
Deconstructing AI Memory Systems @mem0ai (YC S24) | @GoogleDevExpert in ML
Joined June 2016
- Tweets1.1K
- Following1.6K
- Followers563
- Likes3.7K
Not every coding task needs a frontier model. Open models meet the quality bar for much of that work at a fraction of the cost.
We put together a guide on how to evaluate open models for coding agents, covering:
- Choosing a pilot team and workload
- Measuring quality on your own merged PRs
- Tracking cache-adjusted cost per successful task
- Maximizing ease of use: keep your harness + get a map from current models to open alternatives
Read it here: baseten.co/resources/guide/e…
Aashi Dutt retweeted
Meta published a paper that might end the current LLM era.
They proved you don’t need tokenizers to train SOTA models.
Every major AI (ChatGPT, Claude, Llama) reads text in "tokens", chunks of words. This requires massive vocabularies, usually around 100,000 different tokens.
It's how models have always been built.
But researchers at Meta proved this is a massive bottleneck.
They trained a model to read raw "bytes" instead. Only 256 basic units.
The plot twist:
They took a tiny 1-billion parameter model and fed it up to 1 trillion bytes of data.
At first, the standard token-based model was winning in the low-compute regime.
But then something crazy happened.
The token model hit a ceiling. It plateaued. It completely stopped getting smarter.
The byte model just kept climbing.
By converting token probabilities into pure byte-level reasoning (using a new distillation method called "End-Of-Token"), the tiny byte model shattered the performance ceiling.
It matched the standard token model using just 1/6th of the training data.
It crushed storage costs by 80%.
And asymptotically, this tiny 1B-parameter byte model outperforms Llama 3.2 and even the 2-billion parameter Gemma model across the board.
Nobody is talking about the real implication here.
We have spent years forcing AI to read text the way humans do.
It turns out, AI gets much smarter, much faster, when you let it read like a machine.
Aashi Dutt retweeted
Inkling 2 Small took a while... but it's more interesting than I initially thought!
Using attention residuals (like Kimi K3), but without the linear-attention hybrid.
Food for discussion for the end-of-the-year architecture trends analysis!
Aashi Dutt retweeted
Introducing Gemini Decision 1, our new frontier decision model in the Gemini API.
It is $0.10 per 1M input tokens with no additional cost. With support for image, audio, video, and PDF natively, 256K context window, & 1K questions.
You can try it for free in the API!
Aashi Dutt retweeted
We spent $2m so you don't have to.
And so you can learn how to train MoEs.
Mixture-of-Experts LLMs are what powers the AIs you use everyday, and few people in the world know how to train them at scale (a handful at each top AI lab).
This handbook distills everything we learned in meticulously crafted chapters and figures.
Disclaimer: it may take days to read.
We hope to see 1000x more people able to train their own frontier models all over the world! 🤗
Congrats @Nouamanetazi @lvwerra @Thom_Wolf
Aashi Dutt retweeted
Frontier LLMs are showing potential to read, fix, & rewrite neural network weights.
MIT CSAILers call this "Weight-Space Programming." It's what GPT-6 did when it edited various small NNs including Qwen3.5-0.8B, using tools in interpretability research or even direct weight editing. Soon enough, LLMs might natively optimize NN weights at a human or even super-human level: tinyurl.com/378w4ckw
Aashi Dutt retweeted
Hi! this is Rohan Anil, I develop training algorithms among other things - my work trains various frontier models.
At CoreAuto, I was fascinated by the problem of training on compute islands, eventually in space training.
I hope you all enjoy this write up!🪐🛰️
coreauto.com/hitchhikers-gui…
Today we’re releasing our third open-source decision model, joey-1 (2B), the fastest decision model available and among the most intelligent decision models of its size.
joey-1 is trained by the @baselabs team, and is available on our Model APIs, as a dedicated deployment, and with a cookbook to fine-tune on our Training Jobs.
joey-1 is built to maximize speed and minimize cost without sacrificing intelligence for those workloads where TTFT really matters.
Jump on it: baseten.co/library/joey-1/
Aashi Dutt retweeted
If you like reading and are looking to upskill in the field of deep learning, here is a collection just for you.
huggingface.co/collections/a…
Aashi Dutt retweeted
Introducing Unlimited OCR 2 — now open source!
No page limits. No growing attention costs.
