In partnership with
10 AI Stocks to Lead the Next Decade
AI isn’t a tech trend – it’s a full-blown, multi-trillion dollar race, and 10 companies are already pulling ahead.
These are the innovators driving real revenue, attracting institutional attention, and positioning for massive growth.
Get all 10 tickers in The 10 Best AI Stocks to Own in 2026, free today.
AIHOY and happy weekend, dear AInauts!
Maybe you missed some of last week's exciting AI news, tools, and hacks — or you've just found us for the first time. No worries!
Here's a roundup of our highlights from the past week:
And to wrap things up, we've got the most important Quick News for you — everything you need to know, one click away! Ready? Let's go!
OpenAI is pulling the pieces together. Codex, agents, third-party apps, and ChatGPT are starting to look less like separate tools and more like one big work surface. The friendly chat window is becoming the place where you write, build, automate, and hand work to agents without bouncing between tabs all day.
The numbers explain the move. Two million business customers already account for roughly 40 percent of OpenAI's revenue. Companies are not paying for chit-chat. They are paying for productivity. So of course the product is moving closer to the actual work.
Our take: the default wins again. If one interface gives you chat, code, agents, and apps in a clean enough package, a lot of specialized tools start looking fragile. Convenience beats the better niche feature more often than builders like to admit.
Anthropic is saying the quiet part out loud: more than 80 percent of the code going into Claude is now written by Claude. AI is helping build its own next version. That is the loop some people fear and others are quietly betting on.
We map out three paths: faster research, machine loops that start feeding themselves, and the practical limits that may still slow the whole thing down. One thing feels pretty obvious: global coordination is not a realistic brake in this race. Nobody wants to be the lab that voluntarily slows down.
Our take: this is already happening inside the major labs' codebases. No need to panic. But pretending it is not happening is not a strategy either.
NotebookLM can produce surprisingly good infographics, but it has two annoying habits: typos inside the image and a big logo nobody asked for. That bothered us for weeks.
Our fix this week: run the finished graphic through GPT Image 2.0 once. The model cleans up the text mistakes, removes the logo, and leaves the design mostly alone. Two tools chained together, problem solved, almost no manual design cleanup.
Our take: the best AI workflows often look exactly like this. One tool creates the rough output, the next tool removes the weak spot. Try it the next time a graphic export is almost good enough.
Anthropic released Claude Fable 5, and it feels like the next step up: stronger at coding, better at long autonomous runs, more useful for complex chains of work. The benchmark numbers are nice. The real story is what happens when a model can keep working for an hour without needing your hand every three steps.
The catch matters just as much. Top models are getting expensive. Fable is expected to leave the normal subscription and move back to per-token pricing. That is a preview of the next practical skill: knowing when you need the expensive model and when a cheaper one is enough.
Our take: try Fable while it is still included. Give it your hardest postponed problem, not a tiny email task.
The AI bubble has a new favorite word: loops. It sounds like developer mysticism at first, but the basic idea is simple. A loop is a repeatable process that kicks off an agent, checks the output, and runs another pass if the result is not good enough. Less "I write the perfect prompt" and more "I build a small process that keeps doing the work."
For non-engineers, the useful question is not: how do I build the most advanced loop? The better question is: which parts of my work are loop-able at all? Anything that happens regularly, can be checked clearly, and does not require a sensitive human judgment call is a candidate. Everything else should probably stay with you.
Our take: prompting is not dead. But prompting without process is getting weaker. Good workflows will include both the request and the review loop.
The new models do not love bloated prompt novels. That hurts a little, because many of us spent years building exactly those monsters. But the direction is clear: fewer nested instructions, a clearer goal, cleaner context, and a good definition of done.
The real test is not whether the model gives you a nice answer. Give it a real task with the reason, the goal, the material, and the output format. Then let it work. If you gave it enough context, it should deliver instead of asking for help every two minutes.
Our take: the new prompt is shorter, but not thinner. It says less about every tiny step and more about the why. That is where "using a chatbot" starts turning into "leading an agent."
Stop doomscrolling for tech news
TLDR is the free daily email that does the reading for you. Each issue summarizes the most interesting stories in startups, tech, and programming, curated by ex-Google and Anthropic engineers, so you get the signal without the noise.
It's one email instead of 12 open tabs. A 5-minute read with your morning coffee, and you walk in already caught up.
Tech is just the start. We also cover AI, marketing, dev, and more, so you can follow the topics that matter to your work.
Free, daily, and read by 7M+ subscribers. Subscribe and get your mornings back.
AI News Quickie: the HAI-lights from the industry
AI never sleeps. Here are the most important stories from this week.
OpenAI is moving on several fronts
OpenAI confirmed a confidential S-1 filing with the SEC on June 8. No IPO date yet, but the door is now officially open.
Sam Altman and Jakub Pachocki sketched OpenAI’s next phase: automated AI research by 2028, broad economic upside, and access to personal AGI. Big words. Big expectations.
With the Economic Research Exchange, OpenAI is bringing outside researchers closer to the data and tools needed to measure AI’s impact on work and productivity.
OpenAI updated GPT-Rosalind for life sciences. The direction matters: specialized models for industry work, not one general model for every job.
