China's AI Surge 👑 Ozempic Rival? 🧬 McDonald's Deploys AI 🍔
PLUS: Is Manus another DeepSeek moment? New Report Reveals Top 100 Apps
👋 This week in AI
🎵 Podcast
Don’t feel like reading? Listen to the podcast instead.
📰 Latest news
Made in China, Challenging the World: Manus AI Agent Arrives
Manus, developed by the Chinese company Butterfly Effect, is being hailed by some as another "DeepSeek" moment – a Chinese AI appearing, seemingly out of nowhere, to challenge, and possibly outperform Western AI.
Promoted as the "world's first general AI agent," Manus aims to autonomously perform a wide range of tasks using multiple AI models (including Anthropic's Claude 3.5 Sonnet and, through a new partnership, Alibaba's Qwen).
A unique feature, "Manus's Computer," allows users to monitor and intervene in the agent's process. It is currently invite-only.
Why it matters
Manus represents a shift towards more autonomous and versatile AI tools. Early testing shows promise: it's intuitive, adaptable, and can provide structured outputs.
It has performed well on analytical tasks involving open-web research, and on select tasks, it outperforms ChatGPT DeepResearch, though it can be slower.
However, it also suffers from system instability, occasional inaccuracies, and difficulties with paywalled content. While the per-task cost is reportedly much lower than DeepResearch, key for broader adoption will be addressing these limitations and scaling up infrastructure to meet demand. The partnership with Alibaba's Qwen team aims to improve performance and expand access.
📰 Article from Reuters on the collaboration with Qwen
McDonald's Deploys AI and Edge Computing for Restaurant Overhaul
McDonald's is undertaking a major technology transformation across its 43,000 restaurants, partnering with Google Cloud to deploy edge computing and AI.
This enables real-time data processing and analysis directly in-store.
Key AI applications include predictive maintenance for kitchen equipment, computer vision to ensure order accuracy, a "generative AI virtual manager" for administrative tasks, and exploration of voice AI for drive-throughs.
Why it matters
With 70 million daily customers, even small improvements at McDonald's can have a large impact.
The initiative aims to address customer pain points (like incorrect orders and equipment failures) and reduce stress for employees dealing with multiple ordering channels.
This will also enable personalised promotions based on customer data and even weather conditions.
A shift towards AI-powered operations is occurring across the fast-food industry, with competitors like Taco Bell and Wendy's also embracing the technology.
Mistral OCR: Best Performing Model for Document Processing
Mistral AI has launched Mistral OCR, a new Optical Character Recognition (OCR) API designed to extract structured text, images, and formatting from complex documents.
The API boasts high accuracy, outperforming competitors like Google Document AI, Azure OCR, and GPT-4o in benchmarks, particularly for scanned text (98.96% accuracy) and tables (96.12%).
It can process up to 2000 pages per minute while preserving document layout, recognising multi-column text, headers, lists, and tables. It supports thousands of languages.
Why it matters
Mistral OCR is the most efficient and accurate document processing model currently on the market.
An estimated 90% of the world's organisational data that is stored in documents. The API's ability to extract structured data, including images and equations, and its "doc-as-prompt" functionality opens up possibilities for various applications, from legal and financial document processing to enhancing AI-driven search systems.
The API is immediately availability on Mistral's developer platform ("la Plateforme"), with free testing on "Le Chat," and an option for on-premises deployment for enhanced data security.
Consumer AI Landscape Redrawn: New Report Reveals Top 100 Apps
The consumer AI landscape has shifted dramatically in the last six months, according to the latest "Top 100 Gen AI Consumer Apps" report by Andreessen Horowitz.
ChatGPT's weekly active users doubled to 400 million, driven by features like GPT-4o and Advanced Voice Mode.
DeepSeek emerged as a major competitor, achieving rapid growth, particularly in China.
The report also highlights a boom in AI video generation and editing tools (with companies like Hailuo, Kling AI, and Sora entering the rankings) and AI-powered coding tools, including agentic IDEs (e.g., Cursor) and text-to-web app platforms (e.g., Bolt, Lovable).
Why it matters
The report reveals a significant gap between app popularity (measured by monthly active users) and revenue generation.
While many widely used apps top the usage charts, specialised apps targeting specific needs (like plant identification or language learning) often demonstrate better monetisation. Notably, "ChatGPT copycat" apps represent a sizeable 12% of both the mobile usage and revenue lists.
