Mike Rowe: “We’ve been telling kids for 15 years to learn to code.” “Well, AI is coming for the coders.” “It’s not coming for the welders, the plumbers, the steamfitters, the pipefitters, the HVAC, or the electricians.” “In Aspen, I sat and listened to Larry Fink say we need 500,000 electricians in the next couple of years—not hyperbole.” “The BlueForge Alliance, who oversees our maritime industrial base—that’s 15,000 individual companies who are collectively charged with building and delivering nuclear-powered subs to the Navy … calls and says, we’re having a hell of a time finding tradespeople. Can you help?” “I said, I don’t know, man … how many do you need? He says, 140,000.” “These are our submarines. Things go hypersonic, a little sideways with China, Taiwan, our aircraft carriers are no longer the point of the spear. They’re vulnerable.” “Our submarines matter, and these guys have a pinch point because they can’t find welders and electricians to get them built.” “The automotive industry needs 80,000 collision repair and technicians.” “Energy, I don’t even know what the number is, I hear 300,000, I hear 500,000.” “There is a clear and present freakout going on right now. I’ve heard from six governors in the last six months. I’ve heard from the heads of major companies.”

PSA for people using amp. Huge if you use amp on mobile a lot like me! https://t.co/gMlI5bIBdR

Announcing hameval, made by @randal_olson It's a meme site 🤣 have fun! https://t.co/oQ1KWAjspT https://t.co/TxAscaqhJS

Love this transcript project by @noamseg AdviceArena: When product legends give contradictory advice, which wisdom wins? https://t.co/Ivim5bged1 https://t.co/eJG5jqHfjy

Woah! https://t.co/LMgMpT5PRj

@pamelafox It’s here , point Claude code at it and tell it to install the tools and guide you through using it etc https://t.co/NghIsGO0dG

@pamelafox Also @isaac_flath has a SaaS app for this https://t.co/qwkXEqHpzj and it’s really good if you want to jump right into that

It's easy to get lost when reading traces. An effective heuristic is to stop at the first (most upstream) error you find and move on when doing error analysis. You can then modify and curate traces for evals. BTW this is flashcard 9/12 (links in reply). https://t.co/4pPnCig4ez

This MCP is so dam useful I have conversations with friends over iMessage all the time. Its amazing to be able to yeet this into a coding agent or claude. Thanks to @mattt for making it https://t.co/ouHHg7QGQ6 https://t.co/esnAgEgy8V

There are three categories of evals - and we've found people get confused between these! The trick is to try to use code based assertions whenever you can get away with it (even by being a bit clever). https://t.co/BWcKNRfVVw

Is Agentic RAG worth it? RAG systems have evolved from simple retriever-generator pipelines to sophisticated workflows. It remains unclear when to use Enhanced RAG (fixed pipelines with dedicated modules) versus Agentic RAG (LLM orchestrates the entire process dynamically). This research provides the first empirical comparison. Enhanced RAG adds pre-defined components to address specific weaknesses: routers to determine if retrieval is needed, query rewriters to improve alignment, and rerankers to refine document selection. The workflow is fixed and manually engineered. Agentic RAG takes a different approach. The LLM decides which actions to perform, when to perform them, and whether to iterate. No extra components beyond the basic knowledge base, retriever, and generator. The model controls everything. The researchers evaluated both paradigms across four dimensions on QA and information retrieval tasks. User intent handling: Agentic slightly outperforms Enhanced on most tasks, but Enhanced wins decisively on FEVER (+28.8 F1 points), where the agent often retrieves unnecessarily. Query rewriting: Agentic RAG achieves 55.6 average NDCG@10 compared to 52.8 for Enhanced, showing the agent can adaptively rewrite queries when beneficial. Document refinement: Enhanced RAG with reranking (49.5 NDCG@10) outperforms Agentic (43.9). Dedicated reranker modules beat iterative retrieval attempts. Agentic RAG is far more sensitive to model capability. With weaker models, Enhanced RAG maintains stability while Agentic performance degrades significantly. Cost analysis reveals Agentic RAG requires 2-10x more computation time and tokens due to multi-step reasoning. The choice between Enhanced and Agentic RAG depends on your constraints. Enhanced RAG offers predictability, lower costs, and stability with weaker models. Agentic RAG provides flexibility but requires stronger models and more compute. Paper: https://t.co/FqlGngacqo Learn to build effective Agentic RAG systems in our academy: https://t.co/OTrW4oI8ze

