AI Daily News: Apple CEO transition, Musk’s $60 B Cursor acquisition, new AI models and market trends (April 22 2026)

Apple leadership change

Apple announced that CEO Tim Cook will become Executive Chairman of the Board on September 1 2026, while John Ternus, the current Senior Vice President of Hardware Engineering, will assume the CEO role immediately. The transition table shows Cook moving to Chairman (effective September 1 2026) and Ternus taking over as CEO now.

Background of Ternus

Nearly 25 years at Apple, leading hardware since 2013.

Oversaw design of iPhone, iPad, Apple Watch, and Mac.

Hardware division contributes roughly 80 % of Apple’s revenue.

Internally praised as “the person who best understands Apple’s product soul.”

Cook’s legacy

CEO from 2011 to 2026 (15 years).

Market cap grew from $350 billion to $3.5 trillion (10×).

Service revenue rose from tens of billions to nearly $100 billion annually.

Led launch of Apple Watch, AirPods, and other new categories.

Market reaction Apple’s stock rose 2 % after the announcement, with investors confident Ternus can continue hardware innovation.

Elon Musk’s $60 billion acquisition of Cursor

SpaceX disclosed a major AI move: a proposed $600 billion purchase of the AI‑coding startup Cursor, or a $100 billion strategic partnership.

Deal options

Option A: Full acquisition for $600 billion.

Option B: $100 billion strategic cooperation.

Cursor background

Founded two years ago, valuation jumped from $1 billion to $10 billion.

Core product: AI‑code editor built on VS Code.

Monthly active developers exceed 5 million.

Annual revenue surpasses $200 million.

Founder’s view 25‑year‑old co‑founder Michael Truell said the deal is “a crucial step toward building the best AI programming platform,” noting Musk’s understanding of AI’s disruptive impact on software development.

Strategic intent Musk plans to deeply integrate Cursor with xAI’s Grok model, creating an “AI‑native” development environment that directly challenges GitHub Copilot and OpenAI’s Codex.

Kimi K2.6 open‑source release

The “Moon’s Dark Side” team released Kimi K2.6 today, fully open‑sourced with all technical details disclosed.

Core capabilities

Long‑context handling up to 2 million tokens, enabling whole‑book analysis.

Multimodal support for images and video (10‑minute video understanding).

Code generation in 100+ languages, achieving 94.5 % on HumanEval.

Reasoning on math/science tasks with 96.2 % on GSM8K.

Open‑source scope

Full model weights.

Training code framework.

Technical paper already posted on arXiv.

Founder activity Yang Zhilin appeared at the April 10 Premier’s forum and was featured again on “News Broadcast,” highlighting his prominence in China’s AI sector.

OpenAI launches GPT‑5.4‑Cyber for cybersecurity

OpenAI released GPT‑5.4‑Cyber, a large language model fine‑tuned for defensive cybersecurity scenarios.

Model characteristics

Positioned for defensive use.

Trained on vulnerability reports, penetration‑test logs, and malware samples.

Core abilities: vulnerability analysis, threat detection, security recommendations.

Security design reduces false‑positive rejections for legitimate security work.

Release method

Initially limited to vetted security vendors.

Requires signing a security‑use agreement.

Prohibited for offensive cyber activities.

Industry background Following concerns about Anthropic’s Mythos model, OpenAI emphasizes a defensive stance to balance capability and safety.

Anthropic’s Claude Opus 4.7 and Mythos controversy

Anthropic unveiled Claude Opus 4.7, boosting software‑engineering performance, while its Mythos model sparked international security worries.

Claude Opus 4.7 upgrades

Code generation accuracy rose from 92.1 % (v4.5) to 95.3 % (+3.2 %).

Debugging accuracy improved from 88.5 % to 93.7 % (+5.2 %).

Architecture design accuracy increased from 85.2 % to 91.8 % (+6.6 %).

Context window doubled from 1 million to 2 million tokens (+100 %).

Mythos concerns

Designed for security testing with “red‑team” capabilities.

Can simulate attacker thinking to uncover system flaws.

Governments of Japan, Australia, and others are closely monitoring.

Japan’s finance minister plans meetings with financial institutions to discuss risk controls.

Anthropic response The company states Mythos is intended solely for defensive security testing, requiring approval and logging for any use.

22‑year‑old reverse‑engineers Anthropic’s Mythos

A 22‑year‑old independent developer posted a GitHub project claiming to have reverse‑engineered Anthropic’s Mythos architecture and released it publicly.

Technical details

Mixture‑of‑Experts routing: 64 experts, top‑4 activated (partially verified).

Attention: hybrid of multi‑head and sparse attention (speculative).

Parameters: ~500 B total, ~50 B active (unconfirmed).

Training: RLHF with safety‑constraint optimization (speculative).

Community reaction

Project stars exceeded 10 k within 24 hours.

Anthropic has not commented; GitHub received a DMCA takedown request.

Developer claims use is “for academic research only, no commercial purpose.”

