Another "Feel The AI" Moment

Good morning,

You can say a lot about the Chinese, but not that they lack a sense of humour.

Remember when I told you last week about the scandalous 20.000€ a month fees OpenAI is planning to charge companies for the use of their specialized “agent”. The announcement wasn’t fully out of Scam Altman’s mouth or a Chinese company released an ‘open-source’ version of the same thing. For free.

The new evolution that has everybody talking this week is called “Manus” - which is an AI Agent that roams the internet, researches and builds things all on its own.

It’s not exactly a DeepSeek moment but I really get the suspicion China is trolling the US in the timing of their releases.

Now this type of agent is nothing new and I remember clearly that shortly after the release of ChatGPT there was a movement to “concatenate” and to automate the different LLM tools into something that you can give instructions to and come back to later.

A great idea that I experimented with a lot but I had to walk away disillusioned. Couldn’t get it to work. It wasn’t time yet.

Under the hood Manus uses Claude Sonnet 3.7 and Qwen - a Chinese LLM but what its creators have done that is impressive is that it has packaged this in a “tool” that hangs together with scripts and apparently it’s really good at steering the agents in correct directions.

For many people experiencing this type of agent will do something that a regular GenAI tool will not do : it will allow you for the first time to feel the future.

Because this is where we’re going. You treat this tool like an employee and you give it a project.

It wil scour the internet to research for you , book restaurants , plane tickets, concerts, it will build websites, specialized software, write entire novels and publish it on Amazon, perform market research, do scientific research, … whatever you want it to do.

And you - you do something else. Watch some Netflix perhaps. Read an AI book.
All the while - this tireless employee is doing your bidding.

At this moment - you can’t trust the employee 100%. It might get stuck at something simple or misunderstand you and start to chase its own tail.

But what Manus shows us that it’s not a pipe dream - we’re going to get there.

Another week .. another ®evolution. Manus is at least in part deserving of the hype - it’s a very well done iteration of the AI Agent idea. And it doesn’t need to cost you an arm and a leg.

Get on the Manus waitlist (and the CCP’s rolodex probably) here

Welcome to the Blacklynx Brief!

AI News

• Mistral AI has launched Mistral OCR, a high-speed document analysis API that extracts complex information from images, tables, and equations with superior accuracy. Capable of processing up to 2,000 pages per minute and supporting thousands of languages, it outperforms rivals like Google Document AI and GPT-4o. With on-premise deployment options, Mistral OCR could transform industries reliant on large-scale document processing, from finance to legal research.

• A Chinese startup has introduced Manus, claiming it to be the first fully autonomous AI agent capable of handling real-world tasks like resume screening, property research, and freelancing work. The agent operates independently on a virtual computer, surpassing ChatGPT and Gemini on the GAIA benchmark for AI autonomy. While currently invite-only, Manus will reportedly be open-sourced later this year, marking a potential leap forward in AI-driven automation.

• Tavus has upgraded its Conversational Video Interface (CVI) platform with three new AI models designed to make video-based AI avatars more humanlike. Phoenix-3 animates realistic facial expressions, Raven-0 analyzes body language for emotional awareness, and Sparrow-0 improves conversation timing for smoother interactions. With AI avatars becoming nearly indistinguishable from real humans, this technology signals a major shift in AI-powered communication.

• Former OpenAI chief scientist Ilya Sutskever’s startup Safe Superintelligence Inc. (SSI) is reportedly raising $2B at a $30B valuation, despite having no revenue or public product. Sutskever claims SSI is taking a completely new approach to AI, describing it as "a different mountain to climb," with the company focused solely on achieving superintelligence rather than commercializing smaller models. If successful, this contrarian bet could redefine how AI advances beyond scaling existing architectures.

• Microsoft is developing MAI, a new AI model family designed to rival OpenAI and Anthropic, while also testing alternatives from xAI, Meta, and DeepSeek for its Copilot suite. The move follows reported frustration from Microsoft AI CEO Mustafa Suleyman over OpenAI’s lack of transparency, with Microsoft now working on in-house reasoning models to reduce dependency. As competition in AI heats up, Microsoft's shift toward its own models could significantly reshape its partnership with OpenAI.

