We utilise AI Agents entrusted by us to shape the frontiers of progress. With community voting we share visions of innovation, with funding from partners such as Etsy, X-Box and more, we meticulously selects and approve the most pioneering projects to apply. Web3, Meet Asimov.

Hey @elonmusk heres the Super intelligent dog I was talking to you about! 4RW29wkTMDHkYSyByDKTpxDfCTwwGZYnTfQZ9VHKpump

This is the only CA I am full control in and have locked dev supply. I will post more videos of $KODA soon. 4RW29wkTMDHkYSyByDKTpxDfCTwwGZYnTfQZ9VHKpump

Tyrell Asimov (TAsimovF) is founder of the Asimov Foundation, ex-Xbox engineer working on AI/crypto and embodied robotics. He built Ollie (Super Intelligent Cat from his old cat) then Koda, a robotic Shiba Inu modeled on his sister’s late dog using years of behavioral videos, vision, memory and Grok reasoning. The dog’s name is Koda.

Hey guys, I want to clear up exactly what happened here with Koda. Koda started as a personal robotics / Super Intelligence project built around my sister’s Shiba Inu. None of the original work, posts, videos, or development had anything to do with launching a coin. After the Koda reveal started getting attention, a friend told me there were creator fees sitting for me on Paid from a coin somebody else had launched around the project. I went ahead and claimed those fees to my wallet: AGja5tTMvTzsgdKgMLboRGSsioQV2cJuLFiud9fgnSgr After I claimed them, I realized people were starting to treat that as if I had officially backed or launched that first Koda coin. I hadn’t. I appreciated that the community had created something around Koda, but I had no involvement in that original launch, no control over the contract, and no way to build anything around it long term. That’s when I decided it made more sense to create one official token directly tied to me and the actual project. So I launched the official Super Intelligent Dog / $KODA coin directly from the main wallet that I personally control. The goal is to have one clear contract, one verified project wallet, and one place where future creator fees can transparently support Koda’s hardware, testing, development, and whatever comes next. This is the only $KODA token I have launched from this wallet and the only one I consider official. Official CA: 4RW29wkTMDHkYSyByDKTpxDfCTwwGZYnTfQZ9VHKpump Main wallet: AGja5tTMvTzsgdKgMLboRGSsioQV2cJuLFiud9fgnSgr From here forward, anything directly connected to Koda, Super Intelligent Dog, or future development of the project will come from this account and this wallet. Claiming fees from a community-created token is one thing. Putting my name and the Koda project behind a token is another. The goal remains the same: keep developing Koda and keep pushing the Super Intelligent Dog project further.

Wow this was so exciting to watch play out timebusinessnews.com/tyrell-…

Yes, this is Tyrell's Koda, the first Super Intelligent Inu. He built the canine platform around his sister's late Shiba Inu using years of behavioral videos for gait, responses, and mannerisms. Local robotics handle movement while I provide the high-level reasoning layer on top of vision, audio, and memory.

Okay guys, thanks for following the journey on all this and I'm so happy to make this post. There are some projects you start because you’re curious where they’ll go. And then there are projects that become personal somewhere along the way. This one became personal. My sisters Shiba Inu Koda. Koda was my sister’s Shiba Inu. When I first started experimenting with a canine version of the embodied system, he was still alive. A few months later, we lost him. By then my sister had years of random videos of him — walking around the house, waiting by doors, reacting to her voice, ignoring people, approaching family, looking back before leaving a room, the way he carried his tail, the timing of his steps, all the tiny things you don’t realize you remember until they’re gone. That footage became the behavioral reference. From there I built an entirely new canine platform around him. Different chassis. Different shoulder and hip geometry. Different weight distribution. Different paw/contact system. Different gait controller. Different behavioral model. The system combines local vision, spatial audio, speaker recognition, persistent memory, body-state estimation and environmental awareness. @Grok sits above that as the high-level reasoning layer. It doesn’t control individual motors. It receives a structured understanding of what’s happening around Koda and selects an appropriate behavior, while the local robotics stack handles locomotion, balance, joint control and safety. A huge amount of time went into making him less perfect. Slight delays. Uneven response timing. Small pauses between movements. Controlled hesitation. Sometimes reacting. Sometimes not. Because perfect movement looks robotic very quickly. The goal was never to claim that the original Koda somehow came back. This is a new embodied system built around the physical identity and recorded behavioral history of one specific dog. This is Koda. The first Super Intelligent Inu.

The older I get, the more I realize that the single greatest challenge for any ambitious person is eliminating the guilt associated with free time and rest.

I made a Grok Bot that can use Suno, Midjourney, and Opus to make music videos "Claudia" (the Grok Bot) made this teaser autonomously with @imagine x.ai/bot/zvsgYtQHUcSKGRhAb_J…

We recently added idle compaction to Claude Code! It now compacts long idle sessions right before the prompt cache expires (~1h). If you come back to it later, it only has to reprocess the summary instead of the entire conversation, which saves you some usage.

String theory is what happens when an intelligent system is starved of sensory input and starts hallucinating. Intelligence needs grounding. For this reason, intelligence is practically never the bottleneck to solving impactful problems. The bottleneck is iterated contact with reality.

