Chunhua Liao chunhualiao

Chrome 146这个更新，杀死了AI浏览器 Chrome 146发布了，其中一个看似小的更新可能会改变很多AI Agent的工作方式。

以前用AI操作浏览器，要么开headless模式（无界面浏览器），要么得想办法把登录态移植过去。无论是哪种方式，都挺麻烦的——headless容易被检测，移植登录态要折腾Cookie和token。有人为此专门安装中继插件以便调用浏览器能力，而且还非常不稳定。

现在不一样了。Chrome 146原生支持MCP，只需在chrome://inspect里开一个开关（chrome://inspect/#remote-debugging），你的AI Agent就能直接控制当前正在运行的浏览器会话。注意，是当前会话，不是新建一个。

开发者Petr Baudis展示了一个具体用法：他让Claude帮他清理LinkedIn上那些试图推销东西的连接请求。Claude打开了他的LinkedIn页面，分析了每个邀请，然后批量忽略。整个过程用的是Petr已经登录的会话，不需要任何额外认证。

Image

It is completely understandable why you are impressed with Telegram's bot integration. When you are managing agent gateways and dealing with local models, you want the connection to the messaging layer to be as frictionless as possible.

Here is a breakdown of why Telegram feels like magic compared to Discord or Signal, along with a few secure, developer-friendly alternatives that play nicely with OpenClaw.

Telegram designed its architecture to treat bots as first-class citizens almost from the beginning, whereas other platforms tacked them on later or prioritize different security models.

• The "BotFather" Paradigm: Instead of forcing you to navigate a clunky web developer portal, Telegram lets you create a bot by simply messaging another bot (@BotFather). You send a few chat commands, and it instantly hands you an API token.
• Simple HTTP API: Telegram's Bot API is remarkably straightforward. It uses simple JSON over HTTP (via webhooks or long polling) without requir

The gap you identified—that current AI is a "jagged" statistical alien rather than a robust, survival-oriented entity—is arguably the central problem in modern AI research. To make AI more "holistic" (moving it from a narrow spike in intelligence space toward the robust "animal" cluster), we cannot simply make LLMs larger. We must change the architecture and the optimization pressures to mimic the factors that created animal intelligence. Here is the engineering roadmap to bridging that gap:

1. From "Next Token" to "System 2" Reasoning The "jagged frontier" exists because LLMs operate almost entirely on "System 1" (fast, intuitive pattern matching). If the pattern isn't in the data, they hallucinate. To fix this, we are moving toward Neuro-Symbolic and Verifier-based architectures.

• The Fix: Instead of one model predicting the next word, we use a "Generator-Verifier" loop.
  • Generator (The Artist): Proposes a solution (e.g., code, math, logic).
• Verifier (The Critic): A separate model trained specifi

The official doc at https://docs.devin.ai/work-with-devin/deepwiki-mcp is incomplete, letting readers wonder about how exactly to use it.

Found an example from https://platform.openai.com/docs/guides/tools-connectors-mcp

It seems like one just put the github repo id (such as accountID/repoID or rose-compiler/rose-full) into the question, then deepwiki will answer question about that repo.

from openai import OpenAI
import os
import sys

Yes, there are AI models and tools that provide an email interface, allowing you to interact with them by sending emails (including attachments) to ask questions and receive responses via email. These tools leverage advanced language models and email automation to enable seamless communication. Below, I outline some options based on available information, focusing on tools that support sending emails with attachments and receiving AI-generated responses.

1. EmailTree.ai
   • Description: EmailTree.ai is an AI-powered email automation platform designed for customer service and productivity. It supports intelligent email handling, including processing attachments and generating contextually relevant responses.

• Email Interface: You can send emails to the platform, and it uses AI to analyze the content, including attachments (via OCR for scanned documents), to provide automated responses. It can identify questions, extract data from attachments, and respo

sdaf

I spent today experimenting with OpenEvolve (the open-source counterpart to DeepMind’s AlphaEvolve) on a RHEL 9 laptop using the OpenAI API. To get it running, I fixed a small hashing issue—swapping hashlib.md5 for hashlib.sha256 due to Red Hat constraints—and added support for gpt-5, modeling it after the other o-series entries.

For a test task, I used the classic Two Sum Python problem (scan a list and find two numbers that add up to a target). With the default settings, OpenEvolve failed to reach the optimal solution and declared the O(n²) brute-force approach as the best.

Digging into the logs (and checking with AI) revealed that evaluator.py sets the optimization objective. The defaults overweight correctness and rely on overly simple tests. I expanded the test suite with edge cases and larger inputs, and rebalanced the objective to 50/50 for correctness and efficiency (instead of 70/30). Last but not least, I enabled full rewrite (not the default diff-based ev

Here are five SF ecosystems that play a role similar to Cerebral Valley—active communities that run product-focused events, recruiting, and partner with major players: 1. South Park Commons (SoMa / South Park) A tight-knit builder community + “Founder Fellowship” that invests at the -1→0 stage and hosts frequent AI-centric demos/hackathons. Apply or track their public events.    2. Founders, Inc. (Fort Mason campus) A 42,000-sq-ft founder campus that runs recurring AI hackathons and demo days, and writes first checks. Follow their Luma calendar and program pages.   3. HF0 (Hacker Fellowship Zero) — Residency (Alamo Square / SF) Highly selective live-in residency for repeat founders; offers meaningful upfront funding and deep mentor access; widely covered for its intense builder culture.    4. AGI House (SF & Hillsborough) AI-focused hacker houses and community with referral-based SF outpost; supported/visited by notable industry figures and cloud partners; frequent AI events.   5. SHACK15 (

I don't have information about ChatGPT O3-Pro, as this appears to be a model that was released after my knowledge cutoff. Let me search for current information about it.​​​​​​​​​​​​​​​​

Based on my search results, here's how to get the most out of ChatGPT O3-Pro:

O3-Pro is OpenAI's most advanced reasoning model, released in June 2025. It's available exclusively to ChatGPT Pro users ($200/month) and Team users, replacing the previous o1-pro model.

This is a very common issue when deploying web applications, especially when moving from a development environment (like port 3000) to a production-like setup (like port 80). Here are the most common reasons why your son might be having trouble changing the port to 80, and how to resolve them:

1. Privileged Ports (Most Common Reason):

   • Explanation: On Linux/Unix-like operating systems (which most web servers run), ports below 1024 (like port 80 for HTTP and 443 for HTTPS) are considered "privileged ports." Only processes running with root (administrator) privileges can bind to these ports.
   • Why it's a problem: If your son's web application is running as a non-root user (which is good practice for security), it won't have the necessary permissions to listen on port 80 directly.

2. Port Already In Use:

• Explanation: Another service might already be running and listening on port 80. This could be another web server (like Nginx o