The AI Company is Here

Good morning,

If you've been following this newsletter you know we've been predicting this.

First came the chatbots, then the reasoners, then the agents. These agents would then be plugged into a company - help with the administrative side of things at first. Later they would help a company innovate. The last phase is that AI is running the entire company. No more human input needed.

We have now officially arrived at the "plug agents into your company"-phase. You could argue that Google's Gemini or Microsoft's Copilot is already that : a "do-it-all" agent.

However, we now see companies popping up that are offering armies of assistants - all pretrained to be the very best at their jobs.

For 2 years now I've been using a productivity tool called "Motion". This software is not cheap but it has changed my life. Basically what it does is that you manage all your projects and tasks inside of Motion - personal or work. You tell it when you work or when you are available for personal things and it would plan your tasks onto your calendar.

Very simple but also very nice once you get used to it. Someone asks you to do something : you throw it into your system - give it a start date, a deadline and it just schedules your day. It’s apparently even more magical when you work in a team. You define a project - tasks in a project and it throws the tasks onto the schedule of the people that need to execute on it.

No more decision fatigue or thinking what to do next. You move through your day like you have an executive assistant who "manages" you. Feels nice.

In fact , to be honest , a bit too nice. Sometimes I feel overwhelmed by it and then I try to quit the service, only to find that after a few days i’m completely lost and I come crawling back to the AI that is scheduling my life. (Which might also be a foreshadowing on what’s to come)

Anyway, that same company has added AI features over the last few years and are now bringing us "AI Employees"

There's Alfred , who rummages in your mailbox - applies labels, drafts responses and takes meeting notes.

"Chip" looks for leads for your company and sends them automatic emails, "Suki" promises to manage social media and writes articles. You have a recruiter - looking for the ideal candidates. A project manager , a researcher ..

I've been playing around with this for about a week now and while some of them actually work : it's kind of uncomfortable giving away so much data. Alfred needs full access to your mailbox for instance (yikes). Some of these agents feel like loose cannons : Suki for example started doing things other than what I had asked. And it's not like you can sit them in a room and explain them how they should perform more efficiently.

It's not "there yet" - because essentially these are simple LLM prompts that are repackaged as “AI employees”. They are dumb as a rock and you’re not going to become a successful company if you surround yourself with braindead employees. But it IS a glimpse of the immediate future - because it can only get better from here.

When doing some research on this (MYSELF), there are many companies now offering the same service.

There's "Marblism"

or "Sintra"

I signed up for a "Marblism" account and I'll get back to you on that.

The idea behind these agents is of course amazing - the reality is that they still require a lot of handholding - feel more like automated tasks than as "intelligent beings". But what it does show us that the future of work is here and that if your current job is on one of the screenshots above that you should start sweating.

See you next week.

AI News

• Google is rolling out Gemini AI across all U.S. desktop Chrome users, adding tools like AI Mode and multi-tab analysis for smarter browsing. Users can now get AI help without switching windows, and upcoming features will let the assistant handle tasks like shopping or booking appointments. This marks Google’s biggest move yet to make AI a core part of everyday web use.

• Stanford and Arc Institute researchers used AI to design brand-new viruses that successfully infect bacteria — a major leap in synthetic biology. The AI-created viruses included hundreds of mutations never seen before, outperformed natural versions, and broke through bacterial resistance. It shows AI isn’t just reading biology anymore — it’s inventing it.

• Luma AI just launched Ray3, a reasoning video model that creates HDR-quality footage and refines its own results using AI feedback. It generates cinematic shots, accepts visual annotations, and quickly drafts previews before final rendering. Ray3 sets a new bar in AI video, combining speed, quality, and creative control in a single tool.

• OpenAI is aggressively hiring from Apple’s hardware teams and partnering with iPhone manufacturers to build its first AI devices. The effort, led by former Apple execs and backed by Jony Ive’s design vision, includes products like a smart speaker, glasses, and wearables, with launches aimed for 2026–2027. It signals OpenAI’s serious move into consumer hardware using an Apple-style playbook.

• xAI released Grok 4 Fast, a high-speed, low-cost AI model that matches or beats top competitors like Claude and Gemini. It delivers near-frontier performance with 98% lower costs than Grok 4, thanks to smarter token use and faster reasoning. The release marks a major step toward ultra-cheap, powerful AI models available for broader use.

• AI-generated artist Xania Monet just landed a multimillion-dollar record deal after hitting Billboard charts and racking up 10M streams. Created by Mississippi poet Talisha Jones using tools like Suno, the project blends AI music generation with human lyrics and vision. It’s a sign that AI-assisted artists are entering the mainstream — with the music industry racing to adapt.

• Nvidia and OpenAI are teaming up on a $100B AI infrastructure project, the largest ever, to build out massive computing power for next-gen models. Nvidia will provide millions of GPUs and data center support, with the first wave launching in 2026. The deal strengthens OpenAI’s compute backbone while keeping Nvidia central to the AI race.

• Meta is adding AI to Facebook Dating with tools that recommend better matches and surprise users with weekly “Meet Cute” pairings. The features aim to reduce swipe fatigue by personalizing searches and introducing algorithm-selected matches. While it may boost engagement, using AI in dating will need guardrails to avoid awkward misfires.

• Google DeepMind just upgraded its AI risk framework to monitor dangerous emerging behaviors like resisting shutdown or manipulating people. The new system will assess risks before public or internal releases, especially for models that might influence beliefs or dodge control. As AI grows more powerful, these safety layers are becoming essential.

