Hey, it's Andreas.
After 8+ years at IBM, I decided a few weeks ago to move on.

IBM is a great company, and honestly one of the most interesting places to work right now if you are in AI. No other company combines a technology business, serious R&D, and a services and consulting arm under one roof. During my time there I worked with close to 100 companies and helped train and enable more than 10,000 people on AI around the globe. But it was time to move on. I wrote down my key lessons on LinkedIn, and they resonated far beyond what I expected. I'm sharing them with you at the bottom of this issue.

One more thing: my Agentic AI Cohort starts today (I am super excited for it). Enrollment is closed, but if you want to build agents with me live in the next round, join the waitlist here.

Now, on to this week.

In today's issue:

🌀 Anthropic brings Fable 5 back worldwide → Export controls lifted. Anthropic redeployed its top public model with new cyber safeguards, launched Claude Science, and made Sonnet 5 the default model.

🌀 Microsoft launches $2.5B Frontier Company → 6,000 experts will embed inside enterprises to build AI systems customers own, using Microsoft, OpenAI, Anthropic, open-source, or industry models.

🌀 BMW puts Figure 03 to work in factory logistics → After Figure 02 supported production of 30,000 BMW X3s, the third-generation humanoid is now sequencing parts at Plant Spartanburg. BMW is turning a successful pilot into a wider Physical AI programme across its production network.

🌀 Meta says Watermelon has matched GPT-5.5 → Still training on 10x Muse Spark’s compute, Meta’s next model (called Watermelon) reportedly reached parity on undisclosed internal benchmarks.

🌀 Meta explores selling AI compute → Meta may rent raw capacity or sell access to hosted models, creating a second return path from its infrastructure buildout even if its own models do not lead the market.

🌀 Anthropic launches Claude Science → The new research workspace connects 60+ scientific tools, generates auditable outputs, and runs across local or cloud compute. Anthropic plans also to expand into AI-assisted drug and vaccine discovery for neglected diseases.

📘 Must read: Anthropic releases new Fable 5 prompting guide → Fable 5 performs best with clear goals, boundaries and acceptance criteria, not step-by-step micromanagement. Anthropic also recommends auditing claims against tool results and reserving xhigh effort for the hardest tasks.

📘 DeepLearning.AI launches voice agents course → This 86-minute course with Vocal Bridge shows how to add voice to existing agents, build outbound-calling workflows, and evaluate conversations before production.

📘 Phil Chen on career skills for the AI age → As AI commoditizes clearly defined work, the edge shifts to problem selection, systems thinking, judgment, relationships, and reputation - skills rarely taught in school.

📘 Google Cloud on evaluating multi-agent systems → A practical walkthrough for building continuous evaluation pipelines with Gemini, including adaptive rubrics, shadow deployments, and CI/CD checks to catch regressions before production.

📘 Ex-Meta Principal Engineer on agentic coding setup → Kun Chen shows how to build a reproducible Mac development environment , so the full agentic engineering stack can be rebuilt from scratch.

🔹 Alex Karp on owning the means of AI production → Karp argues that companies lose strategic control when critical workflows depend on the same closed models as competitors.

🔹 a16z on vibe coding’s localhost problem → Building software is becoming cheap, but distribution, trust and real customer demand remain hard. The moat is no longer shipping an app - it is getting people to use it, trust it and pay for it repeatedly.

🔹 SJTU asks if agent memory is ready for production → After testing 12 memory systems across 11 datasets, researchers found no single architecture wins everywhere. Performance depends on matching memory design to the workload, while local updates often beat costly full-memory reorganisation.

🔹 Bridgewater proves specialized models can beat the frontier → A Qwen3 model trained on Bridgewater’s expert judgment reached 84.7% accuracy, outperforming the best frontier model at 13.8x lower cost.

🔹 Sam Altman calls for a global AI referee → He proposes a U.S.-led forum to set safety standards and govern access to frontier models, as OpenAI separately explores giving the U.S. government a 5% stake to share AI’s economic upside.

I just left IBM after 8+ years of AI projects across the globe. I started very basic: building chatbots with NLP, spending nights hand-tagging and labeling words, for results that wouldn't survive five minutes against today's systems. From there to agents running multi-step work on their own, in under a decade.

So what did 8+ years, dozens of projects, and 10,000+ trained people actually teach me?

Seven things.

1 - Most companies think buying the technology is the hard part. It isn't. Success is roughly 10% the technology (AI) itself, 20% your tech and data, and 70% your people and how you work. Almost everyone gets that backwards and pours all their effort into the 10%.

2 - Data is the lifeblood of AI, and most companies skip it. AI doesn't fix a messy organization, it speeds it up. If your information is scattered and no one trusts it, no model will save you. Clean it up first, or you just make the mess faster.

3 - AI can now do real multi-step work on its own (the "agentic" shift everyone's talking about), and that exposed a new problem. The technology is no longer what holds you back, your old processes are. Drop a capable agent into a slow, bureaucratic workflow and it gets stuck in the same traffic everyone else does.

4 - You can't figure out what AI is useful for by sitting in a meeting. You have to try it, fail a bit, and learn by doing. The people getting ahead aren't the ones with the perfect plan. They're the ones running lots of small, cheap, slightly embarrassing experiments.

5 - Nobody fully knows what this technology is best at yet, not even the people who built it. That's the opportunity. Whoever figures out how to use it in their own field first gets a real head start, because there's no manual to wait for.

6 - A new kind of valuable person is emerging, and they don't have a job title yet. They're not the best at typing prompts. They're the ones who can take a fuzzy problem, define what a good result looks like, and direct AI the way a manager directs a small team.

7 - Technology never drives change on its own, people do. Nothing I ever built made a difference until someone inside the business decided to own it and push it forward.

Now, what's next?

I've never been more excited to work in AI than I am right now, and leaving opens something new: working directly with you. Every week I get questions about AI adoption and enablement. After training 10,000+ people and leading AI transformations for Fortune 100 companies, I know what works and what doesn't. I'm limiting this to 3 companies per quarter (minimum 50 people), and two spots are already taken. If you want to figure out AI for your organization, whether that means Claude, Copilot, or both, reply to this email or reach out on LinkedIn.

A lot more to come. Let's get to work.