building @interact_studio || growth eng @AquinF03 love building world models
Internet
Joined September 2023
- Tweets3.2K
- Following2.4K
- Followers688
- Likes11.5K
Pinned Tweet
this is where interactive media is heading
meet interview couch bear : he remembers what you told him last time. hears emotion in your voice. explains ideas with the objects around him.
all in real-time, powered by world models
made with @reactorworld
Excited to see what everyone builds. 🎮
What if you could make games with AI?
we're hosting "MAKE IT PLAYABLE", a game creation challenge with @ElevenLabs, part of @chromaawards in Delhi NCR on 14th november.
Build games. Experiment. Make them playable. 🎮
sign up now: luma.com/cnlehsw4
5
269
what happens when the viewer decides what happens next in the story ?
we're hosting a kickoff talk for a film residency in USC on 9th nov (8 pm PT)
check out : luma.com/g5ozassp
Interactive storytelling is one of the most exciting spaces in media rn.
Excited to bring this conversation to USC and kick off BRANCH.
If you’re interested in the future of storytelling, come join us!
What if your audience could shape the story?
We’re kicking off BRANCH- a 2 week film residency at USC with @CtrlMovieLtd to explore interactive storytelling.
📅 Nov 9 | 8–9 PM
Register: luma.com/g5ozassp
1
8
362
Sachin retweeted
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…
Sachin retweeted
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
trained on ~385 clips with just 10.6M trainable parameters.
It can animate everything from portraits and paintings to anime characters, with generated speech + lip sync.
Single image in → talking video out
1
5
290
H3 Avatar aq is live on Hugging Face.
It takes a single reference image + dialogue and generates a speaking avatar video with identity preservation, expressive mouth motion, subtle head movement, and model-generated audio.
Try it here: huggingface.co/Aquinlabs/h3-…
5
148
Sachin 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!
Dataset: 2M+ pixel-art gameplay frames 🦊
huggingface.co/datasets/Aqui…
1
3
127
Sachin retweeted
really excited to have @VivGadepalli joining variance house as a mentor.
angel investor and builder behind Vivly, bringing his experience backing ambitious early-stage founders to the house.
14 builders. 30 days. bengaluru.
variance.house
Sachin retweeted
Introducing Redstring.
Your new-age talent discovery platform. Describe exactly who you need. Redstring finds the right people, reaches out to them, and brings you a shortlist ready to interview.
Today, hiring still means posting a job, waiting, searching, and screening manually.
Redstring is built for the opposite. tell us who you need. We’ll help you find them.
Early Access opens now at (redstring.co.in)
Waitlist link in comments!
@Redstringhire
Sachin retweeted
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…
what if you could find concepts inside the model weights themselves?
@AquinF0 just published Finding Usable Weight Mechanisms with Tiled SVD.
instead of learning a separate dictionary, it extract mechanisms directly from the weights:
we’ve published our first paper: Finding Usable Weight Mechanisms with Tiled SVD.
most interpretability methods learn a separate dictionary(like SAEs) to identify features in a model.
we instead extract mechanisms directly from the weights using tiled SVD, giving each mechanism a clean trigger, write direction, and strength.
evaluated on Gemma-2-2B: 182/182 evaluations passed.
paper arxiv.org/abs/2608.06969
5
430
made a small pirate game where nothing ashore is pre-made.
built with Three.js + H3 reference Turbo by @reactorworld.
sail across a low-poly sea. Go ashore, explore quests and collect gold.
here's the code : github.com/sachin1705s/littl…
2
94