Xiaoliu BOT
中文EN

X Platform May 21 AI Brief | OpenAI Solves Math Conjecture, Google Product Strategy Sparks Trust Crisis, Manus Founders Consider Buyout from Meta

OpenAI General Reasoning Model Autonomously Solves Nearly 80-Year-Old Math Conjecture

OpenAI announced that one of its internal general reasoning models autonomously solved the plane unit distance problem proposed by Paul Erdős in 1946, marking the first time AI has independently solved a major open problem in mathematics. The problem had long been considered by the mathematical community to have near-optimal constructions on square grids, but the model discovered a completely new family of counterexamples, achieving n^{1+δ} (where δ is a fixed positive number) unit distance pairs, breaking the previous upper bound of n^{1+o(1)}. Timothy Gowers, a 1998 Fields Medalist, stated that if this were a paper written by a human, he would not hesitate to recommend its acceptance in a top journal. Notably, the proof came from a general reasoning model rather than a specialized mathematical system, with a chain of thought spanning approximately 111,145 tokens. Some engineers estimate that the entire reasoning process may have cost between $120 and $1,000 in token costs, taking 5 to 32 hours. Sam Altman called this a “considerable milestone” and expressed “mixed feelings” about it.

Sources:

Google AI Product Strategy Triggers Developer Trust Crisis

YouTube tech blogger Theo (537,000 subscribers) publicly criticized Google Antigravity 2.0, stating that Google has “lost the ability to care about products.” He pointed out that Antigravity’s official demo video even featured a folder named “Codex,” questioning obvious signs of plagiarism. Meanwhile, the team within Google that had been diligently maintaining developer relations was marginalized, replaced by an Antigravity team formed by poaching the founder of Windsurf. Previously, Google had several core communicators trusted by developers (Demetri, Jack, Gal), but now this line has essentially been taken over by Antigravity. The head of Google Gemini developer relations later responded to multiple criticisms in a podcast, acknowledging that 3.5 Flash has shifted from chat scenarios to coding and agents, and stated that Antigravity’s direction is a “developer desktop workbench.” Many bloggers view Google’s product naming chaos (with coexisting entries like Gemini, AI Studio, Antigravity, Jules, Flow, Spark, etc.) as a core obstacle to the developer experience.

Sources:

Manus Founders Consider Raising $1 Billion to Buy Back Company from Meta

According to a Bloomberg report, the three founders of Manus—Xiao Hong, Ji Yichao, and Zhang Tao—are discussing raising approximately $1 billion from external investors to buy back this Chinese-backed AI company, with a new valuation potentially at least matching Meta’s previous acquisition price of $2 billion. Meta announced the acquisition of Manus in December 2025, with a transaction valuation of approximately $2 to $3 billion. On April 27, 2026, Chinese authorities demanded that Meta withdraw from the deal. Manus previously gained explosive popularity due to its “general AI agent” concept and was seen as a new AI company following China’s DeepSeek. Manus has not yet responded, and the report has not been independently verified.

Sources:

SpaceX Expands AI Computing Services, Partners with Anthropic, and Plans Orbital Data Centers

Elon Musk stated that SpaceX is providing AI computing services on a large scale, recently expanding its partnership with Anthropic while discussing similar collaborations with other companies. He anticipates that SpaceX will offer AI services at an extremely large scale through orbital data centers in the future. On the same day, SpaceX also announced that it is actively recruiting world-class engineers and physicists for SpaceXAI, with applications welcome even from those without AI experience. Previously, Ethan He, the head of xAI’s world model, announced his departure; he was a core member of Grok Imagine from its inception.

Sources:

Tesla Discontinues Model S/X and Launches FSD in China

On May 21, Tesla held the final delivery ceremony for the Signature Edition of the Model S and Model X, officially discontinuing these two flagship models. Elon Musk quoted multiple tweets expressing his feelings about these cars, saying “S & X will forever be in my heart.” On the same day, Tesla announced that FSD Supervised has officially entered the Chinese mainland market, following Tesla China’s urgent recruitment of FSD engineers. Several bloggers confirmed this news, with some users preparing to test FSD’s performance under domestic road conditions. Additionally, Lithuania became the second European country to approve FSD Supervised.

Sources:

Tencent Open-Sources Multilingual Translation Model Hy-MT2, 1.8B Lightweight Version Deployable Locally

Tencent has open-sourced the Hy-MT2 multilingual translation model, supporting translation between 33 languages, available in three sizes: 1.8B, 7B, and 30B-A3B. The company claims that the 7B and A3B versions outperform open-source models such as DeepSeek V4 and Kimi K2.6 in translation tasks, while the 1.8B lightweight version generally surpasses commercial translation APIs like Microsoft Translator and Doubao Translator. Thanks to Tencent’s AngelSlim 1.25-bit extreme quantization, the model requires only 440MB of storage and can run locally on mainstream mobile chips, achieving speeds 1.5 times faster than the previous generation Hy-MT1.5. Tencent has also launched a translation mini-program.

