Why China Is Winning the AI War

Good morning,

When I sit down in the evening to write this and I start scouring the interwebs for the latest and greatest AI news sometimes my brain feels like it’s about to melt.

There is so much mesmerizing new information every day I can easily write a daily newsletter.

Only, that would burn me out within a fortnight.

So much to focus on this week : did you know that LLMs now have officially passed the Turing test. The Turing test (as devised by Alan Turing) is that you basically have a conversation with an AI system and with a second conversation partner and if you can’t figure out it’s human or not by asking questions or from the conversation , the AI passes the test. The imitation game, it’s called.

It’s the premise of one of my favorite movies : “Ex Machina” (and also “The Imitation Game” obviouslyu”

So yes, back to the news of the week. Choices. We need to make some choices.

And the editorial decision for this newsletter is to try and zoom out a bit. Take a bit more of a helicopter view of what is going on and not go crazy with every new shiny thing that is announced.

That’s where we differ from other newsletters. You can trust us to spot the trends.

Topic of the week therefore is : China.

After the disruption (which even caused me to write an extra edition of the newsletter) they had caused when releasing DeepSeek at the end of January - they continue to crank out models that are better or on par with what the big AI studios in Silicon Valley have to offer.

China’s Strategy : Open Source EVERYTHING

I follow some extremely smart people on X and I came across a hefty tweet on this topic by Balaji Srinavasan.

His thesis is quite interesting.

He says we tend to forget about China in the West and it sometimes seems like there’s a billion people on this planet living in stealth mode. Yet, they’re quietly dominating the world and unleashing products unto the West that are not necessarily to our benefit (TikTok anyone?)

According to Balaji , China is unleashing an “open-source blitz”.

We’re talking high-quality AI models for language, vision, robotics, image generation—you name it. Free or dirt cheap. It all feels … strategic.

According to Balaji, China wants to “take the profit out of AI software.” Why? Because they make their money where they’ve always dominated: hardware. By flooding the world with free software, they devalue the Western AI industry and set up a world where Chinese-built devices—drones, robots, phones, cameras, sensors—are the vessels for global AI.

It’s the same move they made with manufacturing: copy, scale, undercut, dominate.
I see all these BYD cars zipping around and they get raving reviews from owners. They’re clearly coming for the market that Tesla is occupying and they’re doing better. It has to be said.

Back to AI, the release of the DeepSeek models a few months ago spooked markets so badly that around $1 trillion in US tech valuations got vaporized.

And DeepSeek isn’t exactly vaporware. The new release , Deepseekv3 is already pushing GPT-4.5 in benchmarks for reasoning, frontend coding, and tool use.

They’re also coming up with novel training approaches. One China-developed model, Jarvis VA, was trained to play Minecraft—not by copying human players, but by combining world knowledge, visual understanding, object recognition, and player trajectory learning. That multi-layered strategy led to a 40% performance boost over previous benchmarks.

China has the reputation of relentlessly copying everybody and everything , but they do seem capable of true innovation themselves.

It’s All About The Money

And they’re giving us all of that innovation for absolutely free.

And obviously there is no such thing as a free lunch. Nothing is truly free in this world.

So they give it for free while in the meanwhile OpenAI’s projected revenue model is $20K/month for an AI PhD-level agent, $10K/month for a developer-grade AI, $2K/month for a knowledge worker. That model only works if no one else offers 80% of the capability for free. But that’s exactly what’s happening.

Remember the leaked Google memo from 2023 ? The one that said, “We have no moat”?

This memo was a warning from people internally at Google to their bosses saying that independent developers where reaching similar or better results with open-source tools. We have no moat (by the way a moat is the big hole around a castle to keep out attackers). Without a moat - your enemy can just swarm your castle and finish you off.

So now the West faces a tough choice. Ban Chinese models? That risks global isolation. Compete on open source? That’ll take humility and speed. Stick to proprietary models? Maybe—but only if they offer serious value beyond what anyone can get for free.

This trend does not bode well for AI safety , it’s also going to heighten geopolitical tensions especially between the US and China and it is certainly bad news for your stock portfolio (as if it isn’t bad enough).

And on this cheerful note I wish you a warm welcome to The Blacklynx Brief.

AI News

• OpenAI is reportedly closing a record-breaking $40 billion funding round led by SoftBank, potentially doubling its valuation to $300B and marking the largest private investment in history. Despite losing $5B in 2024, OpenAI projects $12.7B in revenue next year and plans to be cash-flow positive by 2029. This funding will also help fuel the Stargate project, reinforcing OpenAI’s lead in the AI race despite rising global competition.

