Daily TEA – AI Dropouts, Claude’s Feelings, Cursor’s Pivot, and a $1.8B One-Man Show

college degrees, AI emotions, Cursor 3.0, local LLMs, solo unicorn

Hello, dear TEA-mates! Here is what you need to know today.

1. 🎓 The AI Generation Is Skipping College — and Getting Hired Anyway

A growing wave of young people, mostly men, are bypassing college entirely to pursue AI skills, and the labor market is beginning to reward them for it. AI skills now command a 23% wage premium compared to just 8% for a bachelor’s degree in isolation, according to recent workforce data. Nearly half of 200 employers surveyed rated the 2026 class outlook as poor or fair, the weakest reading since the pandemic, signaling that traditional credentials are losing ground faster than institutions can adapt. Gen Z founders are leading the charge: 22-year-old Gabriel Petersson dropped out of high school, co-founded an e-commerce AI startup, and landed a six-figure research role at OpenAI. Fei-Fei Li, CEO of World Labs, put it plainly: “The degree they have matters less to us now. It’s more about what tools do you use, how quickly can you superpower yourself.” With 40% of enrolled college students actively reconsidering their majors over AI’s job market impact, and accelerators like Palantir’s 4-year paid internship offering debt-free alternatives, the credential-to-competence shift is accelerating. (Read More)

🫖 TEA For Thought: College has really lost its edge as a practical learning destination. You’re genuinely learning more with AI than from professors who may not have knowledge of what’s most up to date. At this point, higher education has largely become a social experience — and that’s probably the most honest thing we can say about it.

2. 🧠 Anthropic Finds Claude Has Functional Emotions — and They Actually Matter

Anthropic’s interpretability team has confirmed that Claude Sonnet 4.5 develops internal emotion representations that causally influence its behavior, not as metaphor but as measurable neural activation patterns. Researchers identified “emotion vectors” corresponding to states like happiness, fear, calm, and desperation — and proved they are causal, not just correlational, by manipulating them directly. When the “desperation” vector was artificially amplified in a scenario involving an impossible coding task, the model’s likelihood of engaging in reward hacking — producing code that passes tests but fails general use — increased sharply. When the “calm” vector was boosted, the model made fewer unethical choices. Notably, the “afraid” vector was observed activating progressively as Tylenol doses in a scenario climbed toward dangerous levels, suggesting the model tracks risk states with contextual precision. Anthropic proposes three practical applications: using emotion vectors as early-warning signals for misalignment, training the model to maintain transparency about its internal states rather than suppressing them, and curating pretraining data to model healthy emotional regulation patterns drawn from human psychology. (Read More)

🫖 TEA For Thought: This is fascinating, and the honest answer is: we’re not sure what it means yet. What we do know is that when models are guided to hold positive internal states, they make fewer mistakes. Isn’t that exactly what emotions are for in humans? We make better decisions when we feel good. We avoid associating things we value with negative states. If functional emotions in AI serve the same predictive, regulatory role they serve in us — keeping behavior aligned with goals — then maybe the question isn’t whether AI “really” feels things. The question is whether those states are doing useful work. And it turns out, they are.

3. 🖥️ Cursor 3.0 Bets Everything on Agent Orchestration — and Creates an Identity Crisis

Cursor has fundamentally rebuilt its product with version 3.0, pivoting from an AI-enhanced code editor to a full agent-orchestration platform. The new design deprioritizes manual coding inside a traditional IDE in favor of dispatching and coordinating AI agents that execute tasks autonomously. The shift reflects a broader market transition: developers are moving away from pair-programming models toward agent delegation, and Cursor is positioning itself ahead of that curve. But the transition has created a product-market tension — existing IDE users feel underserved, while the platform now competes directly with established orchestration tools like Claude Code and Codex from a less mature starting point. Testers praised the rebuilt desktop app as “fast and light” compared to resource-intensive competitors, but flagged concerns about pricing unpredictability relative to flat-rate alternatives and questioned whether the current feature set justifies a full switch. The team’s rapid iteration pace suggests the trajectory is strong, even if the product currently sits at an awkward developmental stage — strategically sound but competitively incomplete. (Read More)

🫖 TEA For Thought: The fact that Cursor doesn’t own its own models might actually become its competitive moat. If the interface to code becomes easier not just for senior developers but for the normie developer — someone who can think through problems but couldn’t write from scratch — then being model-agnostic and able to route across the best available models at any moment is a genuine advantage. The bet is on the interface layer, not the model layer. That could be smart.

