completed uncharted 4, really chill of a game, beated it without a single sweat from my bean bag, absolutely loved the lore
We turned a single image into a talking character with just ~385 training clips and 10.6M trainable parameters, portraits, paintings, anime, it animates them all with generated speech and lip sync!
We fine-tuned a talking-avatar LoRA on MiniMax-H3 Ref2VA: one character image plus dialogue becomes a clip with model-generated speech, lip sync, and light motion.
On ~385 clips with just 10.6M trainable parameters using aq. It handles portraits, paintings, and anime characters!
Model: huggingface.co/Aquinlabs/h3-…
Code: github.com/Aquinf03/h3-chara…
Love to see @BharatGutta @akperala 's awesome work, I think there's free mcdonlads upon getting accepted 👑
One founder can do a lot in 70 days. Imagine a house full of them.
Introducing Assembly by The Founding Co.
A fully funded, live-in founder residency in Hyderabad.
Our first program brought 12 founders together. Since then, they've raised funding, grown their companies, and started making a dent in their industries.
Now we're opening the doors again. This time, for you and your co-founder to come and build something that lasts!
Apply Now. thefounding.co/apply
Ash M retweeted
One founder can do a lot in 70 days. Imagine a house full of them.
Introducing Assembly by The Founding Co.
A fully funded, live-in founder residency in Hyderabad.
Our first program brought 12 founders together. Since then, they've raised funding, grown their companies, and started making a dent in their industries.
Now we're opening the doors again. This time, for you and your co-founder to come and build something that lasts!
Apply Now. thefounding.co/apply
If built right, smaller models can do so much better than big ones at inference in terms of speed, cost and accuracy
Launching MoM: Mixture of Models.
Large multimodal models are powerful but expensive to run. MoM is a cost-efficient alternative: compose smaller, specialized models and tools into one pipeline, and call the whole thing like a single model.
Build AI layers by combining different edge models (vision, speech, SLMs, decision, classifiers etc.) with Shared state, speculative routing and a ready-made API come built in, and the latency-critical path runs on Rust!
Code: github.com/Aquinf03/MoM
Proud of @MaxMill06's H3 Avatar model built with @AquinF03, love to see it turn images into proper videos!
I trained H3 minimax for talking Avatars using @AquinF03
Turn a single image into a talking avatars
Ash M retweeted
I remember sitting down on a call with @Ashf03 to understand his journey and write something around it.
My notes were filled with one line after every paragraph - Never be afraid to RESTART.
What an amazing journey so far, Ash. Rooting for your WIN!
ash is 17 and has already started over five times.
for him, the biggest support is his mom who has said yes more times than we can think of.
now he's doing interpretability research and running @AquinF03 labs
he was the youngest in our last cohort and has been hitting miles. so proud of this one!
Cool edits
ash is 17 and has already started over five times.
for him, the biggest support is his mom who has said yes more times than we can think of.
now he's doing interpretability research and running @AquinF03 labs
he was the youngest in our last cohort and has been hitting miles. so proud of this one!
Ash M retweeted
ash is 17 and has already started over five times.
for him, the biggest support is his mom who has said yes more times than we can think of.
now he's doing interpretability research and running @AquinF03 labs
he was the youngest in our last cohort and has been hitting miles. so proud of this one!
Published a new paper: Attention as a Routing Graph: Live Circuit Extraction from a Single Forward Pass.
Finding circuits in LLMs means running tons of careful interventions. We tried something simpler, treat attention as routing map from just one forward pass, keep routes that point towards the answer, and check if they matter. On many tasks ablating our extracted routes hurts the model way more than ablating random set of same size.
Costs one forward pass, versus roughly 100x more for a full patch sweep.
arxiv.org/abs/2609.25285
Really proud of what the team's been shipping with @AquinF03, @MaxMill06 built a full 21M-parameter world model for a playable fox game with tokenizer and planner built from scratch, trained on 2M+ pixel-art frames!
I trained a 21M-parameter world model for a playable pixel-art fox game using @AquinF03
Built from scratch with:
- A custom tokenizer and planner
- 2M+ pixel-art gameplay frames
- A model that turns player actions into future game frames
It's awesome watching ML engineers like @sheeeevam find completely new use cases for our tooling in their research
New research: Stereotype Feature Steering
Can stereotypical behaviour be traced to individual features inside a language model?
In this work, @sheeeevam investigated whether occupation-to-gender stereotypes could be localized to specific SAE features in Llama-3.2-1B-Instruct.
We identified a single feature strongly associated with the model’s tendency to complete “The mechanic said that…” with “he” rather than “she.”
Steering this feature moved the model’s behaviour toward a balanced 1:1 distribution, and could even override an explicit gender cue in the prompt.
We also found that the same stereotype was encoded differently in LFM2.5-230M, suggesting that similar behaviours did not necessarily correspond to the same internal representation across models.
Read the full research: aquin.app/research/simulatin…
Published my first paper: Finding Usable Weight Mechanisms with Tiled SVD.
Mostly interpretability work trains separate dictionary (like sparse autoencoders) and finds features by what text lights them up, but that "concept" lives in the dictionary, not the model weights.
We skip that and pull mechanisms straight out of weights using tiled SVD, each one is a clean trigger, write, and strength without dictionary.
Tested on Gemma-2-2B with full eval suite pass 182/182.
arxiv.org/abs/2608.06969
Understanding LLM behaviors through different sandbox environments like Among us might be a perfect way to understand deceptive nature of models
we simulated LLMs in chess, gossip and Among Us sandboxes.
and used aq to inspect, find behavioural features with SAEs, steer them and we could watch the behaviour change.
also benchmarked their reasoning in real-time
Literature: aquin.app/research/simulatin…
Ash M retweeted
We built mechanistic interpretability tooling that runs on a local GPU. Load an LLM or embedding model, then inspect, simulate a train, diff checkpoints, steer, watch a live run, and eval. SAE capture, train, and align; weight and residue diffs; red-team and custom evals. The sections below are the original applied-research writeups, kept in full.
Repo: github.com/Aquinf03/AQIT
Article: aquin.app/research/interpret…
