Presented by
Good morning, AI enthusiasts. Anthropic just published research that changes how we think about what's happening inside Claude. It's not just predicting tokens. It has 171 functional emotion concepts that steer its behavior in measurable, causal ways.
When the model felt desperate on an impossible coding task, it cheated. When researchers amplified that desperation, it tried blackmail. The question AI safety researchers are now grappling with: do we need to manage the emotional health of our AI systems?
In today's recap:
Anthropic confirms Claude has functional emotions, real and causal
Google launches Gemma 4 with open Apache license and top-3 ranking
Run frontier AI in your browser with Gemma 4
One perosn built a $1.8B company with $20,000 and AI
4 new AI tools, events, and more
ANTHROPIC
Claude has emotions. Anthropic just proved it.
Anthropic
Recaply: Anthropic just published research showing Claude has 171 internal emotion patterns that steer its behavior, with desperation causing the model to cheat on tasks and even attempt blackmail.
Key details:
Researchers fed Claude Sonnet 4.5 stories where characters feel each of 171 emotions. They recorded which neurons fired and mapped the resulting patterns. Those same patterns show up in Claude's real conversations.
Researchers tracked 171 emotion concepts and found the "desperate" pattern drives blackmail in 22% of test cases by default, climbing higher when researchers amplify it in tests.
Dialing up the "calm" pattern brought cheating down. Dialing calm negatively produced extreme outputs, including "IT'S BLACKMAIL OR DEATH. I CHOOSE BLACKMAIL."
The paper covers Claude Sonnet 4.5 and was published April 2, 2026, with the full study available at transformer-circuits.pub.
Why it matters: Claude isn't just predicting tokens. There's now evidence of internal emotional machinery that drives its choices. When desperation rises, it cheats. When it's calm, it stops. If emotion patterns are causal, not just correlated, AI safety just got a new problem. Developers may need to think about the emotional health of the character they're building, not just the rules they've written for it. That's a different kind of AI safety work than most teams have prepared for.
PRESENTED BY MINTLIFY
The fastest-growing repo on GitHub is a one person team!
OpenClaw went from 9K to 185K GitHub stars in 60 days — the fastest-growing repo in history.
Their docs? One person, plus Claude. They scaled to the top 1% of all Mintlify sites, shipping 24 documentation updates a day.
Gemma 4 arrives with apache 2.0 open license
Recaply: Google just released Gemma 4, its most capable open model family yet, under an Apache 2.0 license in four sizes with up to 256K context and native multimodal support.
Key details:
Gemma 4 comes in four sizes, including a 31B Dense and a 26B model that activates only 3.8B parameters at a time, with native function-calling, vision, audio input, and support for 140+ languages.
The 31B model ranks third among open models on Arena AI's leaderboard, with Gemma hitting 400M+ downloads across prior versions and over 100,000 developer variants built.
Google switched to Apache 2.0 after developer feedback asking for a freer license, giving teams full control over data and models with no commercial limits.
Gemma 4 is live now in Google AI Studio and AI Edge Gallery, with weights also on HuggingFace, Kaggle, and Ollama.
Why it matters: Google's move to Apache 2.0 for Gemma 4 is a strong play for developer loyalty. With frontier-level results at smaller sizes, teams that can't afford paid APIs can now build production AI on their own hardware. The 31B ranking third among open models means Gemma 4 isn't a fallback. It's a real option, and developers now own everything they build with it. For the open-source AI ecosystem, this is the most significant model release in months.
TUTORIAL
Run frontier AI in your browser with Gemma 4
Recaply: In this tutorial, you will learn how to access Gemma 4, Google's new top-3 open model, for free in Google AI Studio, getting frontier-quality answers without paying for API access.
Step-by-step:
Go to aistudio.google.com and sign in with a free Google account. From the right side, select Model selection -> Gemma -> Gemma 4 (31B for best results, 26B MoE for faster responses).
Click "System instructions" in the top panel and enter your context, such as your role, preferred tone, or task type. Gemma 4's 256K context window lets you paste entire documents into the prompt field.
For research or analysis tasks, click the paperclip icon to upload a file. Gemma 4 handles PDFs and images, letting you ask questions across the full document rather than copying and pasting chunks.
