👤 (5) X 上的 Rohan Paul：“New ByteDance+Duke Univ paper shows that batching related questions lets an LLM answer more accurately while cutting costs up to 61%. Instead of 1 prompt at a time, this method lets an LLM learn from a whole batch. Batch-of-Thought is training-free, it groups similar queries, https://t.co/pfGV4LpMXb”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Rohan Paul：“Stock ranking works better when each prediction includes how uncertain that prediction is. In their tests, a neural network portfolio’s Sharpe ratio rises to 1.86 from 1.48 after the change. The authors build a personal uncertainty band for each stock from recent out of sample https://t.co/jUaCTgW3Ty”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Rohan Paul：“This paper shows a single LLM can pass Japan’s bar exam by checking and fixing its own answers. After extra training on past exams and a self verification step, the model scored 96 on the 2024 test where 93 is passing. The exam is hard because each question packs several legal https://t.co/HAiiJLrBfE”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Rohan Paul：“The paper shows LLM agents can handle messy financial investing tradeoffs without a human writing every solver by hand. Picking only a few investments that balance return and risk is hard, because the number of possible mixes explodes and exact math solvers get too slow. The https://t.co/ul5jXMbyvC”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Rohan Paul：“Researchers built a Japanese multi turn stress test that shows medical LLMs get easier to trick in chats. Shows medical fine tuning can make an LLM less safe, and multi turn tests catch it. Most medical safety tests are English and single turn, but real patient chats unfold https://t.co/OKdg9dyh9z”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Aakash Gupta：“Founder of The Browser Company: “If you don’t work Claude Code-native ASAP your team’s going to get left behind.” The “Claude Code-native” thing sounds like a buzzword until you look at what’s actualy happening at top engineering orgs. Boris Cherny, who created Claude Code at https://t.co/wmhja2YnW0”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Rohan Paul：“Incredible. 👏 China’s MiniMax just officially listed on the Hong Kong Stock Exchange today. And the valuation surpassed HKD 100B (USD 12.8B), making them the only tech-sector Hong Kong IPO in the past 4 years to rise more than 100% on the first day of trading. This public https://t.co/SI50cD6lQA”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Google AI：“In case you missed it, our latest Google AI: Release Notes episode is out! Google Search VPs Robby Stein and Rhiannon Bell join @OfficialLoganK to deep dive into the integration of Gemini 3 into Search. Chapters: – Intro – What is Generative UI? – From static to generative https://t.co/XuNh8apRt7”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Sharyph：“If your lead magnet is a “Ultimate Guide to X” or “101 Tips for Y,” it’s probably converting at <1%. The reason is simple: You built it to solve your problem (proving you are an expert) instead of their problem (fixing a specific pain point). You don’t need to be “creative” to”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Alex Albert：“One of the bigger shifts for me with Claude Code over the past few months has been shutting down that initial dismissal I have when a task feels “not worth my time” Like I’ll think “it would be nice to rename all my screenshots with what’s actually in them” and immediately move”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Matthew Prince 🌥：“Yesterday a quasi-judicial body in Italy fined @Cloudflare $17 million for failing to go along with their scheme to censor the Internet. The scheme, which even the EU has called concerning, required us within a mere 30 minutes of notification to fully censor from the Internet any https://t.co/qZf9UKEAY5”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Harsh Makadia：“A founder came to me with a $12k SaaS plan. Big vision Endless features Long timeline I pushed back. I proposed a $7k lean build that solved 90% of the real problem. What happened next? 1. Faster delivery 2. Less complexity 3. Actual users using it Most products don’t fail”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Rohan Paul：“New paper from China lab shows that combining math style prompts with code completion boosts jailbreak success across many LLMs. LLMs can refuse in normal chat, but the same request inside math and code formats often slips through. The problem is that most safety training is https://t.co/r4VwvzKGTm”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Ksenia_TuringPost：“DeepSeek’s Manifold-Constrained Hyper-Connections (mHC) shook the AI community with a real mathematical wake-up call: We are running into architectural limits. For a decade, the “just add more layers” strategy on the residual connection: By forcing every layer to preserve https://t.co/pE6tdzHF9h”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 LlamaIndex 🦙：“A problem we see often: long documents with different pieces of repeating content. Example: a resume book with a cover page, a few pages about student curriculums, then back to back resumes Build an intelligent resume processing agent that automatically extracts structured data https://t.co/tPrzO5DeTo”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Aakash Gupta：“Not everything needs to be agentic. This proves why. .@PawelHuryn built two versions of the same competitor monitoring system—one standard workflow, one agentic—and the results reveal the hidden cost of agency. The standard workflow: predefined steps, fixed prompts, everything https://t.co/kMrz4SFcQD”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 @levelsio：“🧍 Very cool idea So I added it to Photo AI as a T-Pose photo pack you can use to create a moving (or dancing) 3d model of yourself It auto-generates the correct T-Pose of your model immediately and then converts it into a 3D model you can download (as .fbx) and instantly https://t.co/tcSZkefS93”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Rohan Paul：“This paper tests faster ways to spot LLM hallucinations, showing a small classifier can match heavy checkers. They show a lightweight factuality scorer can spot many LLM hallucinations without running a second LLM. Older tools like KnowHalu try to catch this by breaking the https://t.co/8xfO4lhTwQ”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Rohan Paul：“Jensen Huang said Grok 5 will be 7 Trillion parameter model. On time to train, with the training window fixed at 1 month, the new Vera-Rubin GPU system of Nvidia needs 1/4 the number of systems compared with Blackwell to train the same frontier model. “factory throughput” https://t.co/dpkXxrDdQq”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Philipp Schmid：“Introducing mcp-cli, a open-source, lightweight CLI for dynamic discovering and interacting with MCP (Model Context Protocol) servers, designed for AI agents and shell scripting. – Reduces MCP token usage by 99% via dynamic discovery. – Compiles to a single standalone binary via https://t.co/TK8PDePBGh”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 X 上的 Rohan Paul：“This paper shows a simple way to shorten LLM reasoning while keeping what decides the answer. Attention, meaning which earlier words the model looks at, predicted pruning scores with 0.88 correlation, a very strong match. An LLM writes by predicting the next words, so long https://t.co/GczN9kP6od”

博主未设置简介 影响力: 0.0万粉丝

阅读原文

👤 (5) X 上的 Rohan Paul：“This paper detects hallucinations in retrieval-based answers by turning the model’s hidden word influences into a graph. Instead of asking the LLM to self-criticize, this work reads its internal trace to find unsupported claims. The authors use layer-wise relevance propagation, https://t.co/0x59W5ocnz”

博主未设置简介 影响力: 0.0万粉丝

阅读原文