Bhavika @stacktraceee

Sharing the first article here. You can follow the rest of the series by article number in ascending order :) medium.com/@bhavikapawani…

If you work on applied AI, it’s worth understanding how these models work internally. It’s always better to learn things from first principles. I wrote a series on GenAI internals a few months ago and intentionally kept the math light. Sharing the first article link in comments

Jev, Laya and Kev feel like AI going back to an old idea. Before LLMs, we built specialized models for specific jobs. Then we tried making one general model do everything. Maybe the future is somewhere in between: general intelligence underneath, specialized models on top.

Maybe AI doesn’t need to feel an emotion to understand it. It can study millions of examples of grief, love, fear and learn their patterns better than one human ever could. Which raises a stranger question: Can AI understand something at a human level without experiencing it?

Some things I’ve learned about building evals for LLM apps and agents x.com/i/article/2086…

Had a good laugh watching this video But this is also a great example of just how far AI video generation has come. A few years ago, generating a realistic image itself felt impressive. Video is a much harder problem because you’re not just generating good-looking frames. You need temporal consistency across all of them. The same face has to remain the same face. Objects can’t randomly disappear. Lighting, geometry and camera motion need to stay coherent. If something leaves the frame and comes back, the model still needs some sense of what existed before. Modern video models are getting surprisingly good at this by learning space and time together. They compress videos into latent representations, use attention across spatial and temporal tokens, and use diffusion or flow-based generation to learn how scenes and motion evolve over time. Pretty crazy how fast we got here

DesiEsco🇮🇳 @DesiEsco7

2 months ago

AI will take over the jobs in India meanwhile AI in india-

1K 8K 71K 4.8M 16K

One thing I never expected while going deep into LLMs... The more I study how they're trained and optimized, the more I end up learning how to optimize myself. LLMs were inspired by ideas from neuroscience, even though they work very differently from the human brain. That's why it's fascinating that many of the techniques we use to improve them also seem useful for humans. Curriculum learning is just learning the basics before the hard stuff. Knowledge distillation is reducing a complex topic into a mental model you can actually remember. Retrieval beats trying to memorize everything. Attention reminds you that not every piece of information deserves equal focus. Iterative refinement is why your second or third explanation is usually much better than your first. Context often matters as much as raw intelligence. Maybe that's why learning AI has slowly changed how I learn everything else.

I used to be confused about when to use Skills vs code for AI agents Now I think of it this way: • Early stage → Skills. Optimize for iteration speed • Production → Code. Optimize for cost, scale, and reliability Let the LLM do the thinking. Let code do the orchestration

"Agents are just scripts." That's what people say. Then you build one for production. Writing the script is easy. Maintaining it across countless edge cases is the hard part. Miss one tiny condition or an important piece of context, and the entire execution can fail.

One thing that surprised me while evaluating LLMs: The same step by step prompt that helped a smaller model hurt a stronger one. The stronger model didn't need the extra scaffolding. I was in its way. Prompt quality isn't universal. That's why we need evals, not intuition.

Some things I've been doing lately to save on AI tokens: • Caveman mode • Better model for planning → cheaper model for coding • Run /insights to spot token waste • Create a project index/map for the codebase • Start a new chat before context gets too big

@NTFabiano It’s been 4 months since I quit short-form videos on IG and YouTube. Less stress, more control, better life. Just stopping this one thing improved other areas of my life. Never want to go back to that again.

@shri_shobhit I’m also doing the same, learning AI from first principles before getting back to applied side

@bluewmist Learning isn’t just books, videos count too. What matters is to keep learning.

@SakshiSugandhi Even if the job is CRUD, problem-solving skill shows how you handle unknowns

@OrevaZSN Reduce social media and Netflix. They take a lot of time. You’ll have more time for everything else

@gdb kinda crazy how fast this is moving… doesn’t even feel real yet tbh

@rezoundous Didn’t go anywhere. n8n just isn’t hyped now. AI agents and Zapier are getting more attention

@yacineMTB Feels off. OpenAI is scaling distribution while Anthropic is focused on quality. Not a Yahoo–Google replay.