Blog
This blog is a curated collection of my thoughts on AI, technology, and business—originally shared on LinkedIn and expanded here. You'll find research paper analyses broken down for practical application, non-technical explanations of complex AI concepts, reflections on industry trends, and actionable insights for leveraging AI in real-world scenarios. Each post aims to cut through the hype and deliver substance that matters.
-
Rewriting from scratch is increasingly viable due to AI-assi
Rewriting from scratch is increasingly viable due to AI-assisted coding. Someday, all SaaS might be rebuilt every 3 months, with AI ensuring parity & no regression via AI auto-tests based on hi...
-
Qwen releases QwQ-32B, small reasoning model which rivals wi
Qwen releases QwQ-32B, small reasoning model which rivals with DeepSeek-R1 and o1-mini.
-
You should encourage employees to wear Meta glasses to captu
You should encourage employees to wear Meta glasses to capture tribal knowledge for AI coding agents. And here is why ⬇️
-
People saying with AI you take 3 min to generate code and 2
People saying “with AI you take 3 min to generate code and 2 hours to debug it” or things along those lines are just bad software engineers.
-
18 lessons to develop better products using LLMs
- Use n
🚀 18 lessons to develop better products using LLMs:
-
Developers who dont use AI-assisted coding are already falli
Developers who don’t use AI-assisted coding are already falling behind.
-
Ai Experts
Don’t trust “AI experts”.
-
Dont trust AI experts
🐔 Don’t trust “AI experts”.
-
ColBERT: Contextualized Late Interaction over BERT
🍄 For RAG, and generally any semantic matching task, try ColBERT!
-
The AI product market is overcrowdednot with effective tool
🙋♂️ The AI product market is overcrowded—not with effective tools, but with promises.
-
TLDRs on Googles Gemma, spoiler dont use Gemma 7B yet?
N
⚡ TLDRs on Google’s Gemma, spoiler: don’t use Gemma 7B (yet?)
-
New model from Mistral AI Mistral Large!
TLDR - Second bes
🚨 New model from Mistral AI: Mistral Large!
-
How do you continue training on an already pre trained LLM
🥋 How do you continue training on an already pre trained LLM: TLDRs of the “Simple and Scalable Strategies to Continually Pre-train LLMs” paper
-
Some takeaways I got from the conversation between Lex Frid
🎧 Some takeaways I got from the conversation between Lex Fridman and Sam Altman:
-
Anyscale again! They built a FREE model comparator, where yo
Anyscale again! They built a FREE model comparator, where you can evaluate 3 open source LLMs simultaneously using the same prompt. They support llama2 7b, 13b and 70b among 8 others.
-
Amazing post by Yi Tay about the challenges of training LLM
🌟 Amazing post by Yi Tay about the challenges of training LLMs as a startup! “Training great LLMs entirely from ground up in the wilderness as a startup”
-
If you are looking for a product manager, here is one
If you are looking for a product manager, here is one. Abdel has contributed to shape my view in roadmap planning, feature definition, market analysis and more. This guy likes his job, sometimes to...
-
Did you know? Cosmic rays is one of the greatest threat to
😱 Did you know? Cosmic rays is one of the greatest threat to interplanetary travel with people onboard the spacecraft, and… they can make your LLM training fail.
-
Hugging Face was a sassy chatbot
Hugging Face was a sassy chatbot.
-
Takeways from the Mixtral paper with no chitchat
🐬 Takeways from the Mixtral paper with no chitchat.
-
OCR just got better
🤠 OCR just got better.
-
What you need to know about Groq and LPUs How can Groq run L
What you need to know about Groq and LPUs: How can Groq run LLMs so much faster than the competition?
-
How do we get LLMs to know what a software bug is without m
🤔 How do we get LLMs to know what a software bug is without making them write buggy code? A non technical dive into alignment.
-
Amazing read by M Waleed Kadous from Anyscale httpslnkd
Amazing read by M Waleed Kadous from Anyscale: https://lnkd.in/e77fUTiv
-
We often talk about the sample efficiency of ML models and
🤌 We often talk about the sample efficiency of ML models and compare it to that of humans.
-
Token facts cheat sheet for practical estimations
Tokens
🦎 Token facts cheat sheet for practical estimations
-
A lot of people, when they hear that we are using logarithm
A lot of people, when they hear that we are using logarithm as a trick to go from a multiplication to an addition, think that we are diverting the log from its original purpose, we are “hacking” it.
-
Mathematical intuition is an amazing tool, but it has its li
Mathematical intuition is an amazing tool, but it has its limits.
-
One way a project that was supposed to bring big money ends
One way a project that was “supposed” to bring big money ends up being a money waster.
-
One cool idea behind deep learning is the manifold hypothesi
One cool idea behind deep learning is the manifold hypothesis.
-
Meta AI published a blogpost titled Using AI to bring childr
Meta AI published a blogpost titled “Using AI to bring children’s drawings to life”.
-
Leetcode style coding interviews are great because - They al
Leetcode style coding interviews are great because: They allow companies to efficiently filter applications. They are standard and transparent, which makes them relatively fair. They show yo...
-
Learning rate schedules allow for changing the learning rate
Learning rate schedules allow for changing the learning rate while training, instead of having the same learning rate for every batch of every epoch.
-
ReLU is so dominant in the field because it embraces the fac
ReLU is so dominant in the field because it embraces the fact that all a neural network does is slice and dice the input space, linear transformation after linear transformation, hyperplane after h...
-
To train a classifier, you need a dataset
To train a classifier, you need a dataset. When you don’t have a dataset, you build one by asking humans to classify your examples.
-
A human brain requires less data to be trained on a given ta
A human brain requires less data to be trained on a given task than a neural net, or does it?
-
1 No product that has been called AI as of today is intellig
1) No product that has been called “AI” as of today is intelligent. 2) “AI” that is not machine learning will never achieve intelligence. 3) Deep learning may not be the tool to build intelligent s...
-
The most impactful artists of tomorrow will be the ones who
The most impactful artists of tomorrow will be the ones who know where to sail to on the latent space of art.
-
GauGAN2, NVIDIAs model keeps amazing me
GauGAN2, NVIDIA’s model keeps amazing me. Aside of producing crazy photorealistic landscapes, it can also make some stunning semi abstract images like this one, where I just made a few strokes and ...
-
Machine learning uses optimization, but in a slightly differ
Machine learning uses optimization, but in a slightly different approach than traditional optimization does.
-
Some Friday AI fun
Can a gorilla ride a camel? This is the
Some Friday AI fun: