Manus
Late last week, a new Chinese AI debuted to massive buzz: Manus, a general-purpose AI agent.
Between the lines:
Boosters have been suggesting that this could be a second "DeepSeek moment," with another Chinese company surpassing American rivals.
However, others are taking a different view: unlike DeepSeek, the creators of Manus are not a frontier AI lab - they're using existing models to create a better user experience.
Even without an underlying research breakthrough, I suspect Manus has crystallized what generalized agents can do - much like Cursor did for AI-assisted programming last year - and is going viral as a result.
Elsewhere in model mayhem:
Cohere launched Command A, a new generative AI model for enterprise tasks that runs on just two GPUs and handles 256K tokens.
Alibaba researchers demonstrated R1-Omni, an open-source model they claim can read emotions, building on their HumanOmni technology.
OpenAI's Sam Altman revealed they've trained an AI model that excels at creative writing, though its release plans remain uncertain.
And Foxconn developed FoxBrain, an in-house reasoning LLM trained in just four weeks using Nvidia's Taiwan-based supercomputer.
Gemini Season
Google Deepmind has been on a roll lately, launching countless models and features under its "Gemini" umbrella. Over a half dozen releases have occurred in the last week alone.
What to watch:
Gemini Robotics and Robotics-ER, two new models to help robots with real-world tasks.
Image generation in Gemini 2.0 Flash, including the ability to edit images with natural language.
Gemini Embedding, a new state-of-the-art embedding model.
Free usage of advanced Gemini app features like Deep Research, Gems, and 2.0 Flash Thinking Experimental.
Gemini with Personalization, an experimental capability that brings your Google account data into Gemini's context.
And while not technically Gemini, there’s also Gemma 3, the latest in Google's open-source family of models.
Elsewhere in the FAANG free-for-all:
Microsoft announces Copilot for Gaming, a new AI gaming coach initially available for Xbox Insiders through the Xbox mobile app.
Meta plans to incorporate improved voice features in Llama 4, betting on conversational AI agents over text-based ones.
Snapchat rolls out AI video Lenses powered by its in-house generative model for Snapchat Platinum subscribers.
And Apple executives have raised concerns about personalized Siri, with some suggesting it needs a complete rebuild.
AI Action Plan
After President Trump rescinded his predecessor's AI Executive Order, the White House requested input in developing a replacement "AI Action Plan." With the deadline set for this weekend, both OpenAI and Google have already weighed in.
Where we're headed:
Both companies argue that the growing patchwork of state AI regulations threatens US technological leadership and innovation, particularly in competition with China.
The copyright issue emerged as a crucial battleground, with both tech giants arguing that the fair use doctrine should be expanded to include AI training.
Ultimately, the policy whiplash is stark: we’ve gone from "Safe, Secure, and Trustworthy" AI to "America's global AI dominance," with AI leaders now advocating for minimal regulation.
Elsewhere in AI geopolitics:
The FCC is establishing a new security council to strengthen US cyber defenses and maintain technological leadership over China in AI and 5G.
Spain's government has proposed legislation imposing hefty fines of up to €35M on companies that fail to label AI-generated content.
And the White House is also considering measures to restrict Chinese AI company DeepSeek, including potential app store bans over national security concerns.
Elsewhere in AI anxiety:
French publishers and authors have launched a legal challenge against Meta for allegedly using their books to train AI without permission.
A comprehensive study reveals that AI search engines are highly unreliable at citing news articles, with Grok 3 having a 94% error rate.
Consumer Reports finds that popular voice cloning services like ElevenLabs and Speechify lack adequate safeguards against voice impersonation.
And experts warn that AI-generated voice evidence could create significant challenges for the legal system.
Things happen
Cursor developer Anysphere is in talks to raise at a $10B valuation. Nvidia to announce new GPUs named after dark matter discoverer Vera Rubin. Vibe coding is a fantastic way to explore LLM capabilities. AI tools are spotting errors in research papers. Bengaluru is becoming an AI talent hub as overseas startups compete with Big Tech. Google owns a 14% Anthropic stake but holds no voting rights. AI 'wingmen' bots to write profiles and flirt on dating apps. 36% of IT jobs in January 2025 were AI-related. Is generative AI hype peaking? Meta is testing its first in-house AI training chip. The UK has 29 of Europe's top 150 startup hubs. Microsoft completed training a family of models codenamed MAI. Judge allows authors' AI copyright lawsuit against Meta to move forward. Signal's president calls out AI agents as having security and privacy issues. Sony is working on an AI-powered PlayStation game character. Apple postponed the launch of a smart home hub. Zoom researchers detail a "Chain of Draft" method for LLMs. Microsoft is testing ways to replace OpenAI's models in Copilot. McDonald's new tech initiatives include edge computing to predict equipment breakdowns. Alibaba debuts New Quark, an updated app using its Qwen reasoning model. CoreWeave signs a five-year $11.9B contract with OpenAI. Kryptos sculptor frustrated with false AI-cracking claims. Beijing schools to offer AI classes for primary and secondary students. US DOJ drops bid to force Google to sell its AI investments. Hugging Face cofounder worries AI is becoming "yes men on servers". The gadget boom that started with Alexa has largely fizzled. With AI you need to think bigger. Xata Agent: AI expert in PostgreSQL. AI fakers exposed in tech dev recruitment.
