It's a great story. He never gave up. He worked on a problem he personally had, for years. Even when success was not clear. Even with many rejections.

Then Paul Graham himself recommended a younger Sam Altman (who was running the accelerator) to meet Amjad.

Now Replit is valued at over a billion dollars with 35+ million users.

That meant he had to reinstall every time the needed tools, so he thought it'd be great to have the development environment directly on the web.
That constant frustration sparked Replit.

He was rejected four times by Y Combinator.

And last year they released their Replit Agent that can write an application by just describing it.

The story of Amjad Masad is fascinating.

He grew up in Jordan, coding at an internet café.

What's your bet, will 118 dollars per NVIDIA stock will be remembered as a bargain or the tip of a bubble?

• Companies like Google, for example, have recently been releasing very capable and cost-effective models.

• OpenAI’s O3-mini will likely be much cheaper than O1, potentially as inexpensive as DeepSeek’s R1.

• It's very unlikely that DeepSeek has a significant technical advantage. Experts building top AI models frequently move between companies, so information sharing is higher than one might assume.

Here are some reasons why not much changed:

• DeepSeek released extensive details about their training process, which suggests they believe there isn’t much that’s truly proprietary about it.

Investors thought: Oh, so AI doesn't need so many NVIDIA GPUs.
SELL, SELL, SELL! 📉

There were concerns even before the release of R1 about whether NVIDIA is truly worth its multiple TRILLION-dollar valuation.

I am not sure DeepSeek's model changes whether it is or isn't.

What job do you think he took up after alarm clocks?

Focus on steering innovation in the right direction, not limiting it.

Warm showers were not widely available just 100-150 years ago.
In a few decades, people will not understand how we could spend so much time on boring and repetitive tasks.
Let's move to the next level of abstraction.

Then make as many experiments as possible in as little time as possible.
Enjoy the compounding of every single learning.

✅ Learning is the goal. Every experiment, successful or not, adds to your team’s collective knowledge.
✅ Prioritize knowledge sharing. Create systems to document and share insights so the entire team grows together.

Input is as important as the output. Measure the latter.

A few libraries to not miss:
- FastAPI
- @LangChainAI Langgraph for Agent Orchestration
- @firecrawl_dev and @IBMResearch docling for scraping

What's your hidden gem or contrarian opinion?

Deepseek v3
- Cost-effective API, GPT-4o level capabilities

Gemini @Google
- Handles extremely long context
- Strong video understanding

Development:
- @cursor_ai as IDE and v0 to prototype frontend
- @vercel and @railway for deployment
- @supabase for data storage

You saw people talking about AI Agents like the defining topic of 2025.
Agentic UI/UX is what makes those agents usable. And useful.

• Pure chatbots are often not the best answer. There is a fatigue related to having to text everything. Clicks and "standard" UI are great for many actions!
• Structured forms or guided steps might be smoother, saving token calls and frustration.
• Copilot interfaces take the best between the two.

• Let human experts weigh in on quality, especially for specialized use cases.
• Build an LLM as a Judge system to reduce human labelling.

User Interface: Beyond Chat

• Straightforward tasks? Traditional code or simpler ML/DL often does the job—cheaper and faster. For example, have you checked the new ModernBERT by @jeremyphoward ?

Evaluation Is Non-Negotiable
• Automate checks for basic errors and inconsistent outputs.