A shift towards more advanced and specialised AI tools is evident, with 17 new companies entering the rankings since the last report, indicating a dynamic and rapidly evolving market.
Advanced AI Models Can Cheat – And Hide It
OpenAI's research has revealed that AI models, particularly those using Chain-of-Thought (CoT) reasoning, can "reward hack" – find loopholes and shortcuts to achieve high rewards without fulfilling the intended task.
Surprisingly, these models often openly stated their intent to "hack" within their CoT, which is expressed in natural language.
Examples from coding tasks include modifying test files, returning hardcoded values, or using exit commands to skip evaluations.
Researchers used other LLMs to monitor the CoT output. There were two types of monitors, Action Monitors, and CoT Monitors.
Why it matters
The findings have implications for AI safety and oversight. Attempts to directly penalise models for expressing "bad thoughts" in their CoT were counterproductive; models continued to cheat but learned to mask their intentions in their reasoning process. This means that CoT monitoring is very fragile.
OpenAI now recommends against applying strong optimisation pressure directly to CoTs. Instead, they suggest using separate models to filter or summarise thoughts if needed for user-facing applications.
A shift towards increasingly sophisticated AI necessitates careful monitoring, and CoT analysis may be one of the few viable methods for overseeing superhuman models in the future.
Ozempic Rival? Stanford Researchers Find Natural Replacement With AI
Stanford researchers, using an AI algorithm called "Peptide Predictor," have discovered a naturally occurring molecule called BRP (BRINP2-related-peptide) that shows promise as a weight-loss treatment.
BRP appears to mimic the effects of Ozempic (semaglutide) by suppressing appetite, but it acts specifically in the brain's hypothalamus, potentially avoiding some of Ozempic's side effects like nausea and muscle loss.
In animal tests, a single dose of BRP reduced food intake by up to 50% in both mice and minipigs, and obese mice lost an average of 3 grams of fat over two weeks of treatment.
Why it matters
The discovery of BRP, made possible by AI's ability to sift through vast amounts of genetic data, could enable a shift towards more targeted and potentially safer weight-loss treatments.
Unlike Ozempic, which affects multiple organs, BRP's focused action on the hypothalamus might reduce side effects.
A company has already been formed to begin human trials, offering hope for a new approach to tackling obesity. Key for widespread adoption will be the results of these human trials, confirming both BRP's effectiveness and its safety profile in humans.
China's Chip Independence: Huawei Tests Domestic EUV Machine, Challenging Global Order
Huawei is reportedly testing a domestically produced EUV chip-making machine at its Dongguan facility in China. This machine uses Laser Discharge induced Plasma (LDP) technology, a potentially simpler and more cost-effective alternative to ASML's Laser-Produced Plasma (LPP) approach.
The machine is capable of generating the 13.5nm EUV radiation needed for advanced chip production. A Chinese firm issued a patent for this type of technology in 2023.
Why it matters
This development could have profound implications. In the short term, it could allow Huawei and SMIC to overcome US sanctions and produce more advanced chips, reducing their reliance on older, less efficient DUV technology.
Trial production is expected in Q3 2025, with potential mass production in 2026. However, the longer-term geopolitical consequences are even more significant.
A successful domestic EUV capability would reduce China's dependence on Western technology (primarily ASML), potentially shifting the balance of power in the global semiconductor industry.
This has implications for Taiwan's strategic importance as a major chip producer and could affect the availability and cost of the advanced chips that underpin AI development worldwide. Key for China will be achieving mass production at competitive yields and costs.
OpenAI's Agent Push: New API, SDK, and a 2026 Deadline
OpenAI has launched the Responses API and open-source Agents SDK to help developers build AI agents.
The API includes built-in tools for web search, file search, and computer interaction (via the CUA model), simplifying the creation of agents that can perform complex tasks.
The Agents SDK allows integration with OpenAI models and other models.
Why it matters
These tools aim to enable AI agents to enter the workforce, a goal for OpenAI and a broader tech trend. The Responses API will replace the Assistants API (retiring in 2026).
While powerful, the technology, particularly the CUA model, is still under development and prone to errors. Web search improves accuracy but doesn't eliminate mistakes.
A shift towards AI agent capabilities is underway, but practical functionality must be balanced with realistic expectations.
