Alan Helmick

@robotsdigest Yes I've been teaching my robots to perform actions on a 3B and they have no problem porting the knowledge... But I use a custom sensor pack The simulation to real gap doesn't exist in my world

ViPRA learns robot actions from unlabeled human + robot videos. Instead of requiring expensive action labels, it extracts motion-centric latent actions and predicts future frames, giving robots a prior for how the world changes.

@akshay_pachaar Claude does not pay attention to Claude.MD at all! doesn't pay attention to its memory, makes the same mistakes over and over and if you only put rules in a Claude.MD almost nothing will change! Hooks are the only way. Don't believe what you read!! 🤣

A single 𝗖𝗟𝗔𝗨𝗗𝗘.𝗺𝗱 file just hit 15K GitHub stars. (derived from Karpathy's coding rules) Andrej Karpathy observed that LLMs make the same predictable mistakes when writing code: over-engineering, ignoring existing patterns, and adding dependencies you never asked for. If you've used AI coding assistants, you've hit all of these. But here's the thing: If the mistakes are predictable, you can prevent them with the right instructions. That's exactly what this 𝗖𝗟𝗔𝗨𝗗𝗘.𝗺𝗱 does. You drop one markdown file into your repo, and it gives Claude Code a structured set of behavioral guidelines for your entire project. This is a big deal. - Built entirely around prompt engineering for AI coding assistants - No framework, no complex tooling, just one .md file that shapes behavior Developers are moving past "use AI to write code" and into "engineer the AI's behavior so the code is actually good." The Claude Code ecosystem is growing fast, and the best tools in it aren't always software. Sometimes they're just well-crafted instructions. 100% open-source. I've shared a link to the GitHub repo in the next tweet!

@aaronistan The Iron Man robot! how fun! not incredibly useful for the price but a lot of fun. Fold those clothes you'll get donated to a community college!

Introducing Lume. A lamp that does your chores. Order now. Shipping this summer.

@FarzaTV @theanuragdev Yeah totally rip off the Microsoft Clippy and everything it does. 😂 I'll just change the name a little bit and completely copy the entire functionality. 😂 original doesn't mean scalable though does it, what a world

I built this thing called Clicky. It's an AI teacher that lives as a buddy next to your cursor. It can see your screen, talk to you, and even point at stuff, kinda like having a real teacher next to you. I've been using it the past few days to learn Davinci Resolve, 10/10.

@VectorWang2 Nice work! you know I went the other direction...

Looks sharp😁😁

@tianle_cai Yep that's what I've been working on for the last two years Looking forward to reading your paper github.com/alanchelmickjr…

Can we turn part of an LLM's weights into long-term memory that continuously absorbs new knowledge? We took a small step toward this with In-Place Test-Time Training (In-Place TTT) — accepted as an Oral at ICLR 2026 🎉 The key idea: no new modules, optional pretraining. We repurpose the final projection matrix in every MLP block as fast weights. With an NTP-aligned objective and efficient chunk-wise updates, the model adapts on the fly — complementing attention rather than replacing it. 📄 Paper: arxiv.org/abs/2604.06169 with amazing @Guhao_Feng @Roger98079446 Kai @GeZhang86038849 Di @HuangRubio

@MeRTcooking The old line follower Now I don't have to use painters tape! 🦾

Our MRReP has been featured on IEEE Spectrum Video Friday, "Your weekly selection of awesome robot videos"!!! MRReP: Mixed Reality-based Hand-drawn Reference Path Editing Interface for Mobile Robot Navigation - IEEE Spectrum Video Friday spectrum.ieee.org/video-humanoid… #ROS (1/N)

🚀 Excited to share our work: MRReP🤖! MRReP: Mixed Reality-based Hand-drawn Reference Path Editing Interface for Mobile Robot Navigation! With MRReP, users can directly draw a path on environment using hand gestures, and the robot follows the path! #ROS2 #ROS #Unity (1/N)

I found another elderly robot a job Didn't you guys know that's what I do, I'm a robot employment specialist. #robots As one of the resident Roboticist at Frontier Tower I encourage y'all to join us for true frontier innovation. @frontiertower

@MeRTcooking The old line follower gets an upgrade!

@FarzaTV All of this would be OK if you were kind and professional to everyone that you meet but you were a real dickhead to me for no reason so I really hope anything you do burns and dies Maybe you could code something that teaches you not to be a discriminatory bitch

I decided to join Y Combinator, again. This would be my second time! Not fully sure what I'm working on yet. But, I'm sure I'll find something in time as I wander and ship. I'm a little scared to do the whole build a company thing again ngl, but mostly excited. There's never been a better time to work on the ideas in my head. The batch started this week. Starting a company at 23 vs now starting a company at 30 feels so different. At 23 (when I did YC in 2020), naivety was there. At 30 I guess I know how difficult it all is. It's not surprising to me that most people in YC are aged 19-24. Still, I feel like I have the naivety of a 19 year-old, but, with the mental of a guy who's been through a lot and learned a lot. So, I'm bullish. Let's see what happens. You'll probably see me launching a lot of random stuff over the next few weeks especially. Also, I am blown away by the number of founders in the batch walking up to me telling me they credit being at YC to @_buildspace. It's so wonderful, and warms my heart. I often struggle to stop and understand the value of my past work because I'm so interested in the future. So, this was nice. It's funny, many saw me irl and freaked out thinking I was joining as a YC partner and were very very surprised to hear I was joining as a founder back in the dirt alongside them haha. Most founders never start another company and usually turn into VCs or get a high-tier job at a big company. I do not blame them. And honestly, that would be the easier more secure path for me especially as I begin thinking about family. But, idk. I feel like my ideas are important. And even though I don't have a specific "This is the idea I'm excited about" it's more a feeling of "I should explore my ideas...I would regret it if I didn't". Especially in 2026, at the epicenter of one of the greatest inventions of my lifetime. Every time I think about getting a job (of which I've been offered many great ones) that voice in my head comes back and says to give my nascent visions a shot. So, gonna try :) Maybe I flop, maybe I don't, only one way to find out. I'll be dropping weekly updates on YouTube if you're interested. I put one out last week that talks more in depth around the story of how this YC stuff even happened randomly, why I'm doing this again, my imposter syndrome and how I think about it, and other stuff. I'll link it below. Lets see what happens!! See y'all.