• Anthropic released two new research papers offering rare insight into how its AI model Claude thinks — revealing a shared “language of thought” across multiple languages and advanced planning capabilities like rhyme prediction in poetry. The papers also explain how Claude avoids hallucinations by only speculating when confident, helping shed light on internal model behavior. Understanding how these systems process information is increasingly crucial as AI grows more powerful.

• Alibaba’s Qwen team launched QVQ-Max, a new visual reasoning model that can solve complex tasks like blueprint analysis, geometry problems, and creative sketch evaluation. It builds on earlier models by incorporating adjustable “thinking time” to boost accuracy and shows strong performance in visual and mathematical reasoning. It’s Alibaba’s third major model release this week, further tightening the global AI race between China and the U.S.

• Elon Musk’s xAI has officially acquired social media platform X, forming a new entity called xAI Holdings worth over $100B. The move consolidates Musk’s AI and social media ventures, combining Grok’s advanced models with X’s massive data and distribution reach. It formalizes what’s long been an informal partnership and gives xAI a powerful edge in both training data and user access.

• A new book excerpt sheds fresh light on Sam Altman’s firing and return to OpenAI in 2023, revealing internal tensions, accusations of dishonesty, and ownership concerns tied to the OpenAI Startup Fund. While key figures like Mira Murati and Ilya Sutskever led the charge to oust Altman, employee backlash forced a dramatic reversal. With the full book releasing in May, it could reopen scrutiny — but OpenAI’s momentum may be too strong to derail.

• Apple is working on an AI-powered health coach under “Project Mulberry,” aiming to upgrade its Health app with personalized insights, food tracking, workout analysis, and expert-backed wellness content. Launching as “Health+” in 2026, it’s a long-term play that could make Apple’s wearables more proactive and intelligent. However, given Apple’s AI struggles so far, the pressure is on to finally deliver something breakthrough.

• Amazon just introduced Nova Act, an AI agent that can control web browsers to complete tasks on its own, outperforming top rivals like Claude 3.7 and OpenAI’s agents. Developers can now build web-based agents using Amazon’s new SDK — and the tech is set to power Alexa+, bringing advanced automation to millions of homes.

• Runway launched Gen-4, its latest video generation model built for professional creative use, delivering smoother motion, consistent scenes, and improved control over details. With adoption from major entertainment companies already underway, Gen-4 is positioning AI video as a serious tool for filmmakers and advertisers alike.

• UC Berkeley and UCSF scientists created a brain-computer interface that turns thoughts into speech in real time, with just a one-second delay and personalized voice reconstruction. This breakthrough could restore near-natural communication for people with speech loss from paralysis or neurological conditions.

• Dartmouth researchers completed the first clinical trial of an AI therapy chatbot, showing care on par with top human treatments. The chatbot, Therabot, helped reduce depression by 51% and anxiety by 31%, with users forming meaningful bonds and reporting high levels of trust and comfort during sessions.

• ChatGPT has reached 20 million paid subscribers, driving OpenAI to an estimated $5B annual revenue pace. The spike follows the viral release of GPT-4o’s image features and a $40B funding round, as OpenAI plans to open-source a new model for the first time in years.

• Tinder just launched an AI-powered flirting simulator called “The Game Game,” using ChatGPT tech to help users improve their dating skills. Players practice live conversations with virtual personas and get instant feedback — part of a growing trend of using AI not just for chatting, but for coaching real-life social skills.

• Researchers at UC San Diego found that GPT-4.5 can consistently pass the Turing test, fooling humans 73% of the time into thinking it’s a real person. The study revealed that casual conversation and emotional cues, not factual knowledge, were the key to deception — showing how difficult it's becoming to distinguish between AI and humans.

• Anthropic launched Claude for Education, a student-focused AI assistant designed to teach critical thinking instead of just giving answers. With a new “Learning Mode” and partnerships at major universities, the platform aims to become a trusted academic tool while encouraging ethical and thoughtful AI use on campus.

• Google DeepMind published a 145-page AGI safety strategy, warning that human-level AI could arrive by 2030 and pose major existential risks. The paper called out rivals' approaches and highlighted issues like deceptive AI behavior, proposing strict safety protocols — though getting global compliance remains a massive challenge.

Quickfire News

• OpenAI released an updated version of GPT-4o to paid users, featuring better prompt adherence, enhanced coding and creativity, and more model “freedom.”