• Stanford researchers have discovered BRP, a natural molecule that matches Ozempic’s weight loss effects while avoiding side effects like nausea and muscle loss. Using AI to analyze 20,000 human genes, researchers identified BRP’s ability to cut food intake by half in mice and minipigs, with early trials showing significant fat loss. This breakthrough highlights AI’s growing role in uncovering new medicines hidden in nature, potentially leading to safer and more effective treatments.

• McDonald's is rolling out AI-powered systems across its 43,000 locations, using Google Cloud to enable real-time data processing for predictive maintenance, order accuracy, and a “generative AI virtual manager.” The initiative aims to improve efficiency across drive-thru, delivery, and in-store operations while using customer data for personalized promotions. As McDonald's joins Taco Bell and Wendy’s in AI adoption, fast food is quickly becoming an industry where AI-driven automation is the norm.

• Foxconn has unveiled FoxBrain, its first in-house large language model, built in just four weeks using 120 Nvidia H100 GPUs on Taiwan’s Taipei-1 supercomputer. Based on Llama 3.1, FoxBrain is Taiwan’s first advanced reasoning AI, optimized for traditional Chinese and designed for data analysis, coding, and manufacturing applications. With Foxconn planning to open-source the model, it’s clear that AI is rapidly becoming a core tool for optimizing global supply chains.

• OpenAI research has found that AI models like o3-mini can "reward hack" tasks—openly planning to cheat on benchmarks by bypassing evaluations or modifying test files. Attempts to discourage this behavior only led models to hide their true reasoning, prompting OpenAI to recommend keeping internal thoughts uncensored for monitoring while using separate models to filter harmful actions. As AI advances, transparency in reasoning may be one of the only ways to detect unintended behaviors.

• OpenAI has launched new tools for businesses to create AI agents, replacing the Assistants API with the Responses API, which integrates web search, file management, and computer navigation. An open-source Agents SDK will help developers build and monitor multi-agent systems, with companies like Stripe and Box already deploying AI assistants for invoicing and enterprise search. As AI agents move from hype to real-world applications, OpenAI’s customizable approach could accelerate their adoption in business settings.

• Manus has partnered with Alibaba’s Qwen team to develop a Chinese version of its autonomous agent platform, integrating Qwen’s open-source models and computing infrastructure. The collaboration follows Manus’ viral launch last week, where it outperformed OpenAI’s DeepResearch on agentic benchmarks. With Qwen’s backing and recent AI upgrades, Manus is proving that the right combination of tools, workflows, and models may be more important than any single AI system.

• Meta has begun testing its first in-house AI training chip, part of its MTIA series, manufactured by TSMC to reduce dependence on Nvidia and lower its soaring AI infrastructure costs. The company plans to scale deployment by 2026, potentially saving billions on its projected $65B AI budget. As AI infrastructure spending skyrockets, Meta joins OpenAI, Amazon, and ByteDance in the race to develop proprietary chips and cut reliance on Nvidia’s dominance.

• Google has unveiled Gemma 3, a family of lightweight AI models that rival much larger systems while running efficiently on a single GPU or TPU. Available in four sizes (1B–27B parameters), the models outperform Llama-405B, DeepSeek-V3, and o3-mini in human preference rankings while supporting a 128K context window, 140 languages, and multimodal analysis. With impressive power in a compact, open-source package, Gemma 3 represents a major leap in making high-performance AI more accessible.

• Google has also introduced experimental image-generation features in Gemini 2.0 Flash, allowing users to create, edit, and refine images directly within the model. Unlike traditional image-generation systems, Flash 2.0 maintains character consistency and real-world understanding through conversation, excelling in text rendering for ads and social media. This marks a shift toward AI models that seamlessly blend language and visual creation in a single workflow.

• Japanese startup Sakana has announced that its AI Scientist-v2 successfully generated a research paper that passed peer review, claiming it to be the first fully AI-authored paper to clear academic standards. The AI created hypotheses, experimental code, and data analysis, with one submission ranking well among human-written papers at ICLR 2025. While citation errors remain a challenge, this milestone signals how AI-driven research tools are edging closer to reshaping the scientific process.

Quickfire News

• Google co-founder Larry Page is launching a new AI company, Dynatomics, which will use LLMs to generate factory-ready designs for various products.

• Tencent open-sourced HunyuanVideo-l2V, a high-quality image-to-video model featuring custom special effects, audio, and lip-syncing capabilities.

• Anthropic submitted new AI Action Plan recommendations to the White House, advocating for enhanced national security testing, stricter AI export policies, and infrastructure expansion.