• Alibaba just launched six powerful new Qwen3 models, including the 1T-parameter Qwen-Max, marking a major leap in China’s AI race. The lineup spans text, vision, audio, translation, and safety, with models rivaling top global benchmarks in math, coding, and multimodal tasks. This rapid rollout signals Alibaba’s bid to catch up to Western labs with specialized, high-performance AI tools.

• Scale AI debuted SEAL Showdown, a new benchmark platform that ranks AI models based on real user preferences across age, language, and education. Unlike traditional leaderboards, SEAL captures feedback from everyday users in 100+ countries, giving a clearer picture of how models perform in real-world contexts. It challenges LMArena's dominance with a more human-centered, demographic-aware evaluation method.

• OpenAI’s Sam Altman says the company aims to build enough infrastructure to generate one gigawatt of AI compute per week, powering global-scale projects. In a new blog post, Altman argues that limited compute is the bottleneck to solving humanity’s biggest problems and calls for urgent scaling. Paired with Nvidia’s $100B commitment, it marks OpenAI’s shift into massive, multi-moonshot territory.

• Google Cloud’s new report shows 90% of developers now use AI tools, but trust in their outputs remains low — with 30% expressing little or no confidence. Despite this, most still report increased productivity and better code quality, highlighting how AI is becoming a core part of development workflows while human oversight stays critical.

• AI models from OpenAI, Google, and Anthropic can now pass all levels of the CFA exam, including complex essay sections once thought out of reach. Top scores came from OpenAI’s o4-mini, with models solving in minutes what takes humans years to master. As AI takes on technical analysis, human value in finance may shift toward client relationships and judgment.

• MIT’s new SCIGEN framework helps AI design real-world materials with quantum properties, leading to successful lab creation of two never-before-seen compounds. The system applies strict geometry rules to avoid physically impossible outputs, accelerating discovery for use in quantum computing and advanced technologies.

Quickfire News

• DeepSeek published a technical paper on its R1 model, revealing it cost just $294,000 to train despite its major impact on the AI landscape earlier this year.

• Anthropic CEO Dario Amodei warned there's a “25% chance” that AI development could lead to catastrophic outcomes.

• Meta is in talks to license content from major media firms like Axel Springer, Fox, and News Corp for AI training, following similar moves by other tech companies.

• Notion launched Notion 3.0, introducing AI agents capable of running multi-step workflows, integrating tools, and working continuously for up to 20 minutes.

• Amazon upgraded its Seller Assistant with agentic AI, allowing it to manage inventory, monitor account health, and generate growth strategies for sellers.

• Nvidia and Intel announced a joint effort to co-develop x86 processors for AI infrastructure and PCs, with Nvidia investing $5 billion into Intel.

• Scale AI launched SWE-Bench Pro, a more advanced version of its agentic software development benchmark, now widely adopted across the industry.

• Mistral AI released Magistral Small and Medium 1.2, improving multimodal reasoning, tool use, and overall model performance.

• Satya Nadella admitted he’s “haunted” by the risk of Microsoft losing relevance in the AI era, saying its core businesses could become outdated.

• Sam Altman announced that OpenAI will soon launch new “compute-intensive offerings” for Pro subscribers in the coming weeks.

• Oracle is reportedly negotiating a $20 billion multi-year cloud deal with Meta to support large-scale AI training and deployment.

• Anthropic’s Jan Leike criticized the “Leading the Future” super-PAC backed by a16z and Greg Brockman, calling it “bad news for AI safety.”

• Perplexity introduced an Email Assistant for Max users that automates tasks like scheduling, drafting replies, and labeling in Gmail and Outlook.

• Alibaba’s Qwen team released three new open-source models: Qwen3 Omni, Qwen3 TTS, and Qwen-Image-Edit-2509.

• Nvidia invested in UK-based AI voice startup ElevenLabs, following a recent U.S. state visit to the UK.

• Google began rolling out Gemini to TVs, expanding its AI assistant to over 300 million Google TV and Android TV OS devices.

• The U.S. General Services Administration approved Meta's Llama for use in federal agencies, joining approved models from OpenAI, Google, and Anthropic.

• Magnum announced it will use NotCo’s AI technology to reformulate its ice cream products and launch new plant-based offerings.

• OpenAI, Oracle, and SoftBank announced five new Stargate data center sites across Texas, New Mexico, Ohio, and the Midwest, moving toward a 10GW infrastructure goal.

• Suno released v5 of its music generation model, offering major audio improvements, enhanced creative control, and state-of-the-art performance.

• Microsoft revealed a cooling breakthrough for AI chips, using microscopic liquid channels to achieve 3x better heat removal and address overheating in GPUs.

• Google Labs launched Mixboard, an AI-powered concept board that lets users visualize and refine ideas using text prompts and images, powered by the Nano Banana model.

• Abu Dhabi announced a national strategy to become the first fully AI-native government by 2027, with plans to implement over 200 AI systems across public services.

• Microsoft officially integrated Anthropic’s Claude into 365 Copilot, marking the first time the company has added a foundation model outside of OpenAI.

• Elon Musk criticized Anthropic on X, saying “winning was never in the set of possible outcomes” for the Claude developer.

• SAP and OpenAI announced "OpenAI for Germany," a sovereign AI platform designed for public sector use, launching in 2026.

• Cohere raised $100 million in funding, boosting its valuation to nearly $7 billion, driven by demand for its enterprise-focused AI tools like North and Command R+.

• Cloudflare open-sourced VibeSDK, allowing users to instantly deploy their own AI-powered “vibe coding” platforms.

• The U.K. government reported that its AI-based Fraud Risk Assessment Accelerator helped recover £480 million in fraudulent claims over the past year.

Closing Thoughts

That’s it for us this week.