Varun

[Ben Dicken @BenjDicken]

I'll give you an update in 88 hours.

[Varun @varunsaagar_ai]

@tpktelangana @gkcs_ I agree with you, in 2021 ai government conference this was present as a solution by one of the state and this system is already already in place for long time.

[Varun @varunsaagar_ai]

Done with my feed

[Tornado guy @fanofaliens]

@sundarpichai We are literally so back

[Varun @varunsaagar_ai]

@paulabartabajo_ @liquidai What type of licence is this model?

[Varun @varunsaagar_ai]

Bro it is vector search! Lol what ever you can sell, under the hood same

- Vector db
- semantic search
- embedding

[Tech with Mak @techNmak]

Vector Search is not Memory.

RAG is great for retrieving documents, but it fails at maintaining User State. If you are just dumping chat logs into a vector DB, you aren't building an Agent, you're building a semantic search engine.

I found the open-source library that solves this. It’s called Mem0.

It acts as a self-improving Memory Layer that sits between your user and the LLM.

The Cheat Code for AI Agents => Instead of static retrieval, Mem0 creates adaptive memory that evolves with every interaction.

This solves the three biggest bottlenecks in Agent production:

1./ The Amnesia Problem
Standard RAG retrieves facts but forgets preferences. Mem0 tracks User, Session, and Agent state separately. If a user prefers Python over JS in session 1, the Agent remembers it in session 50.

2./ The Context Window Trap
Mem0 uses a Memory Compression Engine. It distills complex history into optimized prompts. Result: 90% reduction in token usage and significantly lower latency (TTFT).

3./ Multi-Agent Coordination
If you use CrewAI or LangGraph, your agents are usually siloed. Mem0 acts as a shared brain. What your "Researcher" agent learns, your "Writer" agent instantly knows.

It’s fully open-source and integrates with OpenAI, Vercel AI SDK, and LangChain.

Stop building stateless bots.

[Varun @varunsaagar_ai]

Gold miner

[Lewis Tunstall @_lewtun]

We've just published the Smol Training Playbook: a distillation of hard earned knowledge to share exactly what it takes to train SOTA LLMs ⚡️

Featuring our protagonist SmolLM3, we cover:

🧭 Strategy on whether to train your own LLM and burn all your VC money

🪨 Pretraining, aka turning a mountain of text into a fancy auto-completer

🗿How to sculpt base models with post-training alchemy

🛠️ The underlying infra and how to debug your way out of NCCL purgatory

Highlights from the post-training chapter in the thread 👇

[Charles 🎉 Frye @charles_irl]

We reverse-engineered Flash Attention 4.

[Varun @varunsaagar_ai]

@sandboxvibe Wow!! Let me try

[Varun @varunsaagar_ai]

@vibe_terminal Please drop me

[Y Combinator @ycombinator]

"Understand your users. That's the key."

Paul Graham, if he could only give startups one piece of advice:

[Varun @varunsaagar_ai]

😭😂

[Rhys @RhysSullivan]

these gpt-5 numbers are insane 🤯

[Varun @varunsaagar_ai]

@anythingai How can i get the credits 🥺

[Varun @varunsaagar_ai]

What my boy @RachinRavinder is doing here

[bone @boneGPT]

If you want a check from Sam Altman, you're gonna have to do a "seed round"

[Varun @varunsaagar_ai]

Amazing space and so much transparency on how they are implementing the ad recco!

[Business @XBusiness]

https://t.co/RUNtkMKNMP

[Varun @varunsaagar_ai]

@WisprFlow when you guys are going to be available in android 😔

[Varun @varunsaagar_ai]

Got my invite 🎉🎉🎉

[Varun @varunsaagar_ai]

Awesome work! Globalization of token

[Sukjun (June) Hwang @sukjun_hwang]

Tokenization has been the final barrier to truly end-to-end language models.

We developed the H-Net: a hierarchical network that replaces tokenization with a dynamic chunking process directly inside the model, automatically discovering and operating over meaningful units of data

[Sukjun (June) Hwang @sukjun_hwang]

Tokenization has been the final barrier to truly end-to-end language models.

We developed the H-Net: a hierarchical network that replaces tokenization with a dynamic chunking process directly inside the model, automatically discovering and operating over meaningful units of data