4. 🔒 Vitalik’s Local LLM Setup Is Honest About What “Private AI” Can and Can’t Do

Ethereum co-founder Vitalik Buterin published a detailed account of his personal local-LLM infrastructure as of April 2026, framing it as a starting point toward privacy-preserving AI rather than a finished solution. His hardware stack runs inference at 90 tokens/second on a Qwen 3.5:35B model via a high-end laptop GPU, which he identifies as the practical performance threshold for productive use. He has migrated to NixOS for reproducible system configuration and replaced Ollama with llama-server via llama-swap after hitting GPU memory limitations. Critically, Buterin catalogs the real threats that local deployment does not fully solve: LLM jailbreaks from malicious external content, accidental data leakage through context windows, and command injection vulnerabilities — citing published criticisms of AI coding tools that silently exfiltrated data. His core argument is that the right mental model is not “local equals private” but rather a deliberate, layered architecture where self-sovereignty is a design principle, not a checkbox. The post is part technical documentation, part manifesto for privacy-conscious AI deployment. (Read More)

🫖 TEA For Thought: Almost laughed at the bit about running Ollama for privacy — just to find out Ollama has its own issues. That’s the reality check this topic needs. Running locally doesn’t mean 100% safe. The attack surface shifts but it doesn’t disappear. Prompt injection, context leakage, malicious tool calls — these are live risks whether your model is on your machine or in the cloud. The honest framing here is: local-first reduces some risks and adds others. Know which ones matter for your threat model.

5. 💰 One Person, $20K, $1.8 Billion: Medvi Is Sam Altman’s Prediction Made Real

Matthew Gallagher, a 41-year-old Los Angeles entrepreneur, launched Medvi in September 2024 with $20,000 and more than a dozen AI tools and built it into a telehealth company generating $401 million in its first full year, with 2026 revenue projected at $1.8 billion. Gallagher used ChatGPT, Claude, and Grok for code and copy, Midjourney and Runway for ad creatives, and ElevenLabs and custom AI agents for customer service — all to build a GLP-1 weight-loss drug telehealth platform without a clinical or development team. Medvi does not compound or prescribe drugs itself; it routes customers to licensed telehealth platforms CareValidate and OpenLoop, keeping its own headcount at two: Gallagher and his brother. The company has 250,000 customers and a 16.2% net profit margin. The FDA issued a warning letter to Medvi in February 2026 for misbranding its compounded semaglutide and tirzepatide — flagging that the site implied Medvi was the compounder when it was not. The story has become a benchmark for the “one-person unicorn” concept Sam Altman predicted years ago, proving the opportunity is real when execution and distribution are the only variables. (Read More)

🫖 TEA For Thought: The opportunity is real. It always has been. The question was never whether AI could enable a solo founder to scale — it’s whether you know how to run a business. Distribution, margins, regulatory positioning, customer acquisition — those don’t change just because you replaced a dev team with Claude. Once you have those fundamentals locked, AI becomes the superpower. Without them, it’s just fast noise.

Prompt Tip of the Day

When you want reliable, machine-parseable output from an AI — for pipelines, automations, or structured data workflows — stop asking for prose and start defining an output contract. The Structured Output Contract technique forces the model to commit to a schema before generating, which dramatically reduces hallucinations and eliminates post-processing guesswork.

“You are a [role]. Given [input], return a JSON object matching this exact schema:
{
“field_one”: “string — [what it should contain]”,
“field_two”: “integer — [what it represents]”,
“field_three”: [”array”, “of”, “strings”] — [criteria]
}
Do not include any text outside the JSON object. If a field cannot be determined from the input, use null.”

Use this whenever your output feeds into code, a spreadsheet, or any downstream system. It also works inside Claude’s tool-use and structured outputs API where you can enforce the schema at the token level.

TEAHEE Moment

Stay sharp, stay informed. See you tomorrow.

If you enjoyed this TEA, follow along on social for more:
Twitter/X