爱窝啦 AI 日报 AI Daily 2026/10/24: DeepSeek Search API Opens, Baidu Stock Drops Over 5%, AI Talent Bidding War Ends, Million-Yuan Salary Era Closes

Today’s Summary

DeepSeek's search API has officially opened, with its hundreds-of-billions-scale web index posing a direct challenge to Baidu, whose stock fell over 5% in a single day.
AI talent premiums are narrowing, Agent open-sourcing is accelerating, and search infrastructure is opening up to developers — the industry is shifting from cash-burning expansion to deployment competition.
Start today by checking DeepSeek's search pricing and the momoding open-source repo before deciding whether to replace your existing search or Agent module.

🔥 Today’s Top 10 Focus Stories

1. DeepSeek Search API Opens, Baidu Stock Falls Over 5%

The search competitive landscape has shifted abruptly. According to 36Kr, DeepSeek’s API now supports web search and crawling , priced at ¥5 per 1,000 successful requests. Official documentation states the index covers hundreds of billions of web pages globally, with filtering mechanisms targeting SEO manipulation and low-quality AIGC content. Developers can plug in directly today to replace existing search modules.

2. AI Talent Bidding War Ends, Million-Yuan Salary Era Closes

AI hiring has shifted from hoarding talent to selecting it. LatePost interviewed 24 researchers, headhunters, and HR professionals at major tech companies, documenting the shift in large-model companies’ hiring logic . Some companies used to hire first and find roles later; this autumn recruitment season, that approach has noticeably tightened. Narrowing salary premiums is a signal that AI professionals currently in salary negotiations will need to face directly.

3. Volcano Engine Seedance 2.5 4K and Canva Craft Launch on the Same Day

Two platforms each bet on high-quality image generation. According to AIBase, Volcano Engine’s Seedance 2.5 4K and Canva Craft launched on the same day . Seedance generates confirmed output in 10-bit color depth; Canva Craft uses a two-stage architecture and supports 2K resolution and ultra-wide aspect ratios. Seedance 2.5 4K is limited to select enterprise trials, while Canva Craft ranks fourth on evaluation leaderboards, still trailing top-tier models in quality.

4. Fudan Proposes Embodied Self-Evolution Framework, Enabling Robots to Autonomously Accumulate Skills

Robots no longer depend on pre-collected data. Synced (量子位) reports that Fudan has proposed an embodied self-evolution framework that lets robots autonomously accumulate skills after deployment 🤖, bypassing the traditional bottleneck that required data collection before embodied tasks could begin. This is currently at the research demonstration stage.

5. The Tian Keyu Incident Reconstructed: Disrupted Training and the Race for Talent

A personnel conflict reflects the cost of the industry’s rapid expansion. LatePost published an in-depth report reconstructing the full story of the Tian Keyu incident , touching on how large-model companies handle the development of young researchers and competitive pressure during rapid scaling. It contrasts with the concurrent signal of cooling AI talent markets; specific details require reading the original article.

6. Guicang Tests AI App Recognition Performance on iPhone Duo

First on-the-ground records of AI recognition on new hardware are out. Jike user Guicang recorded and published a video of AI app recognition performance on the iPhone Duo — a first-hand, on-site record. Suitable for users evaluating the AI capabilities of the iPhone Duo to reference directly, without waiting for second-hand media coverage.

7. Developer Builds Mobile Personal Agent in Their Spare Time and Open-Sources It

Mobile Agents are starting to handle real everyday tasks. Jike user 糯.米.鸡 fully open-sourced the mobile Agent project momoding , built on OpenMinis. It can automatically clean large files, book high-speed rail tickets on 12306, and set alarms based on pickup-code SMS messages. The interaction design draws from Grok Bot and Instinct. Developers who want to experience an Agent getting things done on their phone can head straight to the repository.

8. Anthropic Key Employee Thariq Reportedly Departed

Departure reported for a prominent Agent advocate with significant community influence. According to a tip relayed by Guicang on Jike , Thariq has long represented Anthropic in publishing Skill and AgentSDK content and context engineering tutorials, and is considered one of the community’s primary communicators alongside Boris Cherny. The information comes from social media secondhand; it has not yet been confirmed by Anthropic officially. Developers tracking the Claude ecosystem can watch for further official statements.