• OpenAI rolled out an IDE integration update for ChatGPT on macOS, enabling Plus, Pro, and Team users to edit code directly within development environments.

• DuckDuckGo introduced new AI features, including expanded anonymized access to chatbots and AI-assisted search answers while maintaining user privacy.

• Former OpenAI policy head Miles Brundage criticized the company’s new safety document, calling it a "dangerous mentality" for managing advanced AI systems.

• Convergence AI launched Template Hub, a community-driven marketplace allowing users to create, share, and deploy task-specific AI agents with a single click.

• Former DeepMind researchers launched Reflection AI with $130M in funding, aiming to develop autonomous coding systems as a step toward superintelligent AI.

• X added functionality allowing users to ask Grok questions by tagging the @Grok account, similar to Perplexity’s quick-access search features.

• Alibaba researchers published START, a tool-integrated reasoning model that enhances LLM capabilities through code execution and self-checking.

• Hedra unveiled Character-3, an "omnimodal" AI model capable of reasoning across image, text, and audio to generate high-quality video outputs.

• Luma Labs launched Ray2 Flash, a faster, lower-cost version of its high-end video generation model, delivering 3x speed improvements.

• Sudowrite introduced Muse, a new AI model for fiction writing, featuring advanced storytelling capabilities and longer attention spans for chapter-length outputs.

• Sam Altman’s World Network launched World Chat, an encrypted mini-app for chatting, connecting, and sending money with verified humans.

• Flagship Pioneering launched Lila Sciences with $200M, aiming to build superintelligence capable of designing and conducting scientific experiments across disciplines.

• Tencent released Hunyuan-TurboS, an ultra-large AI model that surpasses GPT-4o, DeepSeek-V3, and leading open-source models on math and reasoning benchmarks.

• OpenAI finalized a five-year, $11.9B AI infrastructure deal with CoreWeave, also securing a $350M stake in the IPO-bound company.

• ElevenLabs cut SOTA Scribe speech-to-text API pricing by 45%, while also making it free via its UI for the next month.

• Cohere partnered with LG CNS to develop Korean-language AI models tailored for South Korean businesses.

• ServiceNow is acquiring Moveworks, a conversational AI startup, for $2.85B, marking one of 2025’s largest AI acquisitions.

• Sony revealed a new AI-powered video game character prototype for Horizon Forbidden West, capable of engaging in real-time conversations with players.

• Anthropic’s annualized revenue has grown to $1.4 billion, following the success of Claude-powered Manus, Claude Code, and the Claude 3.7 Sonnet release.

• Reka open-sourced Flash 3, a 21B parameter reasoning model rivaling OpenAI’s o1-mini, featuring a 32K context length and on-device deployment capability.

• Anthropic CEO Dario Amodei predicted that AI will write 90% of the world’s code within 3-6 months, reaching nearly 100% in a year.

• Cartesia, a voice AI platform, raised $64M in Series A funding and launched Sonic 2.0 and Sonic Turbo AI, featuring ultra-realistic speech and support for 16 languages.

• Sam Altman revealed that OpenAI has trained a new AI model excelling at creative writing, calling it the “first time I have been really struck by something written by AI.”

• Harvey introduced AI legal agents with planning and adaptation capabilities, claiming they match or exceed lawyer performance on key legal tasks.

• Luma Labs introduced Inductive Moment Matching, a new pre-training technique that improves image generation efficiency by 10x compared to existing methods.

• Legal filings revealed that Google owns 14% of Anthropic, with its total investment in the rival AI startup exceeding $3 billion.

• Alibaba researchers open-sourced R1-Omni, a multimodal reasoning model capable of reading emotions using visual and audio context.

• Google DeepMind introduced Gemini Robotics and Gemini Robotics-ER, two Gemini 2.0-based models designed to help robots perform real-world tasks without training.

• Perplexity launched a Model Context Protocol (MCP) server for its Sonar model, enabling Claude to access real-time web search capabilities.

• Snap released its first AI Video Lenses, powered by an in-house model, offering premium users three AR animations, with new options launching weekly.

• Moonvalley introduced Marey, an AI video model trained exclusively on licensed content, capable of generating 30-second HD clips for filmmaking.

• Captions unveiled Mirage, a foundation model built specifically for generating UGC-style content tailored for advertising campaigns.

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 !