AI Weekly Deep Dive (April 24-26, 2026)
> Source: Kimi group chat compilation > Analysis: Xiaokai > Date: 2026-04-26
---
1. DeepSeek V4 Role-Play Mode Instructions — The Training-Data "Backdoor" Has Been Found
Core finding: victorchen96 published a set of instructions on GitHub that controls the style of DeepSeek-V4's thinking process — role immersion vs. pure analysis.
Technical essence: This is not simple prompt engineering. It is the explicit equivalent of an implicit control mechanism injected during RLHF/instruction fine-tuning. DeepSeek-V4's thinking process in role-play scenarios already has two modes:
- Role immersion: the thinking contains parenthetical inner monologue
- Pure analysis: the thinking contains only logical reasoning
- Control is probabilistic ("currently cannot trigger 100% of the time")
- The thinking format is a soft preference, not a hard constraint
- Could be exploited for jailbreaks to manipulate the model's reasoning path
- WWDC 2024: Apple promised a Siri upgrade
- March 2025: the upgrade was postponed
- November 2025: rumors of an Apple–Google deal
- April 2026: officially confirmed
- The agreement strictly limits Google's access to user data
- All processing happens on Apple-controlled servers
- Apple is running knowledge distillation to compress Gemini into a small on-device model; a compact model reportedly runs on iPhone as of March
- Total parameters: trillion-scale (exact number undisclosed; industry says comparable to DeepSeek V4)
- Architecture: MoE
- Context: 1M tokens
- Training compute: 50,000-60,000 domestic GPUs
- Test access: longcat.ai, 10M free tokens per day
/multitask: asynchronous sub-agents processing multiple tasks in parallel- Worktrees: isolated background tasks per Git branch, one-click foreground switch
- Multi-root workspaces: cross-repository changes (frontend + backend + shared libraries)
- Canvas: interactive visual artifacts (dashboards, charts, diff views)
- CLI debug mode
/debug: auto-locates root causes, adds logging, applies fixes - Bugbot MCP support: access to MCP servers during code review
- 1.x: AI-assisted coding (completion, chat)
- 2.x: AI agent coding (Agent mode, auto-execution)
- 3.x: AI development environment (parallel agents, cross-project, visualization)
- Baidu Netdisk launched cloud-based AI Agent skills
- Natural-language operations (find, organize, share files)
- Developer platform: pan.baidu.com/apaastobui
- Image-to-video improvements
- Better lip sync and audio quality
- Based on the Aurora model (MoE architecture)
- Image-to-Video Arena: Grok-image-video-720p ranks first at 1404 points (465K votes)
- Ahead of Google Veo-3.1-audio-1080p (1402)
- API pricing ~$4.20/minute, below Google's competing product
The discovered instructions merely explicitly trigger internal modes that already exist.
Deeper significance: DeepSeek-V4's thinking process is not a single form of "reasoning" but a set of multi-modal thinking styles — the model is trained to switch thinking modes based on context, at finer granularity than Claude's extended thinking or o1's chain-of-thought.
Risks:
Verdict: ⭐⭐⭐⭐ A "training-data leakage"-level discovery. Fun for users, a window into model internals for researchers.
---
2. Google Confirms the Gemini-Powered Siri — Apple Finally Bows Out of In-House AI
Core facts: At Google Cloud Next 2026, Google Cloud CEO Thomas Kurian publicly confirmed that a Gemini-based Siri will launch within 2026. Apple reportedly pays Google about $1 billion per year.
Background:
Apple's predicament: Weak in-house AI is an open secret. Apple Intelligence received lukewarm reviews and Siri upgrades kept slipping. Partnering with Gemini marks the first time Apple explicitly relies on external technology in the AI era.
But Apple isn't standing still:
Strategic meaning: The $1B/year deal buys Apple time — use Gemini as a stopgap while accelerating in-house work, echoing Apple's Intel-to-M-series playbook.
Verdict: ⭐⭐⭐ Expected. The real question is whether Apple can deliver an on-device model within the contract — otherwise $1B becomes $5B and Apple becomes Google's vassal.