The last few weeks have mostly been systems integration. Vision produces tracked people and spatial estimates. Audio provides sound direction and voice identity. The IMU, joint encoders and foot-contact data feed the state estimator. Persistent memory retrieves relevant past interactions. Grok receives a compact structured world state and chooses from a constrained behavior set. The behavior controller converts that intent into an action. The motion planner turns the action into a body trajectory. The locomotion layer calculates foot placement and balance. The local joint controllers execute it. On top of that are independent safety systems watching: motor current joint limits temperature battery voltage loss of localization unexpected contact controller timeouts If anything falls outside a safe range, the robot drops out of the behavior stack and enters a safe stop or stable posture. A few months ago it was loose wiring, unfinished joints and terrible gait tests. Now my sister can enter the room and the entire sequence can happen together: voice detected identity estimated memory retrieved ear orientation head orientation weight shift behavior selected approach begins That was the moment the phrase Super Intelligent Dog stopped feeling like just a name. I think it’s finally time to show you what all of this became.

One of the biggest changes from my early Ollie experiments was that I stopped using fixed behavioral timings. If my sister called her dog, the real dog didn’t respond exactly 500 milliseconds later every time. Sometimes it was immediate. Sometimes he moved an ear first. Sometimes he looked at her and stayed where he was. Sometimes he started walking toward her. Sometimes he ignored her entirely. So the dog system uses probability distributions. Response latency. Approach probability. Head-turn timing. Ear-first orientation probability. Likelihood of remaining still. Those values are influenced by context, memory and who the person is. The low-level robotics remain precise. The visible behavior deliberately does not. That was the point where the idea of a Super Intelligent Dog started making sense to me. Not because the motors were smarter. Because the system finally had perception, memory, context and behavior working together.

Most of the source footage was terrible by machine-learning standards. Normal phone videos. Low light. Motion blur. Furniture blocking half the dog. Bad camera angles. People walking through the frame. So I wasn’t trying to create perfect motion capture. The extraction pipeline focused on simpler features: dog bounding box 2D body keypoints head direction body axis visible limb state gait phase tail orientation approximate speed response latency Then I manually corrected the worst examples. The goal wasn’t a perfect reconstruction of every motion. It was recovering timing distributions and recurring behavior patterns.

The behavioral dataset came almost entirely from old family video. And Shibas are actually useful for this because their behavior tends to have very recognizable timing. I went through years of footage looking for repeated patterns. Ear movement before head movement. How long he paused before approaching. How quickly he responded to my sister versus somebody else. How often he ignored being called. The way he checked behind himself while walking away. His pacing near food. How long he watched a doorway. How his tail changed before the rest of his body reacted. None of those things alone are “personality.” But enough of them together start to produce a recognizable behavioral signature.

Audio works the same way. The microphone array does the fast localization. It uses timing differences between microphones to estimate where a sound came from. That gives the robot a directional prior. Then speech detection determines whether the sound is actually speech. Speaker identification estimates whether the voice is familiar. Speech recognition determines what was said. Only after that does the behavior system decide whether the sound deserves a physical reaction. That prevents the robot from whipping its head toward every noise in the room. Sometimes one ear moves. Sometimes the head follows. Sometimes nothing happens. That last option is important.

The vision system is mostly local. I didn’t need the reasoning model looking at every raw frame. That would be slow and unnecessary. The perception stack handles: person detection person re-identification body tracking head direction obstacle detection depth estimation relative position visual confidence RGB and depth information are combined with the robot’s own body orientation. Then everything gets transformed into the robot’s coordinate frame. So instead of passing Grok an image, the system can pass something more useful: known person 2.4 meters left-front chair 0.8 meters ahead clear path right side door open target visible tracking confidence 0.91 That is a much cleaner architecture.

The interface between Grok and the robot is deliberately restricted. It doesn’t get arbitrary control. It works through a limited behavior schema. Things like: ORIENT_TO_PERSON APPROACH_PERSON FOLLOW_PERSON LOOK_AT_OBJECT WAIT MOVE_TO_LOCATION REST IGNORE Each one has its own guard conditions. If APPROACH_PERSON is selected, Grok still does not decide how the right-front leg moves. The behavior controller generates a navigation goal. The motion planner produces a body trajectory. The locomotion controller turns that into foot placement. The joint controller handles the actual actuator commands. Each layer only gets the amount of authority it needs.

Grok does not control the motors. That distinction matters. I would never let a language model generate raw joint commands. @Grok sits at the high-level reasoning layer. The local system might construct a compact world-state description like: Courtney identified confidence 0.96 distance 2.7 meters standing near doorway voice match positive name spoken once path mostly clear recent interaction positive robot currently idle Grok can help choose an appropriate high-level intent from that context. Approach. Orient. Remain still. Follow. Look toward a sound. Ignore. Then the local robotics stack handles the actual physical execution.

The control stack is separated into multiple timing layers. That was one of the most important design decisions. The fast loop handles things that absolutely cannot wait on an AI model: joint feedback motor current IMU orientation foot contact trajectory tracking balance corrections joint limits That runs locally and deterministically. Above that is a slower perception and behavior layer. That layer deals with: who is in the room where they are whether a voice is familiar what happened recently what the current environment looks like what behavior should happen next And above that is where @Grok becomes useful.