In partnership with
Want to get the most out of ChatGPT?
ChatGPT is a superpower if you know how to use it correctly.
Discover how HubSpot's guide to AI can elevate both your productivity and creativity to get more things done.
Learn to automate tasks, enhance decision-making, and foster innovation with the power of AI.
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!
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
Sometimes AI gets better not when you keep prompting it nicely, but when you force it into review mode. The trick isn't to berate Claude or ChatGPT. The trick is: make the model reason, constrain, and back things up — instead of just handing you another polished draft.
We turned this into a skeptical senior reviewer prompt. It hunts for real weaknesses — but only with location, risk, and the smallest possible fix. That guardrail matters: without it, AI will happily invent problems just because you told it to find some. Works for copy, strategies, code, and plenty of half-baked ideas.
Agents are getting more useful. That also makes them more vulnerable. Hidden instructions can now be embedded in text, images, audio, video, PDFs, or web pages. As long as an agent only reads, that's annoying. The moment it can touch email, calendars, CRM, Drive, GitHub, or a browser, things get serious fast.
Our take: don't shut agents down — but build them more conservatively. Separate reading from acting. Treat external content as untrusted. Give agents fewer permissions than would be convenient. Yes, that's a pain. But security is almost always a trade against comfort. Before: more convenient. After: more secure.
AI automation feels like liberation: fewer clicks, less routine, more output. Then it starts producing meaning: raw copy, decisions, summaries, customer replies, research. And that's exactly where the work starts again: Is any of this actually right?
The new bottleneck is judgment. You're no longer just checking whether the flow runs. You're checking whether the output is accurate, current, sounds like you, and is fit to go out to real people. Automate everything and let no one actually decide anything, and you haven't built a productivity machine. You've built a faster slop machine.
Microsoft and NVIDIA used Build to show the hardware side of the agent shift: new Windows PCs with RTX Spark, up to 128 GB of memory, and enough local AI horsepower to run large models directly on the device. The cloud isn't going anywhere, but everyday workloads are slowly migrating back to your own machine.
This is less gadget news than a setup question. When AI runs locally, privacy gets easier, always-on operation gets cheaper, and files-over-tools becomes even more valuable. The computer stops being just a tool. It becomes more like a colleague with its own workbench. Nice.
We keep seeing the same gap in teams: plenty of people use AI — but few have any idea what's actually possible. The way we see it, there are three useful rungs right now: AI alongside you, AI works while you review, and AI works without you. Most people are stuck on rung one. That's fine, but it also explains why AI still feels like a glorified chatbot to so many.
Rungs two and three are where things get interesting: Files-over-Tools, Codex, Claude Cowork, local models, early always-on agents. Hermes by Nous Research is our current favorite — it reasons locally, learns across sessions, and is finally testable as a desktop app. Not every task needs the top floor. But you should know those options exist.
📰 AI News Quickie: The HAI-lights from the industry
AI never sleeps! This week was an infrastructure week: local AI PCs, desktop agents, dynamic workflows, new Microsoft models, Anthropic security, Codex plugins, robotics, and the next round of AEO reality.
Short version: the tools are sliding deeper into operating systems, hardware, and real workflows.
OpenAI and Codex push toward a work platform
Codex is taking on more roles beyond just coding: OpenAI describes plugins for various workflows and shareable Sites. Codex is becoming less of a terminal toy and more of a workspace for teams.
According to release notes, Codex can now operate Windows applications and use remote access from a phone. If that holds up, debugging shifts from the desk to review mode.
OpenAI models and Codex are landing on AWS. The more clouds that host frontier models, the less any single chatbot tab matters.
OpenAI is rolling out a new ChatGPT memory layer called Dreaming for Plus and Pro users. The key point: memory is no longer just stored — it's being condensed and kept current in the background.
Tying in with our biosecurity piece, OpenAI continues pushing biodefense and is expanding access for trusted partners with GPT-Rosalind. Performance is being measured against real research workflows via LifeSciBench.
OpenAI is building the next Stargate piece in Michigan: a 1 GW data center campus going up with Oracle, Related Digital, and Crusoe.
Florida is suing OpenAI over a tragic ChatGPT case. The underlying question: where assistance ends and responsibility begins.
Bernie Sanders wants to tie major AI companies more tightly to the public good through public ownership stakes, triggering a political reflex. Unrealistic? Maybe. But AI capital is now firmly a government topic.
Anthropic, Claude, and security get more operational
Anthropic is expanding Project Glasswing to around 150 organizations in 15+ countries. Critical infrastructure is becoming the testing ground for AI-assisted defense.
Claude Code is getting Dynamic Workflows and builds its own harness with subagents for complex tasks. That's the difference between a chat and a work machine! But heads up — it can get expensive …
Anthropic has confidentially filed an S-1 with the SEC. IPO option open, timing open, but the signal is clear: Claude wants to play infrastructure, distribution, and capital markets all at once.
Mythos is finding a very large number of security vulnerabilities. Finding them is no longer the hard part — closing them is, and the bottleneck is shifting toward cost and patch capacity.
The US is launching an AI cybersecurity push via Executive Order to embed safety more deeply in government infrastructure: hardened federal systems, AI-assisted defense, and voluntary frontier model reviews.
Microsoft, GitHub, and Windows All Want to Be the Agent Hub
Microsoft AI is launching seven MAI models and a Superintelligence Lab in one go. Redmond wants to be less of a wrapper and show more of its own model foundation.
Microsoft is bringing Scout to Microsoft 365 as an always-on agent. CEO Satya Nadella internally called out a memo that framed Scout as an "addictive app".
