Krish Goel (@krshgl) / X

Krish Goel

1,069 posts

Krish Goel

@krshgl

Memory Research @LittlebirdAI. Learning to think, and teaching it to machines, sometimes robots. Tweeting notes-to-self and musings.

New Delhi, India

Joined April 2020

Pinned
Krish Goel
@krshgl
May 13, 2025
5 undergrads, no lab, no supervision, and 1 year of "figuring it out" later, we're hyped to present - "Chronocept". It's the world's first benchmark to instill a sense of time in machines. This our lives' best work, and we hope to say this about everything we do hereinafter.
97K
Krish Goel
@krshgl
Aug 4, 2025
Replying to @TheCineprism
The Interview - the real Gs know what I am talking about.
41K
Krish Goel
@krshgl
Aug 13, 2025
Replying to @senbodev
Learning by relying on structure decided by someone else is so dystopian. I’d much rather be non-creative than learn art this way. The same approach killed my interest in Physics and Math in high school - subjects I’d have otherwise thrived in. AI solves exactly this
9.2K
Krish Goel
@krshgl
May 13, 2025
Replying to @krshgl
Temporal reasoning is a gap in AI. Humans intuitively know when facts expire, or "how long a dentist's appointment lasts"; machines don’t. Chronocept fixes that. It teaches models how validity of information evolves over time. TL;DR:
arxiv.org
Chronocept: Instilling a Sense of Time in Machines
Human cognition is deeply intertwined with a sense of time, known as Chronoception. This sense allows us to judge how long facts remain valid and when knowledge becomes outdated. Despite progress...
5.3K
Krish Goel
@krshgl
Aug 25, 2025
Replying to @ishi_psd
That’s not a friend, that’s a fine-tuned language model
1.1K
Krish Goel
@krshgl
May 13, 2025
Replying to @krshgl
This ones for the team - @sanskxr02, @mdev_1 @HarshKumar3014, Vishesh and the annotators. You guys rock! Still lowkey mad that it was just us. Just 5 remote dudes with Python, notebooks, Google Docs (and Overleaf😔), and a weird ass obsession to model time for some reason.
3.4K
Krish Goel
@krshgl
May 13, 2025
Replying to @krshgl
Chronocept comprises two benchmarks - - Benchmark I: short atomic statements (1.2K+ samples) - Benchmark II: multi-sentence paragraphs (650+ samples) Datasets: huggingface.co/datasets/krish…
8.8K
Krish Goel
@krshgl
May 16, 2025
So the last post blew up, thank you for the validation. Hi everyone, I am Krish. Graduating this weekend, and looking to spend the next few years going deep on knowledge graphs, reasoning in retrieval, meta learning and SSL - by founding myself, in research labs, or with
2.6K
Krish Goel
@krshgl
May 15, 2025
Replying to @vijayshekhar
Thank you so much sir. I think it’s a function of being inspired by the greats early on 😁
2.2K
Krish Goel
@krshgl
May 13, 2025
Replying to @krshgl
If you're building: - RAG systems - Fact checkers - Knowledge bases - Proactive assistants and agents (knowing when to speak, not just what) Chronocept can help. It's designed to plug into ML systems at the application layer, and help make time-aware decisions.
2.7K
Krish Goel
@krshgl
Jul 20, 2023
Replying to @Ashneer_Grover and @AIGFofficial
I don’t think your math has factored-in things like crushing debt, suicide and chronic depression from these “skilled” games. That’s the real hidden math. So much for creating responsible and ethical “dhandas” and doing good for the society.
3.6K
Krish Goel
@krshgl
Jun 9, 2025
Replying to @IndianTechGuide
I was offered a 10 month ISRO internship in my final year (would have lasted 2 months post graduation). Required relocation, opting out of placements, gave zero stipend and didn’t guarantee a job afterwards. I understand working for experience, but this is borderline exploitative
2.7K
Krish Goel
@krshgl
May 13, 2025
Replying to @krshgl
Each Chronocept annotation is a sentence broken into 8 semantic axes and fit a skewed-normal curve over log-time - - Location: when validity peaks - Scale: how long it lasts - Skewness: how it rises/fades No classes. Just a continuous probability distribution over time.
3.7K
Krish Goel
@krshgl
May 13, 2025
Replying to @krshgl
We trained every model, wrote every line, and designed every annotation protocol ourselves. Our suggestion: BERT embeddings with an FFNN (for short sentences), or with a Bi-LSTM (for longer contexts) work best. Baselines: github.com/krishgoel/chro…
2.9K