• Butterfly Effect, the Chinese startup behind Manus AI, is seeking new funding at a $500M valuation as it struggles with high Claude API costs.

• OpenAI is delaying 4o image generation for free users and imposing rate limits due to overwhelming demand that is reportedly “melting” GPUs, according to Sam Altman.

• CoreWeave reduced its IPO target from $4B to $1.5B, with Nvidia stepping in as an anchor investor ahead of the company’s Nasdaq debut.

• Archetype AI introduced “Lenses,” a new physical AI application for its Newton model, designed to turn sensor data into actionable insights.

• PwC launched agent OS, a platform enabling AI agents to be integrated into enterprise workflows up to 10x faster than traditional methods.

• Lockheed Martin partnered with Google Public Sector to bring genAI tools into its AI Factory ecosystem, aiming to improve national security applications.

• Google released its Gemini 2.5 Pro Experimental model to all users, offering free access to the current No.1-ranked model on the LMArena leaderboard.

• OpenAI CEO Sam Altman hinted on X that the company may be developing a computer, replying they plan to make a “really cute one.”

• Some users spotted a new ‘thinking’ slider in ChatGPT, allowing the AI to auto-adapt, think a little, or think harder for more in-depth responses.

• Gemini 2.5 Pro Exp. scored a 130 on the Mensa Norway IQ test, the highest of any AI model, far above the average human score of 100.

• Baidu’s ERNIE 4.5 defeated OpenAI’s GPT-4.5 in multiple rounds of Chinese chess, winning decisively and reportedly “taking it easy” at times.

• Extropic AI revealed new details about its probabilistic computer chips, claiming up to 10,000x efficiency gains over traditional hardware, with plans to compete against Nvidia.

• OpenAI raised $40B from SoftBank and other investors at a $300B post-money valuation, marking the largest private funding round in history.

• Sam Altman announced that OpenAI will release its first open-weights model since GPT-2 in the coming months, with pre-release developer events to boost usability.

• OpenAI added 1 million users in an hour following the viral success of GPT-4o’s image generation, surpassing even the growth seen during ChatGPT’s launch.

• Manus launched a beta membership program and mobile app for its AI agent platform, with plans starting at $39 and $199/month depending on usage tiers.

• Luma Labs released Camera Motion Concepts for its Ray2 video model, letting users control camera movements using natural language commands.

• Apple pushed its iOS 18.4 update, bringing Apple Intelligence features to European iPhones, alongside visionOS 2.4 for the Vision Pro with new AI capabilities.

• Alphabet’s Isomorphic Labs, focused on AI drug discovery, raised $600M in a round led by Thrive Capital, also an OpenAI investor.

• Zhipu AI launched AutoGLM Rumination, a free AI agent for deep research and autonomous task execution, intensifying China’s AI agent competition.

• OpenAI rolled out GPT-4o’s image generation to free-tier users, expanding access to the viral tool across its entire user base.

• Meta’s VP of AI Research, Joelle Pineau, announced her departure after 8 years, leaving a key leadership vacancy at the head of the FAIR team.

• Alibaba is planning to launch Qwen 3, its next flagship AI model, later this month—following three model releases in just the past week.

• Sam Altman said OpenAI is experiencing GPU shortages, warning users of delays and slower service as the company works to secure more capacity.

• Meta researchers introduced MoCha, an AI model that creates realistic talking character animations from speech and text input.

• MiniMax released Speech-02, a new text-to-speech model delivering ultra-realistic voices in over 30 languages.

• Meta plans to launch $1000+ “Hypernova” AI smart glasses by year’s end, featuring a screen, hand-gesture controls, and a neural wristband.

• OpenAI released PaperBench, a new benchmark evaluating AI agents on replicating state-of-the-art research. Claude 3.5 Sonnet ranked highest in the tests.

• ByteDance, Alibaba, and other Chinese firms placed $16B in orders for Nvidia’s upgraded H20 chips, attempting to get ahead of U.S. export restrictions.

• Google appointed Josh Woodward, previously of Google Labs, as the new head of consumer AI apps, replacing Sissie Hsiao in leading the Gemini assistant.

• OpenAI announced an expert commission to advise its nonprofit, saying it will align financial power with technology that can scale human potential.

• The UFC and Meta formed a multiyear partnership to create AI-powered immersive experiences, including integration of Meta AI, AI glasses, and social tools.

Like newsletters?

Here are some newsletters our readers also enjoy:

Closing Thoughts

That’s it for us this week.

If you find any value from this newsletter, please pay it forward !

Thank you for being here !