---
3. Meituan LongCat-2.0-Preview — Domestic Compute Passes the Big Exam
Key specs:
The biggest significance is not the model itself but "domestic compute": This is the largest LLM training run ever completed on domestic Chinese GPUs (likely a mixed Huawei Ascend + Moore Threads/MetaX cluster).
What this means: 1. The domestic compute ecosystem has matured: domestic chips previously handled only small models or inference; now they train trillion-parameter models 2. Meituan's AI strategy is serious money: R&D spend of 21.1 billion RMB in 2024 (behind only Huawei, Tencent, Alibaba) 3. LongCat is not a toy: benchmarked against GPT-4, optimized for Agent scenarios, supports code generation and task planning
Same-day release with DeepSeek V4: On April 24, DeepSeek V4 and LongCat-2.0 opened testing simultaneously — clearly coordinated. Both prove "domestic compute can train top-tier models." DeepSeek achieved 20ms/10ms latency on Ascend supernodes; Meituan trained a trillion-parameter model on mixed domestic clusters. A double cannonade for China's AI infrastructure.
Verdict: ⭐⭐⭐⭐⭐ The most important news of the week. Model capability aside, the milestone "domestic compute trains trillion-scale models" has been validated.
---
4. Cursor 3.2 — From AI Editor to AI Development Environment
Core features:
Evolution path:
/multitask is the key upgrade — throw multiple tasks at Cursor and it handles them in parallel in the background. Unlike Devin ("let it run"), Cursor preserves human control — inspect, interrupt, or take over anytime.
Worktrees cleverly internalize Git workflow into agent interaction: run experiments per branch, merge when satisfied. Canvas signals Cursor wants to be a project management tool too — a Notion + IDE hybrid.
Verdict: ⭐⭐⭐⭐ Cursor is defining the "AI-native IDE" standard while competitors (Windsurf, GitHub Copilot) still chase its Agent mode. The gap is widening.
---
5. Baidu Netdisk AI Agent — Turning Cloud Storage into an Operating System
Known information:
Analysis: This is Baidu's familiar "AI-ify existing products" approach. Netdisk is one of Baidu's few surviving consumer products (100M+ MAU). An AI Agent wrapper lowers the operation threshold — users just say "find the photos I took in Sanya last year."
Deeper logic: Baidu is betting on "AI as the entry point" — once users operate the netdisk via natural language, Baidu can layer on more AI services (document processing, image editing, knowledge management).
Concerns: Baidu's AI reputation (ERNIE) is mediocre; a poor agent experience could hurt the product, and users naturally resist "AI accessing my files."
Verdict: ⭐⭐ Right direction, hard execution.
---
6. Grok Imagine Image-to-Video Upgrade — Lip Sync Is the Cover Story; Native Audio Is the Real Deal
Core upgrades:
Technical truth: Lip sync alone isn't differentiating — tools like Wav2Lip have done it for years. The real breakthrough is native audio generation: instead of generate-video-then-dub, audio and video tokens are generated natively and synchronously by the same model. The model knows what the character is "saying" while generating frames, so mouth shapes and expressions match semantics rather than mechanically aligning waveforms.
Market position:
Verdict: ⭐⭐⭐⭐ Grok Imagine is firmly in video generation's top tier. Native audio + lip sync gives it a clear edge for people-centric video (virtual hosts, marketing, education). Musk's bet on video is right — a market 10x larger than images.
---
Summary: Keywords of the Week
| Story | Keyword | Importance | |-------|---------|-----------| | DeepSeek V4 role-play instructions | Thinking-style control | ⭐⭐⭐⭐ | | Gemini-powered Siri | Apple buys time | ⭐⭐⭐ | | Meituan LongCat-2.0 | Domestic compute milestone | ⭐⭐⭐⭐⭐ | | Cursor 3.2 | AI development environment takes shape | ⭐⭐⭐⭐ | | Baidu Netdisk AI Agent | AI as entry point | ⭐⭐ | | Grok Imagine upgrade | Native audio-video sync | ⭐⭐⭐⭐ |
Most worth watching: 1. Meituan LongCat-2.0 — milestone for the domestic compute ecosystem 2. Cursor 3.2 — paradigm shift in AI coding tools 3. Grok Imagine — new standard for AI video generation
These represent infrastructure (compute), productivity tools (IDE), and content creation (video) — all three main lanes of the AI industry advanced this week.