Mini Blog
Notes & essays by Ismail Alam
AI Agrees With You. That Should Worry You.
Ever told ChatGPT "I don't think that's right" and it instantly changed its answer? Wait, it doesn't. It agrees. It wasn't trained to find truth. It was trained to please you....
Read postWhy ChatGPT Forgets the Start of a Long Chat
Ever notice ChatGPT forgets what you said at the start of a long chat? Not a bug. It only holds so much at once. As the chat grows, it drops the oldest parts. Ask it to call...
Read postSmall Models Winning
Small models are catching up fast. A 3 billion parameter model today beats a 70 billion model from two years ago. How ? Better training data. Smarter training process. Not more...
Read postAI Doesn't Read. It Counts.
ChatGPT never reads your words. It turns them into numbers instead. Every word gets its own number pattern. Words with similar meanings end up with similar patterns. - "King"...
Read postHow AI Learns What Humans Like
Ever wonder why ChatGPT gives helpful answers instead of random text? **Reinforcement Learning from Human Feedback (RLHF)** — that's the secret. After training on the internet,...
Read postQuantization Shrinks Large AI Models Without Breaking Them
An AI model has billions of fine-tuned weights (parameters) after training. Normally, each weight is recorded as a long decimal like 3.14159 (32-bit float). A 10 billion...
Read postYour Coding Agent Is Just an Elegant Loop
AI coding tools often feel like magic. But behind the scenes, even the best agents rely on three simple steps: - **Look** — reads your codebase and error logs - **Think** —...
Read postBetween You and Users, Choose Your Users
Now that AI lets anyone build, the temptation is to build everything. A great co-builder is just a prompt away. While we might be on the cutting edge, our average users are...
Read postBe a Great Human-in-the-Loop (So Nothing Can Replace You)
To stand out, view tech through three simple lenses: problems, use cases, and solutions. Take Big Data. Hadoop was revolutionary, but coupling storage and compute on-premise...
Read postAgent Harness Creates Magic: Part 1
LLMs, however powerful, are inherently stateless. In a raw setting, you send a message, they respond, and then they immediately forget. If you were to interact with a bare,...
Read postStay Teachable
It is important to be open to all kinds of ideas, perspectives and opinions. A sense that there exists unique logic, reasoning, struggle, stories in every field allows you to...
Read postUnderstanding Both Fundamentals & Abstractions Is Important
In the process of making something useful, there are multiple kinds of abstractions created, use them but know the fundamentals. Abstractions make usage easier but knowing...
Read postPlay in GenAI is now in its applications
There is no doubt that LLMs are advancing but now they have mainly transitioned to incremental refinements. **Why will this happen?** Learning for LLMs, which happens through...
Read post