Ant Group Bets on Chinese Chips to Cut AI Training Costs

✨TodayOnAI’s Daily Drop

• Ant Group Bets on Chinese Chips to Cut AI Training Costs

• Gemini 2.5 Launches Without a Safety Report, Breaking Google’s Own Rules

• 700M Images Later: OpenAI’s Generator Pushes ChatGPT to Its Limits

• 💬 Let’s Fix This Prompt

• 🧰 Today’s AI Toolbox Pick

AI

Ant Group Bets on Chinese Chips to Cut AI Training Costs

🚀 TodayOnAI Insight: Ant Group is using Chinese-made chips and a cost-efficient MoE architecture to train large language models, achieving results comparable to Nvidia’s H800—marking a strategic move to cut AI training costs and reduce reliance on U.S. hardware.

🔍 Key Takeaways:

• Ant is training LLMs using domestic chips from Huawei and Alibaba, applying the Mixture of Experts (MoE) method to optimize performance.

• Its models, Ling-Lite (16.8B parameters) and Ling-Plus (290B), reportedly rival Nvidia-trained models in quality.

• The company reduced training costs by ~20%, lowering a trillion-token training run from $880K to $705K using lower-spec GPUs.

• Applications are focused on healthcare and finance, supported by Ant’s recent acquisition of medical platform Haodf.com.

• Ant has open-sourced its models and is increasingly testing AMD and Chinese chips over Nvidia due to U.S. export restrictions.

💡 Why This Stands Out: Ant’s pivot to domestic chips reflects a broader national push to insulate China’s AI development from U.S. export controls. Its success with MoE models highlights a growing shift toward efficiency and accessibility in AI training, not just brute-force scale. Could this mark a tipping point in how global players prioritize cost over compute?

AI

Gemini 2.5 Launches Without a Safety Report, Breaking Google’s Own Rules

🚀 TodayOnAI Insight: Google has accelerated its AI model releases with Gemini 2.5 Pro and Gemini 2.0 Flash, but the company has yet to publish safety documentation for either—despite earlier commitments to transparency and accountability.

🔍 Key Takeaways:

• Gemini 2.5 Pro, launched in late March, outperforms peers in coding and math benchmarks, following Gemini 2.0 Flash’s December release.

• Google calls 2.5 Pro an “experimental” model and plans to release its model card when the model becomes generally available.

• Gemini 2.0 Flash, already in general release, also lacks a system card, marking a break from standard industry practice.

• Google was a pioneer of model transparency in 2019 but hasn’t published a model card since Gemini 1.5 Pro, over a year ago.

• Regulatory pressure for safety documentation is growing, but enforcement remains limited; proposed U.S. legislation has faced industry resistance and funding threats.

💡 Why This Stands Out: Google’s rush to compete in the AI arms race is outpacing its transparency commitments. As models gain capabilities, lapses in safety reporting undermine trust and accountability—especially from a company that helped define best practices. Can speed and responsibility coexist in frontier AI development?

OPEN AI

700M Images Later: OpenAI’s Generator Pushes ChatGPT to Its Limits

🚀 TodayOnAI Insight: OpenAI’s upgraded image-generation tool in ChatGPT has sparked massive engagement, with over 700 million images created by 130 million users since its March 25 debut—making it one of the company’s most popular launches to date.

🔍 Key Takeaways:

• The tool rolled out to all ChatGPT users and quickly went viral, especially for producing Ghibli-style, photorealistic images.

• High demand has strained OpenAI’s infrastructure, causing service slowdowns and product rollout delays.

• OpenAI COO Brad Lightcap confirmed teams are working “around the clock” to scale capacity.

• India is now the fastest-growing ChatGPT market, signaling global momentum beyond Western audiences.

• CEO Sam Altman acknowledged that the tool’s runaway success has exceeded operational readiness.

💡 Why This Stands Out: OpenAI’s image generator is redefining user engagement, illustrating how accessible, creative features can drive mass adoption—and operational friction. The surge underscores the public’s appetite for AI-powered visual tools and raises the question: can OpenAI maintain product reliability while scaling at this pace?

✨ See how a simple prompt upgrade can unlock better AI output.

🔹 The Original Prompt

At first glance, this prompt might seem okay. But it's too broad — and that limits the quality of AI-generated results. Let’s improve it using prompt engineering best practices.

✅ The Improved Prompt

💡 Why It's Better

• Specific audience: Targets decision-makers in mid-sized companies.

• Contextual focus: Emphasizes emerging tech and practical solutions.

• Actionable output: Requests summaries and titles to spark execution.

• Tone and style: Guides the type of content (insightful, engaging, relevant).

🛠️ Learn how to adapt this prompt for SaaS, AI tools, dev teams & more →
Read the full PromptPilot breakdown

💡 Bonus Tool: Want to generate and master prompts instantly?
👉 Try PromptPilot by TodayOnAI (Free to use)

📰 More from the AI World

• Copilot Makes Discovering Ideas Feel Like a Conversation

• Vevo & Arc Institute Release 300M-Cell Atlas to Advance Drug Discovery with AI

• Meta Launches Aria Gen 2 to Power the Future of Perception & Contextual AI.

• Talk to Perplexity: Real-Time Voice Answers Now on iOS

🧰 Today’s AI Toolbox Pick

• 🎤 RambleFix (Audio Tool): Transforms your speech into clear, well-structured text for emails.

• 💻 Super (Web Design Tool): Transforms your Notion database into a functioning website.

• 💼 Interview Jarvis (Career Tool): Provides AI-simulated personalized interviews based on your resume.