@FarzaTV That would be OK if you were a discriminatory bitch toward some people and nice towards others the fact that you don't treat everyone the same kind of makes everything you do a piece of shit.

@VectorWang2 @bowenwen_me 🔥

Fast Foundation Stereo + SAM2 basically = zero shot Foundation Pose😂 No CAD model, no object image, just click the target. Run directly on my 3070 at 13fps, with a 30$ stereo camera (calibrated in 10min) Thanks @bowenwen_me for his contribution to the community!

I packaged up the "autoresearch" project into a new self-contained minimal repo if people would like to play over the weekend. It's basically nanochat LLM training core stripped down to a single-GPU, one file version of ~630 lines of code, then: - the human iterates on the prompt (.md) - the AI agent iterates on the training code (.py) The goal is to engineer your agents to make the fastest research progress indefinitely and without any of your own involvement. In the image, every dot is a complete LLM training run that lasts exactly 5 minutes. The agent works in an autonomous loop on a git feature branch and accumulates git commits to the training script as it finds better settings (of lower validation loss by the end) of the neural network architecture, the optimizer, all the hyperparameters, etc. You can imagine comparing the research progress of different prompts, different agents, etc. github.com/karpathy/autor… Part code, part sci-fi, and a pinch of psychosis :)

The single greatest challenge I find in building physical AI and robots is teaching them to dance 😂🕺🏼@frontiertower

@Ryan_Resolution @IsaacSin12 @ThomasSchicksal @QILIU9203 @VectorWang2 xlerobot.store <- super improved version. We're behind on orders, so three weeks out. Guaranteed better than any other XLE version on the market or your money back. customer reviews of our upgraded electrical system say it makes all the difference. We call it XLPowerE

We just upgraded XLeRobot 🚀 Built by the MakerMods team @IsaacSin12, @ThomasSchicksal and @QILIU9203. • Easier to build • Improved chassis • Reduced 3D-print time and material • Designed in collaboration with the original author @VectorWang2 Fully open-source Full build guide GitHub release coming soon

@MichaelElabd Yep I built a memory system and it cost me $120 a day so I've been working to bring that down to the monthly, it's 3 weeks FROM Beta but I saw your post, i'll have test benchmarks by next Friday. Input? github.com/alanchelmickjr…

Honestly, I think memory is the biggest blocker to continual learning right now. Here's what keeps me up at night: How do we use memory to not repeat the same mistakes? how do we teach models to selectively remember and forget? When to surface the right context? Humans do this naturally through consolidation, interference management, and contextual binding, however, we haven't figured out how to replicate it. The gap between human hippocampal systems and current LLM memory architectures reveals a fundamental challenge: We've basically built two extremes: models that bake everything into parameters (rigid) or retrieve stuff mechanically with RAG (fuzzy). True continual learning requires that we crack this code of intelligent retrieval; not just what to store, but what to suppress, when to reinforce, and how to let old knowledge gracefully fade without catastrophic interference. Loved reading this survey because it offered a birds-eye view of memory architectures in LLMs and multi-modal models (also loved the brain-inspired taxonomy, nice touch!) Will do my best effort to systematically map it out. The Three-Part Framework: They structure memory around the neocortex-hippocampus-prefrontal cortex analogy: Implicit memory / the neocortex covers parametric knowledge baked into model weights, including techniques for memory editing (like ROME and MEMIT that surgically modify weights to update facts), knowledge injection via adapters like LoRA, and memory unlearning for removing harmful content. Explicit memory / the hippocampus examines external retrieval systems; RAG architectures, vector databases, knowledge graphs. They detail how memory can be organized at different granularities (documents, chunks, sentences, graph structures) and optimization time (training-free, joint pre-training, sft, etc). Agentic memory / prefrontal cortex explores how autonomous agents maintain short-term memory (CoT++) versus long-term memory (external databases of facts, historical trajectories, user feedback, etc). I love this framework for thinking about memory but I think the biggest contribution of this survey beyond its categorization is identifying open problems: memory contamination / hallucination, the computational burden of large-scale retrieval, when to retrieve vs rely on parametric knowledge, and the challenge of memory consistency across long interactions. All areas I would love to see more papers in!

@drfeifei I've gotta learn how to make a resume... this woman has the coolest stuff. 🦾 i'm over here building the $5000 or less robot and you guys are building the brain I need... Johnny5 will be alive... And yes I told everyone don't get hung up on Tele-op they'll have this out by '26

Interactive 3D world model is a highly intuitive representation for learning robotics actions in dynamic and complex environments. Here is our most recent work on this 🤖

@FarzaTV used by 30 community college students for 2 years talkverse.ai Safe was for learning, anti cheating tutors Note taker allowed you to do your homework during class and then study the lecture later Defunct now, they all used the chat to cheat 😩 So FUN is #1, reward

New essay. Build education machines. Enjoy. farza.com/build-edu-mach…