Unlimited OCR 2 parses documents of virtually unlimited length in one go, maintaining high accuracy and efficiency as they grow.
We tested it on 5,000-page documents. It held up.
Aashi Dutt retweeted
Introducing Reinforcement Learning from Collective Rewards (RLCR).
RLCR is a recipe that trains language models as agent organisations, which share a message board and workspace. We use it to train Avalon-M1-Mini (120B) and study its behaviour.
We find a number of interesting results, including:
🛠️ Learnt organisational hierarchies and divisions of labour
🕵️ Increased collective verification, through hires of verifiers and testers
👥 The rise of collective thinking, where agents think more about other agents and how to help them
We believe RLCR can provide a laboratory for studying collective intelligence and multi-agent safety.
Blog post and technical report 👇
Today we release mammoth-1 and elephant-1!
We post-trained mammoth-1, the #1 ranked best decision model in the world, and elephant-1, which is the best model under 15B (and beats Jev).
We're open-sourcing all model weights and provide a technical report detailing all our training and data methods.
Aashi Dutt retweeted
Today we launch SGLang-TorchTPU, built with @Google TPU team and Meta @PyTorch team.
We’re excited to see it powering production inference at Meta!
SGLang-TorchTPU is a TPU plugin for SGLang that brings:
- Top open models on TPU: Muse Spark Open, GLM-5.3 Flash, Kimi K3, Ling 3.1 Flash, MiniMax H3
- SGLang’s production features for large agentic workloads
- Custom MegaKernel: Muse Spark Open hits 770 tok/s decode at bs=1 on InferenceX 8K/1K
SGLang-TorchTPU is available now: github.com/sgl-project/sglan…
Big day for the @PyTorch ecosystem at #PyTorchConf2026 🚀
We’re officially open-sourcing TorchTPU—enabling native, hardware-accelerated PyTorch execution across Cloud TPUs and custom infrastructure.
Here’s everything we launched to supercharge your training & inference workloads 🧵👇
goo.gle/4Bl0zdG
Aashi Dutt retweeted
1/ Excited to release The ultimate guide to multi-harness RL
The same model behaves differently in every agent harness. So we built an open way to train any model with RL on any task set, inside the harnesses people actually use, like Claude Code, Codex, and OpenCode, without changing a single line of harness or training code.
Trained across four harnesses, LFM2.5-2.6B went from 42% to 54% with 31% fewer tool calls. 🧵
This was fun to watch 🙋🏻♀️
Ben Affleck reveals he writes Python, understands convolutional neural networks, worked extensively with GPUs, and used his celebrity status to get private looks at Google and OpenAI’s video models
“I’ve always been kind of into computers since I was young. Then, when film started to move from analog film to digital, I became more interested in that aspect of it. The visual-effects workflow for many years has included machine learning, so I can write pretty shitty Python scripts and stuff like that.
“With convolutional neural networks, which were the precursors to what the transformer can do, which is much more computation simultaneously, you would do things like look at what’s called a tensor. That’s the numerical translation of a visual image in numbers, like the batch number, the frame number and the red, green and blue values of each pixel in each frame. It’s just that simple. That numeric is called a tensor.
“You’d use a convolutional neural network to identify patterns that reveal what’s called edge detection or feature extraction, which is identifying patterns well enough to know, this is where the window ledge is, so we can more easily take the green-screen image out and replace it with something.
“That was familiar to me early on because, prior to Artists Equity, I had a small visual-effects company. I’ve worked with GPUs a lot too. The visual-effects guys said, ‘Hey, you should see. There are a couple: Google and this other company, OpenAI, are doing really interesting stuff with transformers in video.’
“I’ve learned that I can actually just call up and go, ‘Hey, it’s Ben Affleck. Can I come see what you’re doing?’ Sometimes people say yes, to my astonishment.”
2
58
Aashi Dutt retweeted
Meet Mistral Large 4, aka Le Chonk.
• 1T parameters, natively multimodal. 49B active.
It is the best open weights model from US or Europe on aggregated benchmarks.
• State-of-the-art on critical workloads, including cyber defense, manufacturing and finance and it surpasses closed frontier models on visual grounding.
• Forged in Europe end-to-end and is deployable from Europe via our own Mistral Cloud infrastructure.
• Available to all via API today. Working with cybersecurity partners privately.
Open weights release end of October.
🤖 Made with AI



