Work IQ APIs arrive on June 16 and open up Microsoft 365 context for agents. That's exactly where the next enterprise step hinges: connecting company knowledge cleanly.
Microsoft's Web IQ delivers grounding APIs built on the Bing index. Search results are becoming raw material for agents.
Frontier Tuning enters Private Preview and is designed to let models work in company-specific ways. Sounds understated, but it's every ops team's dream.
GitHub is launching a Copilot Desktop App as an agent-native command center. Multiple sessions, PRs, sandboxes, review. GitHub doesn't want to complement the IDE — it wants to flip it.
Microsoft and Mayo Clinic are developing a frontier model for healthcare. Medical AI is becoming even more a question of data trust than model quality.
Microsoft is showing off Majorana 2, a scalable quantum processor. AI and quantum are moving closer together in the enterprise stack.
Microsoft is introducing Project Solara as a platform for agent-first devices. That's the more interesting hardware subtext of the week: devices are no longer designed around apps, but around agents.
Build 2026 brought Windows developer news around Trusted Development and a broader Windows AI wave. That's the security foundation for when agents start acting on your machine.
Local Models, Hardware, and Developer Tools Are Getting Serious
Ok, this block is a bit nerdier than usual 😄… but the bottom line: new hardware and new local models are making it easier to build autonomous AI infrastructure. Perfect for anyone with privacy concerns or who simply wants to stay in control.
Hermes Desktop enters Public Preview, making local agents less of a terminal toy and more accessible to the rest of us.
Google releases Gemma 4 12B as a laptop-friendly multimodal model. Local AI is getting bigger and more practical for everyday use.
Google follows up with Gemma 4 QAT, adding quantized checkpoints. Translation: on-device models aren't just getting smaller — they're being trained specifically for phones, laptops, and consumer GPUs.
MiniMax releases M3 as an open-weight model with 1M context, multimodality, and strong coding and agent capabilities. If the claims hold up, this is a serious signal for open agent models.
Perplexity announces a hybrid approach where the data center moves to your machine. That's exactly the cloud-local mix a lot of people are after.
Factory.ai launches the Factory Router for cheaper frontier coding. Routing is becoming the new cost lever: it doesn't always have to be Opus or GPT — just the right model for the task at hand.
Email platform Resend (we love it!) ships an official Claude Code plugin. Small piece of the puzzle, but exactly these kinds of integrations are what make agents genuinely useful day-to-day.
NVIDIA and Microsoft introduce RTX Spark as a new PC category for personal AI agents. The Surface RTX Spark Dev Box is a developer box built for local AI workloads.
Microsoft's Surface Laptop Ultra combines RTX Spark and local AI performance. The laptop is once again raising the question: what runs on your machine, and what runs in the cloud?
NVIDIA brings DGX Station for Windows as a desk supercomputer. Trillion parameters on an enterprise desktop? Sounds wild — and it's expensive.
NVIDIA Vera is the new CPU for agents. Even CPUs are no longer being pitched around apps — they're being pitched around agents!
NVIDIA releases Nemotron 3 Ultra, a 550B MoE model for faster, cheaper agents. Safety and ASR models for guardrails and voice agents are also included. Infrastructure is getting broader than just GPUs again.
JetBrains releases Mellum2 on Hugging Face as a fast 12B Mixture of Experts model for text and code.
Robots, Image Generation & Creative Machines
NVIDIA introduces Cosmos 3 as a world model for Physical AI. The idea: robots should be able to see, plan, and simulate better before they act.
NVIDIA unveils an open Isaac GR00T reference design for humanoid robotics. When reference designs go open, robotics moves out of the glossy lab and into the developer community.
OpenAI Robotics is hiring again for useful robots. Physical AI is no longer a side quest — it's where frontier models get a body.
TwelveLabs launches Rodeo as an AI copilot for video editing. Video editors get search, cutting, and assembly close to an agent-style workflow.
Reve 2.0 bets on layout-based image generation instead of pure text-prompt magic. When layouts become the backbone, designers finally get real control over composition rather than just pretty accidents.
Ideogram 4.0 arrives as an open-weight image model with strong text rendering and 2K output. Open weights in the design space put closed tools under pressure.
Krea 2 Turbo delivers 2-second images with style references and LoRAs. Less "wow, an image" and more "fast enough for real design iteration."
Runway brings Aleph 2.0 into a new Edit Studio. Precise video edits are exactly the part where AI video has been most frustrating — so this matters.
Media, markets, and society sort out the consequences
Gemini's new avatar feature is rolling out more broadly. Personal AI avatars feel like a novelty — until they get very serious in sales, support, and learning.
Google commits to new water stewardship goals for data centers, while New York's legislature has passed a one-year moratorium on new large data centers. Now the governor has to decide.
UK publishers can now opt out of Google's AI Overviews. When publishers reclaim control, the search peace gets shaky.
Black Forest Labs brings in Martin Scorsese as an advisor for visual AI, while Amazon's AI-generated animation series was quietly pulled after heavy mockery.
The Guardian warns about smart glasses and exam cheating. Schools won't be able to avoid tighter controls — but the bigger question is what's even worth testing anymore.
Amazon's AI-generated product images are showing up in search results. Shopping is getting more visual — and less trustworthy.
Box reports that AI has created 13 new job types. Good reminder: the jobs debate is rarely just about losses — it's often a reshuffling.
That's a wrap! But no need to be sad — the AInauten will be back soon with fresh material for you.
See you next week, with a fresh round of news, hacks, and insights!
Your AInauten
Fabian & Reto