Legal risk Reverse‑engineering may breach Anthropic’s terms of service, and the developer’s EU location adds jurisdictional complexity.

Alibaba opens Qwen 3.6‑27B

Alibaba Cloud released Qwen 3.6‑27B, a 27‑billion‑parameter dense multimodal model.

Performance comparison

SWE‑bench Verified: 77.2 vs 72.5 (previous 397 B MoE model) → +4.7 %.

GPQA Diamond: 87.8 vs 85.3 → +2.5 %.

MMLU: 88.5 vs 86.7 → +1.8 %.

Code ability: 94.2 % vs 91.8 % → +2.4 %.

Key breakthrough

27 B dense model outperforms 397 B MoE model, showing architectural efficiency gains.

Comprehensive programming capability surpasses prior flagship models.

Inference cost reduced by 60 %, making it suitable for SMB deployment.

License Apache 2.0, permitting commercial use.

Alibaba’s “Qianwen Xiao Jiuwō” AI digital assistant

Alibaba unveiled the “Qianwen Xiao Jiuwō” digital human, positioned as an ecosystem‑level AI assistant.

Core capabilities

Voice ordering and dish recommendation for food‑delivery apps (e.g., Ele.me).

Travel itinerary planning and ticket booking (e.g., Fliggy, Amap).

Product recommendation, price comparison, and ordering for shopping (e.g., Taobao, Tmall).

Bill inquiry, transfers, and financial advice via Alipay.

Schedule management, meeting notes, and document handling for office (e.g., DingTalk).

Technical traits

Cross‑app service stitching, no need to switch apps.

Voice‑first interaction with dialect recognition.

Emotionally expressive avatar that can be customized.

Launch status Available on the Qianwen app for both iOS and Android.

Tesla China integrates Doubao and DeepSeek models

Tesla China announced that its in‑car voice AI service, now approved by Shanghai regulators, will integrate ByteDance’s Doubao large model and DeepSeek Chat.

Integration details

Doubao (ByteDance) – voice interaction and content recommendation, accessed via Volcano Engine.

DeepSeek Chat (DeepSeek) – complex Q&A and reasoning, also via Volcano Engine.

Feature upgrades

Multi‑turn voice conversations.

Intelligent route recommendation combining real‑time traffic and user preferences.

Personalized in‑car entertainment recommendations.

Voice‑based vehicle fault diagnosis.

Privacy handling Voice data processed locally; only de‑identified data uploaded for model improvement.

SAIC‑Motor’s “JiaYue” AI‑native vehicle series

SAIC‑Motor (Roewe) launched the “JiaYue” AI‑native vehicle line, featuring three concept cars (JiaYue 06, 07, 09) at the Shanghai Auto Show.

AI capabilities

Voice interaction powered by Doubao large model (via Volcano Engine).

Smart driving via Momenta’s City NOA system.

Multimodal cabin AI from Volcano Engine.

OTA updates through Alibaba Cloud for continuous evolution.

Model lineup

JiaYue 06 – compact SUV, ¥150‑200k, launch Q3 2026.

JiaYue 07 – midsize sedan, ¥200‑300k, launch Q4 2026.

JiaYue 09 – midsize‑large SUV, ¥300‑400k, launch Q1 2027.

Strategic significance First traditional automaker to fully adopt ByteDance’s AI capabilities.

ByteDance 2025 financial dip due to AI investment

ByteDance’s 2025 financial report shows net profit falling over 70 % year‑over‑year, primarily because of massive AI spending.

Key figures

Net profit: $40 billion (2024) → $12 billion (2025) – ‑70 %.

Revenue: $150 billion → $180 billion (+20 %).

AI R&D spend: $10 billion → $35 billion (+250 %).

Investment focus

Large‑model training (Seed team).

Compute procurement (GPU/TPU).

Global talent acquisition of top AI researchers.

Business highlights

Overseas revenue up ~50 % YoY.

TikTok Shop GMV up ~70 %.

Advertising revenue continues to grow.

Investor reaction Private‑equity valuation trimmed from $400 billion to $300 billion, though long‑term investors remain bullish on AI transformation.

Beijing AI model registration surge

Beijing’s Economic and Information Bureau reported 225 AI models registered locally, representing about 30 % of the nation’s total.

Breakdown

General models: 45 (20 %) – examples include Baidu Wenxin, Alibaba Tongyi, Zhipu GLM.

Industry‑specific models: 120 (53 %) – covering finance, healthcare, education, legal.

Vertical models: 60 (27 %) – targeting code, design, marketing, customer service.

Industry dynamics

Zhipu AI listed on Hong Kong main board, becoming the world’s first AGI‑base model company.

ByteDance released Seedance 2.0 video‑generation model, topping multiple benchmarks.

Baidu’s Wenxin Yiyan daily active users surpassed 100 million.

Alibaba’s Tongyi Qianwen API calls exceed 10 billion per day.

Policy support Beijing established a ¥100 billion AI industry fund, focusing on large models, compute, and data annotation.