Introducing ralph-research plugin. I just adopted the ralph-loop for implementing papers. Mindblown how good this works already. The entire plugin was one-shotted by Claude Code, but it can already code AI paper concepts and run experiments in a self-improving loop. Wild! https://t.co/jPFD9RzCae

What an impressive journey @daniel_dhawan, from sleeping on a friend’s couch to building the #1 AI app builder within a span of 2 years - @rork_app A @fdotinc & @a16z portfolio company https://t.co/jmtiHY9kyT

Building in 2026 https://t.co/w8K4JRzqhD

Building in 2026 https://t.co/w8K4JRzqhD

trying a ralph loop to write my rag ebook with more material https://t.co/cYSStTWa3m

one is mapped to cursor composer one is mapped to push to transcribe one is mapped to start a loom video https://t.co/J6MRlyzWqb

👏🏻Congratulations！Step3-VL-10B was selected for HuggingFace Daily Papers today. Check it out here：https://t.co/zro7vQUqzh https://t.co/VVxMVGL0P9

👏🏻Congratulations！Step3-VL-10B was selected for HuggingFace Daily Papers today. Check it out here：https://t.co/zro7vQUqzh https://t.co/VVxMVGL0P9

V-DPM 4D Video Reconstruction with Dynamic Point Maps https://t.co/TdBRHDzrsd

discuss: https://t.co/DTXKcIwk84

app: https://t.co/roiCHYhEx6

HeartMuLa A Family of Open Sourced Music Foundation Models https://t.co/TJzg6eMEXZ

paper: https://t.co/2ALsoeOEji

model: https://t.co/W7HamtcMzs

STEP3-VL-10B Technical Report https://t.co/TiSlflEB58

SeedFold Scaling Biomolecular Structure Prediction https://t.co/XFVd625BEW

Transition Matching Distillation for Fast Video Generation https://t.co/YXFty6ul0W

discuss: https://t.co/YWnc9zFtNO

RigMo Unifying Rig and Motion Learning for Generative Animation https://t.co/FkionIZZLr

discuss: https://t.co/doYeEa9MO3

What makes preference data truly effective for LLM alignment? 🤔 Introducing AIR: A systematic framework that deconstructs preference datasets into 3 core components (Annotations, Instructions, Response Pairs) and reveals evidence-based optimization principles. No more trial and error! 🤗 Paper: https://t.co/dFxRHSux2W 📄 arXiv: https://t.co/QJl1T1uGyi Why it matters: 1️⃣ Simplicity Wins in Annotations: Basic point-wise scoring with generative models (like Llama-3.1-70B-Instruct) + greedy decoding outperforms complex methods. Less is more—excessive design introduces noise rather than clarity. 2️⃣ Smart Instruction Filtering: Select instructions with low response variance across LLMs. This forces models to learn fine-grained preferences (like logical rigor) rather than relying on obvious differences. 3️⃣ Balanced Response Pairs: Optimal pairs combine moderate score gaps (Δ=2-3), high absolute quality (≥8), and 1:1 On/Off-Policy mixing—achieving clear contrast without overfitting. The results? +5.3 average gain across 6 benchmarks (WildBench, Arenahard, etc) with just 14k curated pairs from 17 open-source LLMs covering coding, math, and chat tasks. AIR transforms preference learning from "scale blindly" to component-aware design—a blueprint for building smarter, more aligned AI systems. 🚀 #AI #LLM #RLHF #PreferenceLearning #Alignment