Physical AI / Multi-modal / Fitness. Building @deepreach_AI
California, USA
Joined February 2024
- Tweets316
- Following560
- Followers195
- Likes234
Make distillation great again
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
4
122
Chris Liu retweeted
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math.
It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training.
We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵
Blog: rikiyat.github.io/blog/weigh…
Readers added context they thought people might want to know
LLM was given weights explicitly as text in prompt and used code/forward passes (per authors' logs) to ablate Qwen looping direction, a known thinking-mode issue per model card. Primarily fixed repetition in that mode, not general 'natively' improving weights/math.
rikiyat.github.io/blog/weight-sp…
github.com/RikiyaT/lm_und…
huggingface.co/Qwen/Qwen3.5-0…
huggingface.co/RikiyaT/Qwen3.…
6 takeaways from You Can See Everything:
1) Trust is everything. Say what you mean, and own every word.
2) Ambition needs realism. Know what you can deliver and what it takes to get there.
3) Never compromise your integrity. Small exceptions become irreversible patterns.
4) Great teams challenge each other. Hire people who dare to say no—and leaders who listen.
5) Money is an amplifier. It scales both competence and chaos.
6) Product is king. A story can raise a round. Only a product can build a company.
2
70
Full metal jacket.
DeepReach DR-H01.5 has landed in the US! 🇺🇸
The world’s most stable egocentric stereo capture device. Rugged metal body, built for real-world data capture.
Connecting Physical AI to the real world.
deepreach.ai/devices
#PhysicalAI #Robotics #EgocentricData #RealWorldData #DeepReachAI
3
5
518
Again, bandwidth will be the new bottleneck.
Introducing Griffin, the first model to pass the video Turing test.
48% of people who talked to it live thought it was a real human. Previous systems have had a pass rate <3%. It is #1 on NVIDIA's benchmark for full-duplex AI video.
It’s the first Human Interaction Model (HIM).
Readers added context they thought people might want to know
The 48% figure and "video Turing test" claim are from Tavus's own study of 54 one-minute calls, not independently verified or using a standard protocol. Griffin-Lite leads NVIDIA's VideoFDB benchmark on their public leaderboard.
cellcog.ai/blog/tavus-gri…
research.nvidia.com/labs/amri/proj…
tech-ish.com/2026/10/02/tav…
1
1
81
3/ Better human → robot interfaces may matter as much as better policies.
DEXOP (arxiv.org/abs/2509.04441) from @pulkitology is explicitly designed around transferring dexterous human manipulation to robots. Good capture device is a scaling law of its own.
1
168
5/ World models are starting to become useful training infrastructure rather than just video generators.
RoDyn (arxiv.org/pdf/2510.09036 ) by @ZiweiWangNTU adds geometry/robot dynamics to a 2.5D world model and reports a 42% improvement in real-world imitation-learning success over pure 2D baselines.
It shows that a useful robotics world model needs to understand interaction, not just generate realistic pixels
1
50
6/ Collect physics once, synthesize diversity afterward
RoboDream (arxiv.org/abs/2606.02577 ) by @yuewang314 ‘s team takes a real robot trajectory, then regenerate it with new objects, scenes and viewpoints. Even more interesting: “prop-free teleoperation” lets humans demonstrate motion in empty space and synthesizes the environment afterward.
1
1
36
2/ EgoVLA pushes the same idea even further: human activity itself can become VLA training data.
EgoVLA (arxiv.org/abs/2507.12440 ) by @xiaolonw’s team predicts human wrist/hand actions from egocentric video, then uses IK + retargeting and a small amount of robot data to produce the robot policy. It’s a work back to 2024 which is really visionary.
1
40
1/ Human egocentric video may become a serious alternative to robot-collected pretraining data.
Motion-Focused Latent Action by Jincheng Yu’s team (arxiv.org/abs/2606.18955 ) pretrains exclusively on unlabeled human ego videos, then adapts to a robot with only 50 trajectories.
Robot data may be needed for embodiment grounding, not necessarily for learning every skill from scratch.
3
1
60
Can’t agree more with hot takes from @HarryXu12. Looking forward to your next breakthrough! #IROS #robotics
1
4
129
After the shortage of chips / energy / token, we will have the shortage of bandwidth in 6 months
1
1
49
Bitter lesson again: GPT6 reached SOTA on robotic benchmark with miles beyond existing robot models
Scale matters.
anonymous-report-421.github.…
2
6
222
Amazing work by @Figure_robot. This again proved the importance of in-the-wild data collection for foundation models. The biggest challenge is the access — most businesses won’t allow employees to collect data with headsets . @DeepReach_AI solved this with its global network and has been working with multiple frontier labs.
Introducing Index
Today we're coming out of stealth with Index, the largest & most diverse robot dataset in the world
→ 30min of video uploads/sec
→ 16M video uploads
→ Paid $15M to date
→ 264k downloads
We're committed to spending $1B the next 12 months on data & compute
3
5
207