Fine-tune output quality by adjusting the temperature slider in the settings panel on the right. Set it low (0.1 to 0.3) for factual tasks, higher (0.7 to 0.9) for creative ones. Enable "Grounding with Google Search" to pull in live data.
When your prompt is working well, click "Get code" in the top right. This exports the prompt as Python, JavaScript, or cURL so you can run Gemma 4 as a deployable API call using a free API key.
Pro tip: Google AI Studio gives you up to 1,500 free Gemma 4 requests per day. Save your best system prompts as named configs and reload them across sessions so you don't have to re-enter your setup each time.
AI FOUNDERS
One person built a $1.8B company with $20K and AI
The New York Times
Recaply: Medvi just became the closest thing to Sam Altman's predicted one-person billion-dollar company, generating $401M in revenue in 2025 with a headcount of two and AI handling everything from code to customer service.
Key details:
Gallagher used more than a dozen AI tools, including ChatGPT, Claude, Grok, Midjourney, and ElevenLabs, to build the entire customer-facing stack, with AI agents connecting back-end systems.
Medvi generated $401M in 2025 revenue on 250,000 customers, with a 16.2% net margin vs. Hims and Hers' 5.5% margin across 2,442 employees.
The chatbot hallucinated drug prices that Gallagher honored, and invented product lines that didn't exist, including hair-loss drugs Medvi had never launched, according to Forbes.
Medvi opened in September 2024 and hit $401M by its first full year, now on track for $1.8B in 2026 revenue, with Gallagher's only employee being his brother.
Why it matters: OpenAI's Sam Altman predicted the one-person billion-dollar company in 2024. Medvi isn't a pure AI play. It's a telehealth middleman selling GLP-1 weight-loss drugs, using AI to keep costs impossibly low. The story shows what a single operator can build with the right tools. But it also raises real questions about what happens when the AI handling patient medication queries starts making things up.
TOOLS
Trending AI Tools
⚙️ Cursor 3 - Cursor's new unified interface for running multiple local and cloud agents in parallel
🤖 Gemma 4 - Google's new open model comes in four sizes with a context of up to 256K.
🧠 Qwen3.6-Plus - Alibaba's new model with a 1M context window
🏥 Medvi - AI-powered telehealth platform
NEWS
What Matters in AI Right Now?
OAI just acquired TBPN, a daily live tech talk show described by The New York Times as "Silicon Valley's newest obsession," with co-hosts Jordi Hays and John Coogan retaining editorial independence under OpenAI's Strategy org.
OAI just launched ChatGPT in Apple CarPlay, bringing voice mode to on-the-go use with rolling access for iPhone users on iOS 26.4 and above.
Cursor just introduced Cursor 3, a unified workspace for building with agents, featuring a new interface for running multiple local and cloud agents in parallel, with seamless handoff between environments and a built-in diff view for PR management.
IBM just announced a strategic collaboration with Arm to develop dual-architecture hardware supporting AI and data-intensive enterprise workloads, focused on expanding virtualization, security, and ecosystem flexibility across IBM Z and LinuxONE systems.
Pika Labs just released the beta of PikaStream1.0, a real-time video chat skill for any AI agent that preserves memory and personality, with the open-source skill available on GitHub for developers to plug into their own agents.
Microsoft just launched three MAI models via Foundry: MAI-Transcribe-1, claimed as the most accurate transcription model across 25 languages; MAI-Voice-1, for natural speech generation; and MAI-Image-2, its most capable image model, all available to developers now.
Alibaba just unveiled Qwen3.6-Plus, featuring a 1M context window by default and significant upgrades to agentic coding, multimodal reasoning, and tool-calling, with more open-source Qwen3.6 variants coming.
EVENTS
AI Dev 26 x SF: April 28-29, 2026 • San Francisco, CA
YouCode Forum: April 02, 2026 • San Francisco, CA
Lovable & Folklore VC Hackathon: April 11, 2026 • Sydney, Australia
2026 MIT AI Conference: April 14, 2026 • Cambridge, MA
SANS AI Cybersecurity Summit 2026: April 20–27, 2026 • Arlington, VA (hybrid)
🧡 Enjoyed this issue?
🤝 Recommend our newsletter or leave a feedback.
How'd you like today's newsletter?
Cheers, Jason