9. AI Media Clickbait Problem Called Out with a Real Example

Low-quality AI-written headlines are eroding reader trust. Jike user glgjssy posted a screenshot highlighting a clickbait example from a leading AI media outlet , with an article headlined “Just Now — The Global Internet Has Been Conquered by AI! The Darkest 48 Hours in History Is Here.” This type of AI-generated clickbait article continues to circulate on traffic-driven platforms. For professionals who rely on AI news to make decisions, evaluating source quality has never been more important.

10. ChatGPT Desktop App Reportedly Plagued by Heavy Disk-Write Bug

The disk-write issue in OpenAI’s desktop client has gone unfixed for several weeks. V2EX user crysislinux documented the latest trigger condition : setting pet to none causes the problem to stop immediately, as the app continuously writes over ten MB per second at startup. The issue was first reported on October 17; OpenAI has not yet released a fix. Users currently on the desktop version can use this setting as a workaround.

⚡ Product & Feature Updates

WorkBuddy Has a Problematic Default Config for Long Tasks — Fixable in Minutes

The default configuration causes long tasks to cut off midway. Juejin author 程序员晓凡 compiled the most common pitfalls in WorkBuddy’s initial setup : continuing to run after screen lock is off by default, the default storage path points to a low-space location on the C drive, and the reply style leans overly formal. These can all be adjusted within a few minutes of installation, reducing the chance of tasks failing partway through without warning.

◎ Industry Changes & Personal Impact

AI Is Already Displacing Entry-Level Roles — Companies Must Arrange Training Ahead of Time

AI encroachment on entry-level positions is happening now, not in the future. The Telegram channel AI Exploration Guide relayed a discussion on the pace at which companies are using AI to replace human labor : companies cannot notify workers today and let them go tomorrow — advance notice and transition training are needed. This view comes from community secondhand accounts, and execution will vary by company. For professionals, assessing how replaceable your own role is right now is more practical than waiting to be notified.

⌘ Top Open-Source Projects

NVIDIA/OpenShell: Secure Runtime for Autonomous AI Agents

NVIDIA has open-sourced an isolated runtime environment purpose-built for Agents. OpenShell is written in Rust and provides a secure, private runtime for autonomous AI agents , gaining 1,119 stars today and reaching a cumulative total of 18,926 stars. Suitable for developers who need to deploy Agents in local or private environments with data security requirements.

SnailSploit/Claude-Red: Attack and Defense Research Library Targeting the Claude Skill System

A security research toolkit specifically designed to target the Claude Skill system. Claude-Red packages each attack technique in SKILL.md format , covering attack surfaces including SQL injection, EDR evasion, and shellcode, gaining 226 stars today and reaching a cumulative total of 8,613 stars. Intended as a reference for researchers conducting AI system security testing in compliant, authorized environments. Do not use against unauthorized targets.

◉ Social Media Highlights

Google Gemini Integrated into the Cursor, Ask Questions by Wiggling

Gemini is embedded directly into the mouse pointer interaction flow. The official Google Gemini account posted a demo video of Magic Pointer on Googlebook : on the new Googlebook, wiggle the cursor and click to ask Gemini questions, compare content across screens, or create on the spot 🖱️ — no app-switching required. Check Google’s official announcements for availability and launch dates.

Cohere Rerank 5 Launches, RAG Retrieval Token Usage Cut by Up to One-Third

A new, more token-efficient option for document retrieval reranking. Developer Gorden Sun relayed that Cohere’s official blog introduced Rerank 5 in two tiers — Pro and Fast , with a new Context Focus feature that automatically highlights the most critical sentences and passes only the core evidence to the large model. According to Cohere, this reduces token consumption by up to one-third or more, with accuracy improving in tandem. Available via the Cohere API or major cloud platforms.

😄 AI Curiosities

The Most Advanced AI Paired with the Most Nostalgic UI

Lanxi used Codex to write a Chrome extension that skins Doubao and ChatGPT with a fully restored QQ 2007 interface . All 139 personalized avatars are back, and when the model finishes a reply it plays the iconic “di-di-di” chime — audible even if you’ve switched to another window. The real kicker: it also supports level-up progression, gaining a level every 30 minutes, with stars, moons, and suns unlocking new skins in sequence. In 2026, GPT waits for your quantum mechanics questions inside the UI of 2007.