Welcome back! Sam Altman said in 2024 that AI would make the one-person billion-dollar company possible. A guy selling weight-loss drugs from his living room in LA just proved him right. Elsewhere, OpenAI made its first move into media, Google dropped its most permissive open models to date and Perplexity shipped the most practical demo of computer-use AI we have seen so far.
In today’s Generative AI Newsletter:
Medvi: How did a solo founder build a $1.8B company with AI and $20K?
OpenAI: Why did an AI lab just buy a talk show?
Google: Can Gemma 4 break the Chinese grip on open-source AI?
Perplexity: Is a browser agent really about to do your taxes?
Latest Developments
A Solo Founder Used AI To Build A $1.8 Billion Company
Matthew Gallagher spent $20,000 and two months building Medvi, a telehealth startup selling GLP-1 weight-loss drugs online. A year and a half later, the company is on pace for $1.8 billion in annual sales. His only full-time hire is his brother.
The details:
The stack: Gallagher used ChatGPT, Claude and Grok for code, Midjourney and Runway for ad creatives and ElevenLabs plus custom AI agents for customer service. Doctors, prescriptions and shipping are outsourced to telehealth platforms CareValidate and OpenLoop.
The numbers: Medvi brought in $401 million in revenue in its first year. The company now uses contract engineers and account managers but the core operation was stood up by one person with off-the-shelf AI tools.
The context: Altman predicted in 2024 that a solo billion-dollar company would arrive. Gallagher is the first credible example, and the business itself has nothing to do with AI as a product. It is a drug distribution operation that happens to run on AI infrastructure.
The lesson here is not that anyone can replicate this. GLP-1 demand is extraordinary and the regulatory window for online telehealth prescribing may not stay open forever. But the operational model is the sign. What used to require a 50-person team, office space and millions in seed funding can now be assembled by one person with the right tools and enough urgency to move.
Special highlight from our network
Practical courses on ChatGPT, Claude, Cursor, Perplexity, Deep Seek and more, taught by instructors who actually use them in production. Whether you want to build faster, automate your work or figure out which tools are worth your time, there's a course for it.
What's inside:
Prompt engineering for professionals
Claude Code from zero to shipping
AI-powered marketing, data science and business strategy
Beginner to advanced tracks across every major platform
No filler. No theory-only lectures. Just the skills that are actually moving careers forward right now.
OpenAI Just Bought A Talk Show
OpenAI has acquired TBPN, the daily live tech talk show that pulls around 70,000 viewers per episode and has become required watching for Silicon Valley founders and executives. The deal is reportedly worth low hundreds of millions and marks OpenAI's first media acquisition.
The details:
The show: TBPN goes live every weekday on YouTube and X, hosting major tech CEOs and figures. Co-founders Jordi Hays and John Coogan launched it 17 months ago and the company was on pace for $30 million in revenue this year.
The structure: TBPN's 11-person team will report to OpenAI chief of global affairs Chris Lehane. The company drops its advertising business but retains editorial independence.
The reasoning: OpenAI president Fidji Simo said the standard communications playbook does not apply to a company driving a technology shift of this scale. TBPN gives OpenAI a direct channel to the founders and operators who watch it daily.
This is a reputation play. OpenAI has taken sustained public criticism this year over safety, pricing and competitive tactics. Rather than fighting that battle through press releases and blog posts, they have bought the room where the tech conversation actually happens. Whether TBPN can maintain credibility under corporate ownership is the obvious question, but the editorial independence clause suggests OpenAI understands that killing the show's authenticity would defeat the purpose.
Google Open-Sources Its Most Capable Models Yet
Google DeepMind released Gemma 4, a family of four open models under an Apache 2.0 licence for the first time. The lineup spans 2B, 4B, 26B (mixture-of-experts) and 31B parameters, all with vision and multi-step agent capabilities built in.
The details:
The range: The two smaller models add native audio input for speech recognition and can run entirely offline on a phone. The two larger models are available through Google AI Studio.
The licence: Apache 2.0 means developers can modify, deploy and sell commercially with zero legal friction. Previous Gemma releases used a custom licence that pushed enterprise users towards alternatives like Qwen and Mistral.
The benchmarks: The 31B and 26B models sit close to rivals like Kimi K2.5, GLM-5 and Qwen 3.5 in intelligence benchmarks while coming in at a fraction of the size.
The timing matters. Chinese labs have dominated the open-source AI frontier for the past year, but this week has seen two major US releases push back: Arcee AI's Trinity-Large and now Gemma 4. Google is trending in the opposite direction from its Chinese competitors, who are increasingly moving towards closed systems.
Perplexity Computer Will Do Your Taxes
With the April 15 US tax deadline approaching, Perplexity dropped a demo of its Computer agent preparing a federal tax return. The video has racked up 1.9 million views and is probably the most tangible demonstration of what computer-use AI actually looks like for normal people.
The details:
How it works: Upload your tax documents, answer a few questions about your situation and watch the agent fill out the relevant IRS forms. Perplexity claims Computer retrieves and applies the current tax code to every filing.
The appeal: This is not a chatbot summarising tax advice. It is a browser-based agent navigating forms, entering data and applying rules. The demo makes the abstract concept of "computer use" immediately concrete.
The caveat: Perplexity is clear that users should double-check the output. This is preparation assistance, not a replacement for professional review on anything complex.
Computer-use agents have been a talking point since Anthropic first shipped the capability last year, but most demos have felt like developer previews. This is different because the use case is universal. Everybody has to file taxes. Showing an AI doing the single most universally dreaded administrative task of the year is a smarter marketing move than any benchmark chart.
TOP 5 Tools of the Week
ElevenLabs: is a conversational agents platform that deploys natural, human-sounding voice and chat agents in 32 languages with ultra-low latency. Connected to your knowledge and tools, they automate complex workflows and deliver fast, reliable resolutions.
Higgsfield: is an AI-powered video generation platform that enables creators and marketing teams to produce high-quality, social-native video content from text and images, accelerating creative production at scale.
Encord: is an AI data platform for computer vision and multimodal teams to label, manage, curate, and evaluate training data and models faster.
Vanta: is an AI-powered security and compliance platform that helps companies automate trust, streamline audits, and scale with confidence across SOC 2 and emerging AI risk frameworks.
Alta HQ: is an AI-powered outbound automation platform that uses autonomous agents to research prospects, generate personalized messaging, and execute end-to-end sales campaigns, helping teams scale pipeline generation with minimal manual effort.
Light Bytes
ByteDance Seedance 2.0: AI video generator now broadly available and sitting at the top of Artificial Analysis' video leaderboards.
Alibaba Qwen3.6-Plus: New reasoning model rivalling Opus 4.5 on coding agent benchmarks with native 1M-token context and multimodal inputs.
Microsoft MAI-Transcribe-1: New speech-to-text model topping accuracy benchmarks across 25 languages, now in public preview.
Sanctuary AI robotic hand: New video shows the company's hydraulic hand autonomously manipulating a lettered cube ten consecutive times without dropping it, all at the fingertips with no palm support. A successful zero-shot transfer from simulation to hardware.
Cursor 3: Full interface rebuild with multi-workspace support, letting developers run fleets of local and cloud coding agents in parallel across multiple repos.
Microsoft lost OpenAI: Analysis of how Microsoft lost its largest customer and the trust of the US government in one of the most preventable corporate mishaps of the decade. Doh.
