satyam

2025 feels special. Here’s some stuff that’ll be flowing in me this year Things I plan to play with: ⁃Low level C, Rust - for the fun of it ⁃Reinforcement Learning, Generative Models - build novel things here ⁃Protein design: using AI tools and invitro methods - make…

Umm.. wrote 35k LOC - pure vibe coding. Just sitting back and telling AI what to do while i am roaming around doing other things. This would have taken 1 month of full focus work without AI. I am in shock - not hyping AI but definitely feeling THE shift.

AI Code Agents are not just ONE OF THE MANY use case for agents. They are the core abstraction upon which rest of the AI use cases would be generated.

so i guess that in-context rl or test time rl based solutions to coding would be fundamentally superhuman. Imagine for instance, emergent discovery of new abstractions to solve a problem repeatedly as new features are added

What are some articles, blogposts, essays, etc. that you still think about over a year later? Here are a few of mine:

I use a lot of “Teach me X by inventing new ideas with me”. O3 is first model that seems to actually be genuinely inventive so much so that I am planning to deep dive into some of the concrete ideas we came up with. tldr: ❤️ Its a very different world starting today

real world will be a limit unless it is running these experiments in in-silco environments

Very impressed by govt’s admission that India is nowhere near China (or US) in terms of hard tech / innovation. I love the fact that govt is actively encouraging startups to aspire to be best in the world, and not just try solve problems for the rich Indians.

Playbook to defeat frontier ai labs without billions of dollars initially: - build an app on top of their models - solve an important, large problem for humanity - resume training on top OSS models to reduce dependency, lower costs for certain tasks, increase performance -…

Vibe marketing > vibe coding

Vibe prototyped this new way of doing UI for AI agents. It uses a kanban board as the interface to run AI-Human hybrid teams with. check it out! Is this the future of work? v0-ai-trello.vercel.app

“Until death, all defeat is psychological”

There is so much low-hanging fruit in inference-time scaling including: * Reasoning in continuous latent space instead of token space * Learning to reason about tool use, minimizing the need for agent scaffolding * Domain-specific and task-specific verifiers/validators. Some of…

The sudden popularity of deepseek is really annoying. Miss the days when they were only known to those who enjoyed their detailed reports..

Isn't it ironic? A hundred EV companies have bloomed under communism, while capitalism subsidizes one moron. A startup has trained an AI for $5.5 million under communism, while capitalist AI requires $500 billion in government support. 1/11