Harsh Kumar
@theharshon
Followers
1K
Following
9K
Media
644
Statuses
5K
Founder & CEO @incupie | Building Artificial Super Intelligence (A.S.I.)
Bengaluru, Bharat
Joined April 2015
That is why we are focused on compression.
Agent memory is a compression and caching problem.
4
3
19
As someone building my own AI foundational model. I am optimistic about acceleration. But think about this. Time compression does not guarantee a breakthrough. Simulating 100,000 years of effort in medicine does not mean we will get 100,000 years of progress.
0
0
0
Green screen ✅.Light ✅. Upping my content game so I can share my work better & more often. Excited to keep building creating and posting here. #buildinpublic
5
2
12
We vision a powerful universal digital assistant at incupie. That serves as a personalized companion for everyone.
6
7
10
If you’re 22 and thinking about AI, skip the noise. Build small agents that solve boring real problems. Then scale.
6
0
11
The AI field is full of people chasing novelty. Some of the smartest work right now is in making old ideas actually work at scale.
3
0
9
Paying $200/month for a browser isn't ideal in the current market . Without significant, unique benefits. A Chrome extension can do the same, in my opinion. @AravSrinivas.
okay but how good is perplexity comet really??.
4
3
8
The faster the model learns, the more you realize how slow we are to update our assumptions.
3
0
11
A business aims to profit while ensuring happiness for its owner, vendors, employees, customers, and community.
3
0
6
Everyone wants to talk about weights and parameters. Few want to talk about values encoded in the data. That’s where the model really learns.
2
0
5
A model doesn’t need to “understand” in the human sense. It just needs to behave usefully enough that you forget it doesn’t.
3
0
5
Gradient noise in early training is underrated. Some of our best-performing runs came from things that shouldn’t have worked, . But did.
0
0
5
Most people spend hours, sometimes days, wiring nodes, triggering prompts, and debugging workflows to get a single AI agent working. Sometimes, it doesn’t work.
0
0
4
Every time we train a large model, we throw a stone into the dark, hoping it echoes something meaningful. Sometimes it sings.
0
0
3
A true dynamic UI is alive. Not just reactive but proactive, predictive, and purposeful. It's the UI version of behaving intelligently.
1
0
7
Re-read Norvig & Russell today. Still the best line: .“The question is not whether machines can think, but whether they can behave intelligently.”. We’re still answering that.
1
0
3
LLMs hallucinate. Humans hallucinate. The only difference is that the LLMs don’t pretend they’re sure about it.
0
0
2