Hey, it’s Andreas.
I’m currently rethinking where to take this newsletter next and how to make it more useful for you.

Over the past few days, I’ve been reviewing feedback, going through emails, and collecting ideas about what this newsletter should become. Unfortunately, I also lost a large number of replies after my OpenClaw bot, "Lamine," accidentally deleted most of my emails - a painful reminder to be careful with permissions. Some of those messages included genuinely thoughtful notes, event invitations, and university speaking requests, which was frustrating.

So, if you ever replied to this newsletter and didn’t hear back from me, or if you’ve been meaning to share feedback, please send it right now here: [email protected]

I care a lot about making this better and I’d genuinely love your candid input:
What do you want to see more of?

In today’s issue:

Let’s get into it.

🌀 Google upgrades Stitch → Google upgraded its “vibe design tool”. It’s quite fun and extremely good. I spent 5 hours with it yesterday. I think there's no better approach for front-end design.

🌀 OpenAI plans major hiring push → OpenAI reportedly wants to double its workforce to 8,000 employees by year-end.

🌀 Mistral unveils Forge → Mistral Forge lets enterprises and governments build custom AI models from scratch on their own data.

🌀 Microsoft releases MAI-Image-2 → Microsoft’s AI Superintelligence team just dropped MAI-Image-2, a text-to-image model that reached No. 5 on the Arena AI leaderboard.

🌀 Aristotle launches Aristotle Agent → Aristotle Agent is a fully autonomous mathematician that can work on hard research problems for up to 24 hours without human intervention.

🌀 U.S. charges three Super Micro-linked employees → The U.S. Justice Department charged three people tied to Super Micro with illegally routing about $2.5B in advanced AI servers to China, in a major export-control crackdown.

🌀 Elon Musk unveils Terafab → Musk introduced Terafab, a Tesla-SpaceX-xAI chip manufacturing push aimed at producing 1 terawatt of AI compute per year.

🌀 Anthropic gives Claude desktop control → Anthropic released a research preview that lets Claude operate a Mac directly and assign tasks from phone via Dispatch. Combined with the recent Vercept acquisition, it looks like Claude is steadily turning into a real remote work agent. 

📘 Microsoft launches AI Envisioning Days → Free video series for software companies on how to build, deploy, and monetize AI apps and agents.

📘 365 Data Science opens its AI learning platform for free → For a limited time, 365 Data Science unlocked full access to 126+ AI and data courses, including the possibility to gain certifications.

📘 Traditional RAG vs Agentic RAG → Good visual if you are building AI agents and want to understand why Agentic RAG is replacing traditional RAG.

📘 Master Claude Code cheat sheet → A dense one-page guide covering Claude Code workflows, commands, setup, skills, MCP, and prompting.

🔹 OpenAI chases the automated researcher → In a new conversation with chief scientist Jakub Pachocki, OpenAI makes clear that automated research is now one of its biggest bets. Worth reading if you want to understand where frontier labs think the next leap in agent capability will come from.

🔹 Terence Tao on how top mathematicians use AI → A strong conversation on where AI helps in science, and where judgment still does the heavy lifting. Good reminder that discovery is not just about faster verification loops, but about taste, intuition, and staying with ideas long before they can be fully proven.

🔹 Anthropic engineer shares a Claude Code playbook → One thread collecting Claude Code guides on skills, agents, and practical workflows. Strong starting point for anyone trying to move from casual use to serious building.

🔹 Anthropic on what 80,000+ people actually want from AI → Anthropic’s global study shows most people do not just want productivity gains from AI - they want help living and working better. The deeper insight: hope and fear are rising together, which feels like an early sign of the shift from co-intelligence tools to systems people increasingly hand work over to.

🔹 Andrej Karpathy on the post-code era → Karpathy says he has not written a line of code since December 2025. His podcast with Sarah Guo is a must-watch for everybody into agentic coding.

🔹 Jensen Huang on NVIDIA and the AI revolution → Lex Fridman’s new podcast with Jensen Huang is worth the listen if you want a first-hand view into how NVIDIA sees the next phase of AI infrastructure, scale, and system design.

Today I want to give the stage to some of my IBM colleagues, because I know how much work went into this.

IBM just opened the trial for IBM Bob. For those who have not come across it yet, Bob is IBM’s AI software development partner and AI-first IDE, built to help teams work faster across real codebases with awareness of intent, repository context, and enterprise security standards (you can try it out here for free). A big team was working on that for quite some time and there was plenty of internal testing ongoing over the last year.

Now, as many of you know, I am a big fan of Claude Code, and I still see it as one of the leaders in this space. But what makes IBM Bob interesting to me is not that it tries to win the entire AI coding conversation. It is that it seems built for a part of the market that is often misunderstood and usually reduced to simplistic talking points: legacy modernization at scale.

That distinction matters a lot. Because the hardest enterprise problem was never simply writing code faster, and it was never just translating old code into a newer language. In fact, translating code is one thing, while modernizing a platform is something else entirely. The real challenge sits in the surrounding system, from runtime and integrations to transaction integrity, resilience, security, and the operational architecture the business actually depends on. And IBM Bob is built with a strong focus on surrounding systems, incorporating a wealth of IBM proprietary & domain knowledge.

The most interesting arena for tools like this is not the clean greenfield demo. It is the crown jewels: mainframes, deeply embedded Java estates, regulated environments, and large codebases where the cost of misunderstanding context is incredibly high. IBM’s public positioning around Bob is already leaning in that direction, describing it not just as a coding assistant, but as a development partner that supports work from design to deployment and across modernization efforts. 

That is also why a lot of the current AI code debate still misses the point. The market keeps rewarding demos that make code generation look easy, while enterprises are dealing with a very different reality. Their bottleneck is not a missing snippet. It is the complexity of the system around the snippet. The teams that matter most are not asking who can write ten lines faster. They are asking which tools can help them operate inside environments that carry decades of business logic, compliance pressure, architectural coupling, and modernization risk.

That is the lens through which I see IBM Bob.

Not as a generic entrant in the AI coding race, but as a potentially important tool for one of the hardest and most valuable parts of enterprise engineering.

If you work in enterprise AI or around legacy systems, I bet this problem will sound familiar. That alone makes it worth a closer look.

Good to see the team getting this finally out into the world.

Try it here for free: https://bob.ibm.com/trial
Learn more: https://bob.ibm.com/