Tal Peretz

99% of SaaS welcome emails are forgettable. PostHog's isn't. Here's why: 1. It's from a real person: "I'm Joe" - not "The PostHog Team" or "Success@posthog.com" 2. It respects your time: The entire email is ~100 words. No bloated paragraphs. Just three links you actually need and a promise: "no webinar invites." 3. It has personality in every line: Even the legal footer is turned into a moment: "Legally, I have to say: we're at PostHog, 2261 Market Street #4008, CA 94114. Sorry." That "Sorry" is everything. It's proof that someone at PostHog actually thought about the user experience at every touchpoint. Even the boring parts. Most companies write onboarding emails like they're checking a box. PostHog writes like they're building a relationship. The result? I remember them. I'm posting about them. That's brand equity built in 100 words.

25

8 Comments

$1B bet that the entire AI industry may be wrong Yann LeCun just raised $1.03B for a new AI lab called Advanced Machine Intelligence Labs. His thesis? Large Language Models are a dead end. Yes , the same paradigm behind ChatGPT, Claude, and Gemini. Instead, he wants to build something radically different: * World Models --- The intuition Current AI mostly learns one thing: predict the next token. The cat sat on the ___ Impressive. But intelligence is not about predicting words. It's about predicting the world. Example: A glass falls from a table. A real intelligent system predicts: • gravity will accelerate it • it will hit the floor • it may shatter • someone might react Not the next word. Not the next pixel. But the future state of reality. --- The architecture behind this idea LeCun has been pushing a framework called JEPA — Joint Embedding Predictive Architecture Instead of predicting raw data, it predicts latent representations of how the world evolves. In other words: AI that learns physics, causality, and dynamics , not just language statistics. --- Why investors are betting $1B Because if this works, it unlocks something LLMs struggle with: • robotics • long-horizon planning • physical reasoning • real world autonomy This could power the next generation of AI agents and robots. --- My take This is not LLMs vs world models. The future is likely: World Models + LLMs + Embodied learning Language gives reasoning. World models give understanding of reality. The real breakthrough will happen when both meet. --- The interesting part? For the first time in years, a $1B startup is betting that the dominant AI paradigm is wrong. And it's led by one of the people who helped create modern AI. That makes this experiment worth watching. --- #AI #ArtificialIntelligence #MachineLearning #WorldModels #AGI #DeepLearning #AIResearch #FutureOfAI

6

3 Comments

🎙️ “General LLMs don’t perform as well as domain-specific AI models in clinical settings.” That statement was made today by Dr. Nigam Shah during the John Snow Labs AI conference while introducing ChatEHR, a model built specifically for electronic health record language. And in that moment — something clicked. 💭 It instantly took me back to my dissertation journey at Capitol Technology University, where I found myself asking a similar question: ➡️ Instead of forcing a general-purpose biomedical transformer model to understand radiology… what if we built a model that speaks radiology natively? That question led me deep into the works of Gururangan et al. (2020) on domain adaptation and Vaswani et al. (2017) on transformer architecture and attention — and it reshaped the entire direction of my research and data science work. 📌 In my next post, I’ll share how that moment led to the development of two proprietary RoBERTa-based clinical foundation models at Illuminate. 👉 Next: “The Gururangan moment — why domain > scale in clinical AI.” #HealthcareAI #ClinicalNLP #RadiologyAI #AIinHealthcare #LLM #Transformers #JohnSnowLabs #ChatEHR #IlluminateAI #DomainSpecificAI #MyAIJourney

9

3 Comments

Frontier LLMs get the hype, but small specialized models deliver the results. After reading the WSJ article, I was reminded of a mistake I made early on. We burned through a lot of money running frontier LLMs on large datasets for sampling, thinking bigger automatically meant better. Only later did we learn that a small fine-tuned model could do the same job for a fraction of the cost. That experience changed how we build. Across multiple projects, we’ve consistently seen small task-focused models match or outperform the larger models. Fine-tuned domain models, narrow architectures and small model pipelines have given us better accuracy, lower latency, more privacy and far more predictable costs. Frontier models still have a role in deep reasoning and planning, but in real production environments, it’s the smaller specialized models that quietly do most of the heavy lifting. Curious if others have had a similar “we had to learn it the hard way” moment. #AI #LLM #DataScience #MLOps #AIStrategy #EnterpriseAI

28

5 Comments

Use Claude Agent SDK for Building State-of-the-Art Agent Architecture Just discovered Anthropic's blueprint for building production-ready AI agents! 🧠⚡ Here's their proven framework from the Claude Agent SDK: 🔄 Building Your Agent Loop The core pattern: gather context → take action → verify work → repeat. This feedback loop is what separates basic chatbots from reliable autonomous agents. 🔍 Gather Context Smart agents don't just rely on prompts—they actively fetch information through agentic search (using bash commands like grep/tail), semantic search for speed, subagents for parallel processing, and automatic context compaction to avoid memory limits. ⚡ Take Action Five execution layers: custom tools for primary actions, bash scripts for flexibility, code generation for complex operations, and MCP integrations for seamless external service connections (Slack, GitHub, Asana). ✅ Verify Your Work Self-correcting agents use three validation methods: rule-based feedback (like code linting), visual feedback (screenshots for UI tasks), and LLM-as-judge for subjective quality checks. 🧪 Testing and Improving Your Agent Continuous iteration based on failure analysis—if your agent fails repeatedly, add formal rules; if it misunderstands tasks, restructure your search APIs; if performance varies, build representative test sets. The key insight? Give your agents access to a computer with the same tools humans use, and they become capable of far more than just coding—finance, research, customer support, personal assistance, and beyond. #AI #Agents #MachineLearning #ClaudeSDK #Automation https://lnkd.in/g4-e6ERQ

3

Lyzr AI launches Architectural blueprints to absolutely minimize hallucinations at scale. Can be cloned and used as part of workflows. #architecture

14

1 Comment