ChatGPT is getting a rebuilt memory system called “Dreaming V3” and, after an efficiency jump, it is rolling out to free users too.
🏛️ Anthropic and Claude are getting stronger, pricier, and trickier
Anthropic introduced Claude Fable 5 and Mythos 5. Fable is publicly available, while Mythos stays more restricted.
Fable 5 costs $10 input and $50 output per million tokens, according to Anthropic. A reality check for everyone treating frontier models like a flat-rate subscription.
Claude Managed Agents now get schedules and env vaults. That makes agents feel less like demos and more like operations.
Claude Fable 5 has also been available in Microsoft Foundry and GitHub Copilot since June 9. Anthropic gets to sell the power twice: directly and through Microsoft’s enterprise channel.
With Claude Corps, Anthropic is sending 1,000 young professionals into US nonprofits for a year, of course with Claude at their side.
Through the DXC alliance, Anthropic is pushing Claude deeper into banks, airlines, and other regulated industries. In other words: where the customers with real budgets sit.
Anthropic is testing Claude as a chemist and shows NMR analysis with Opus models in a research post.
🖥️ Google, Apple, and the new device war
Apple introduced Siri AI and new Apple Intelligence on June 8. For EU users, the old question remains: when do we actually get it?
Apple Intelligence is getting new everyday tools for photos, Safari, passwords, and communication, closer to the operating-system layer.
Apple’s new Safari feature Describe an Extension shows where the browser is headed: you describe a small function, and the system builds the helper into the toolbar. Shortcuts are getting easier too with Describe a Shortcut.
Apple is making AI feel less like “assistant answers questions” and more like daily context. Call Context is supposed to surface relevant details during calls, including things like reservations.
Safari will be able to monitor websites with Notify Me and alert the user. Clever feature. But who is still using Safari? Asking for a friend...
Google Live Translate pushes AI closer to real conversations because it is supposed to preserve tone and pace. Translation is moving closer to an actual interpreter.
Google is bringing Gemini models to Apple developers. The practical part: Apple apps can combine local Apple models with cloud-based Gemini models.
Google announced Gemini 3.5 Live Translate. Near-real-time speech-to-speech translation is one of those AI use cases that lands in daily life immediately.
Google is upgrading NotebookLM with Agentic Chat and Code Execution. “Summarize my sources” is slowly becoming “work inside my source room.”
NotebookLM also gets new output formats such as PDF, Excel, PowerPoint, and CSV. That saves exactly the annoying cleanup work after research.
Local models, open source, and agent infrastructure
This block is a little more technical, but the builders and hobby coders among us should find a few useful nuggets here.
Google introduced DiffusionGemma, an experimental text diffusion model with up to 4x faster output. According to Google, it uses 26B MoE and 3.8B active parameters and is meant to help with inline editing and code infilling.
Google is open about DiffusionGemma quality still trailing standard Gemma. That trade-off honesty is useful, because faster is not automatically better.
Google released Gemma-4 models with Quantization-Aware Training. For you, that means stronger local AI on more normal hardware. The QAT models reduce memory needs and support Ollama, LM Studio, and MLX.
NVIDIA introduced Nemotron 3 Ultra for long agent runs. The claim: more throughput and lower agent costs. This matters because NVIDIA is leaning into long-running agents. The hardware company is selling chips and the agent story at the same time.
Friendly reminder: Zapier MCP connects Claude, ChatGPT, and other tools with 9,000+ apps and 30,000 actions. Once connected, your agent can work directly in Gmail, Slack, or Salesforce.
Microsoft’s markitdown is practical. It turns PDFs, Office files, images, and audio into clean Markdown, which is ideal when you want to make messy content AI-readable.
headroom compresses tool outputs and similar material before they hit the model. That can save 60 to 95 percent of tokens at the same answer quality.
Reve 2.0 is betting on structured layouts for image generation instead of pure text-prompt magic.
🛡️ Security, regulation, and infrastructure
Google.org is launching a $50 million initiative for skilled trades workers, according to Axios. Data centers need people with tools.
NVIDIA’s Confidential Computing story shows why Private Cloud Compute is more than branding. With encrypted inference, the question is whether sensitive data can travel through AI backends at all.
Apple’s new AI features come with SynthID watermarks for generated or edited images. That makes manipulation a little less invisible.
With Fable and Mythos, Anthropic publicly separates general access from Trusted Use. That will shape the debate: who gets the strongest models, and who decides?
OpenAI emphasizes international coordination for AI safety in its new plan. Coordination is easy to demand and hard to build.
The American Federation of Musicians is suing major labels over AI licenses for Suno and Udio. The question is shifting from “can this be trained?” to “who gets paid?”
Apple is again making Private Cloud Compute central to Siri AI and Apple Intelligence. Apple’s strongest argument against ChatGPT and co. may not be better models. It may be trust.
Apple’s Passwords app is supposed to repair weak logins automatically.
NVIDIA is supporting Apple’s Private Cloud Compute with Confidential Computing.
The thread this week: loops, agents, and superapps are growing up. The workflows that win will be the ones you actually repeat tomorrow.
Have a good weekend!
Best regards,
Reto & Fabian
