabilities in Vision-Language Models: Compositional Harm and Un-alignment Auditing
New research reveals vision-language models' fragile compositional safety: high benchmark scores don't mean robustness, and un-alignment method CMaG cuts safety scores by 65 points while keeping outputs coherent, calling for internal auditing.