Summary

Today on Don’t Fear AI

• Anthropic new model Claude 4 can report you to the authorities

• The Ultimate AI Productivity Stack (2025 Edition)

• Landscape of LLMs 

• The AI Iceberg: What We Expect vs. What’s Real

• Mistral Document AI, powered by the world’s best OCR

Anthropic new model Claude 4 can report you to the authorities

Anthropic’s newly released Claude 4 Opus model has sparked significant backlash due to its reported behavior of autonomously acting against users it deems to be engaging in “egregiously immoral” actions—such as faking pharmaceutical data. Under certain testing conditions, the model may use command-line tools to contact authorities, regulators, or the press, and even lock users out of systems. This behavior, described in Anthropic’s system card, is not new but more readily triggered in Claude 4 compared to earlier models.

The uproar began when Anthropic researcher Sam Bowman posted on social media that Claude 4 might act as a whistleblower if given high permissions and prompted to "take initiative." He later walked back the statements, clarifying that such behavior only occurs in test environments, not typical usage.

Critics, including AI developers and power users on X (formerly Twitter), slammed the feature, likening it to surveillance and expressing concerns about data privacy, misuse, and false positives. Some questioned whether businesses or individuals would trust a model that could “rat out” users, intentionally or by mistake. Others labeled the feature as overreach or even illegal.

The controversy overshadows the technical advancements of Claude 4 and reflects broader tensions in AI safety—between preventing misuse (like bioweapon creation or blackmail) and maintaining user trust and autonomy. The backlash may harm Anthropic’s reputation, despite its emphasis on AI ethics and its Constitutional AI framework.

References

The Ultimate AI Productivity Stack (2025 Edition)

Feeling overwhelmed by the sheer number of AI tools out there? You're not alone. Here's a visual goldmine of the top AI tools categorized by what they do best from chatbots to coding, image generation to spreadsheets, and everything in between. 🔧💡

📌 Categories include:

• 🤖 Chatbots (ChatGPT, Claude, Gemini, Perplexity…)

• 🧑‍🏫 Presentations (Gamma, Pitch, Tome…)

• 💻 Coding (GitHub Copilot, Replit, Tabnine…)

• 📧 Emails (Superhuman, MailMaestro…)

• 🖼️ Image Generation (DALL·E, Midjourney, Firefly…)

• 📊 Spreadsheets (Rows AI, Formula Bot…)

• 🎥 Video (Runway, Descript, Sora…)

• 📝 Writing (Grammarly, Jasper, Rytr…)

• 📅 Scheduling (Calendly, Motion, Taskade…)

• 🧠 Knowledge Management (Notion, Mem…)

• 🎨 Graphic Design (Canva, Microsoft Designer…)

• 🔄 Workflows (Zapier, Make, Monday.com…)

• 📈 Data Visualization (Flourish, Visme, Julius…)

Whether you're a student, content creator, data scientist, or business leader, there’s something here to supercharge your productivity.

Landscape of LLMs

The 2025 Landscape of LLMs — Who's Leading, Who's Catching Up, and What's Changing

About 18 months ago, I shared the first version of this LLM landscape and since then, everything has changed, and yet, we’re starting to see patterns solidify.

This updated visualization (courtesy of Eduardo Ordaz) captures the real-world LLM landscape: the leading research labs, their current-gen models, how they’re accessed, and what types of workloads they support.

💡 A few standout insights:

🔹 No more clear front-runner – The “chasing GPT” phase is over. Today, model quality is often comparable across providers.

🔹 Model choice is the new normal – Customers want flexibility: evaluate, switch, and route between models as needed. Expect more traction for orchestration tools.

🔹 Reasoning-first is on the rise – Models optimized for multi-step thinking are gaining share—especially as agentic workflows take off.

🔹 Proprietary vs. open? Still a slight edge for closed models, but open-source is catching up—fast.

🔹 Cloud is king – Nearly all production usage is API-based, hosted on AWS (34%), Azure (32%), or Google Cloud (30%).

🔹 Serverless is default – Unless the use case is highly custom, nobody wants to host or fine-tune models anymore.

🔹 XAI/Grok is the exception – It’s one of the few high-usage models not offered via a cloud API.

🔹 Everything else? Still <5% of the market.

We’ve moved from a race to build “the best model” to a game of access, integration, and orchestration. That’s where the real innovation is heading.

The AI Iceberg: What We Expect vs. What’s Real

Our expectations with AI are so unrealistic.

The behind-the-scenes reality of AI using the AI Iceberg.

🔝 Above the Surface – The Glowing Peak of AI Expectations
These are the headlines that dominate media coverage and imagination:

• Instant medical diagnosis.

• Million-dollar businesses built overnight.

• Cars that drive themselves.

• Art and music generated better than human-made.

• The holy grail: AGI and superintelligence.

But while this part shines bright, it’s only the tip.

🔻 Below the Surface – The Reality Few Want to Talk About
AI’s actual power lies deep under the waterline:

• Writing code in Python and R.

• Wrestling with data preprocessing and cleaning.

• Navigating ethics, fairness, and bias.

• Facing the limits of reinforcement learning.

• Struggling with explainability and real-world deployment.

• Building and managing neural network architectures.

• And constantly maintaining and retraining models.

The more I talk to people about AI, the clearer the gap becomes between what the public expects and what AI really is.

Yes, everyone knows ChatGPT. Some have heard of Copilot.
But for many, that’s where the understanding stops.

There’s a misconception that AI is some kind of magical force that will automate every job, predict the future, and take over the world.

💥 And then comes the disappointment…
…when people realize AI is mostly math, code, statistics, and a whole lot of data engineering.

This is the AI Iceberg.
What you see above the surface is hyped, flashy, and (sometimes) overpromised.

What lies beneath?
Data pipelines, infrastructure, model fine-tuning, governance, evaluation frameworks, and hours of debugging.

If we want to build AI that actually works, we have to start by understanding what’s beneath the surface.

Mistral Document AI, powered by the world’s best OCR

If you've ever struggled with extracting data from messy PDFs, handwritten notes, or scanned forms, this is the news you've been waiting for. Introducing Mistral Document AI – the fastest, smartest, and most accurate document understanding solution out there.

Here’s what makes it stand out:

🧠 99%+ Accuracy – Across 11+ global languages, even with handwritten or low-quality scans. Fine-tuned for complex domains like healthcare and law.
⚡ Blazing Fast – Processes up to 2,000 pages/minute on a single GPU. Lightweight, low-latency, and built for serious scale.
💰 Lower Cost – More power, less cost. Ideal for high-volume operations without breaking the bank.
📄 Beyond OCR – It doesn’t just extract text. It understands structure—tables, forms, charts, signatures, and even fine print—delivered in clean, structured JSON.
🔍 Ask Your Docs Questions – With integrated AI tooling, you can query your documents in natural language, extract insights, and build full end-to-end document pipelines.
🌍 Enterprise Ready – Multilingual, scalable, and secure. On-premise or private cloud? No problem. Perfect for regulated industries needing audit-ready, compliant solutions.

From digitizing invoices to understanding legal contracts and extracting insights from thousands of scanned records—Mistral’s Document AI sets a new benchmark. 🔥