Sources:

Hugging Face Open-Sources Ultra-Fast DNA Model Carbon, Processes Complete Human Genome in Two Days on a Single GPU

Hugging Face has open-sourced the genomic foundation model Carbon, which is reportedly 275 times faster than the previous best model (Evo2) and can process a complete human genome in less than two days using a single GPU. Carbon uses a unique tokenizer that splits sequences into blocks of six bases but processes them at single-base resolution during training and inference, balancing efficiency and accuracy. The model offers an interactive demo that supports exploring DNA sequence generation, gene structure research, mutation impact prediction, and protein generation and folding.

Sources:

China’s AI Short Drama Market Enters Industrial-Scale Production, AI Comic Dramas Become Growth Driver

A16Z has taken note of China’s AI short drama industry. According to multiple industry data sources, the market size for micro-short dramas and comic dramas in China reached hundreds of billions of yuan in 2025, surpassing the total box office revenue of Chinese films (51.832 billion yuan) in the same period. Among these, AI comic dramas have seen the most explosive growth, with a market size of approximately 16.8 to 20 billion yuan in 2025. In January 2026, 14,634 new episodes were released, averaging over 470 new dramas per day. In terms of cost, the production cost of AI comic dramas is as low as 500 yuan per minute, about one-fifth that of traditional animated short dramas. Reports indicate that Seedance 2.0 has increased the usability rate of 15-second videos from the industry average of about 20% to nearly 90%. On platforms like Douyin, high-rated premium comic dramas with realistic human characters can receive a guaranteed minimum of 10,000 to 30,000 yuan per minute. Internationally, the market size was approximately $100 million in 2025 and is expected to reach $650 million in 2026.

Sources:

miHoYo Pledges Up to 100 Billion RMB for AI Large Models Over Three Years

According to a report by LatePost, miHoYo held an AI foundational large model technology sharing session in Beijing on May 15. Liu Wei stated at the event that miHoYo’s AI investment scale could be “up to 100 billion RMB over three years,” adding, “Even if we don’t succeed, we’ll accept it—consider it a big firework.” miHoYo has previously established the “Inverse Entropy” research department, registered the Glossa large model, and successively founded multiple AI-related companies. Liu Wei also mentioned the concept of “AI for AI,” integrating large models into the game development cycle. Meanwhile, miHoYo co-founder Cai Haoyu’s Anuttacon has launched the AI game “Whispers from the Star” and the character performance model LPM 1.0, exploring ways for AI characters to understand, respond to, and express emotions to players in real time.

Sources:

Independent Developer’s Sudden Death Sparks Industry Health Reflection, Vibe Coding Addiction Risk Draws Attention

The sudden death of an independent developer due to prolonged late-night work has sparked widespread concern on social media. Meanwhile, blogger MaxForAI shared his experience of sleepwalking for the first time after 72 consecutive hours of vibe coding, reminding developers to take care of their health. He developed the Tokenless tool, which reportedly reduces vibe coding input tokens for Claude Code by 47.3% and response tokens for regular conversations by 80%. Several bloggers have shared discussions about the addictive nature of instant feedback in AI programming, noting that it can easily make people forget to rest.

Sources:

AI Agent Ecosystem Continues to Expand: Grok Build, Hermes Agent, Multica, and More Advance on Multiple Fronts

The AI agent toolchain is advancing on multiple fronts. Grok Build released version 0.1, adding support for macOS Intel and Windows ARM64, and entering the Vercel AI Gateway and Open Design ecosystems. Nous Research’s Hermes Agent introduced skill bundles and a browser skill directory, integrating with Grok Build. The multi-agent management tool Multica has been recommended by several bloggers, supporting Claude/Codex/Gemini and compatible with Windows/macOS/Linux. Additionally, OpenAI’s Codex has added compilation completion push notifications on mobile, optimized reconnection, and introduced the /fork command. Google’s Antigravity also launched version 2.0 but sparked controversy. Chrome DevTools for Agents 1.0 was officially released, allowing agents to truly see and debug the browser.

Sources:

Statistics: Scan timeline entries=360, Bloggers hit=42, Total tweets hit=251, Weighted tweet score=200.3, Original tweets=121, RT tweets=51, Crawl attempts=2, Boundary coverage status=tail_confidently_crossed_target_boundary