Kirill Danilyuk

Grok is actually amazing. The selling point for me is the Ara voice + argumentative personality. It pushes back hard and the voice itself is superb, it handles my Russian accent like a charm. And it feels like a real conversation in a way that's hard to describe.

There is a computer vision course I wish I had taken properly before starting out in perception: web.stanford.edu/class/cs231a/ No, it is not CS231N, the famous CNN-heavy course by @karpathy. It is CS231A, its geometry-first counterpart. You need to know the fundamentals: camera models, calibration, SfM, stereo/mono depth estimation, scene flow, etc. Of course, end-to-end transformers can learn a lot of this on their own, but at least you need to know what is going on and why. Amazing course notes and assignments. If you complete the course and can reimplement something like LSS (Lift, Splat, Shoot), you'll be well positioned for the job regardless of current trendy architectures. Do not skip this part.

The bottleneck for AI engineers has shifted from producing code to verifying it. And verification is getting really hard: ML bugs don't crash anything, they just silently train the wrong thing. And no, you can't catch what you don't already understand. Years ago at Yandex I interviewed ML candidates with exactly this task: here's the training code, find the subtle bugs. Back then it was a filter for senior engineers, now it's the core of the trade. What a time to be alive.

Progress podcast's signal to noise ratio is extremely high. And this is only episode two. Keller and Ryan are ideal role models for any serious entrepreneur: people who take real risk, build real businesses, and compound company value year after year for decades. @fuelfive, outstanding work on guest selection, please keep it up!

@NikolausWest @rerundotio @maticrobots Also I noticed Rust attracts enthusiastic folks who view programming as a craft rather than just a way to earn money. Ruby and Rails got pretty much the same kind of people in the 2010s.

@kd_at_x @rerundotio @maticrobots Agree with all of this (unsurprisingly) and for those that still believe hiring will be important in the agent world I'd also very much second "people really love learning and writing in Rust"

Mark my words: Rust is about to explode in robotics — and coding agents are the main driver. @rerundotio and @maticrobots were right about Rust from the beginning. The #1 requirement for agentic coding (e.g., in Claude Code) is a tight feedback loop with clear error signals. Rust's compiler is perfect for this: if it compiles, it's (mostly!) memory-safe and data-race-free. C++ lets agents generate code that compiles fine and segfaults at runtime. That's a terrible feedback loop. Rust's package ecosystem (crates.io + Cargo) is another huge advantage. Agents can discover, add, and build dependencies with zero human intervention. There's nothing even close in the C++ world, I'm sure even C++ guys are not going to argue with this. "But C++ has all the libraries!" - this is true today. But the libraries that matter at runtime (inference, sensor drivers, motion planning) are increasingly available. And you don't need PyTorch at runtime. "But there's way more C++ to train on!" - there's also way more C++ footguns to learn from. Quality > quantity. Rust's smaller, more modern corpus arguably produces better output for agents with fewer hallucinated patterns. OK, what about hiring? Agents dramatically reduce the importance of team size. One strong engineer with agents writing Rust > a team writing C++ and debugging segfaults in runtime. Also, people really love learning and writing in Rust!

@paveliakovenko Yes, for a really good world model they need to simulate other sensors rather than just pure vision. Sound and voice should also be simulated!

@kd_at_x 💯💯💯 Also it’s pretty cool that they are making lidar cloud simulation as well. However I don’t know the quality of it, as it could be not really easy thing to do

There is one thing I'm really certain about after my years in self-driving / physical AI — there is no safe self-driving and robotics without an extremely high-fidelity neural simulation. The dreaded long tail of self-driving is not solvable by just "driving more". Right now, simulation and world modeling is really the Holy Grail of robotics. Kudos to @Waymo and @GoogleDeepMind, this is an amazing application of Genie 3. waymo.com/blog/2026/02/t…

Just a couple hours ago, @karpathy mentioned the revival of RSS and making an RSS reader for popular HN blogs—so I vibe-coded one in an hour with a nice TUI in Rust (I know zero Rust). It works exactly how I want: extremely minimal, distraction-free, with powerful navigation—and any feature you want is delivered on the fly. Perfect for exploring these blogs, most of which I'd never read before. Not sure if I should open source it or just share the CLAUDE.md. We live in crazy times.

"Built for agents". If there's one phrase that is very 2026, that's the one.

@external_idea @Tesla @robotaxi I hadn't. I looked it up, it is a move in the right direction. Same with sensors getting dirty: the best solution is to optimize car aerodynamics to keep debris and droplets away from sensors.

@kd_at_x @Tesla @robotaxi Did you see Tesla's patent to stay visual only by reducing sun glare? Humans drive on visual and sound as does Tesla. The complexity of Lidar is truth.

A couple days ago I took my first @Tesla @robotaxi ride in Austin. As someone who builds self-driving technology, this felt different. I've taken countless rides in Waymos and other autonomous vehicles. I know how the LiDAR and map-based systems work. Don't get me wrong, it is sophisticated tech, but there's no magic in that type of self-driving. Every maneuver, even the impressive ones, traces back to a carefully curated HD map. It surely works, but you can always explain it away. The Tesla felt different, starting with the visualization. Watching the map build itself in real-time — lane lines, pedestrian crossings, curbs, road edges and the whole perception stack — all from cameras alone, feels like the real magic. That's the kind of AI that works everywhere. It took me 25 minutes to get to my destination and I noticed no interventions from the safety driver. One highlight was how it handled an unprotected left turn across a railroad crossing — the crossings light still red after a train had just passed. The car performed great, see the video. I can hardly imagine such behavior was hardcoded. While I'm not dismissing the LiDAR-based approach due to its safety, I'm genuinely impressed by what a camera-based approach can achieve.

@external_idea @Tesla @robotaxi Yes, it ultimately boils down to the planning part. Tesla has a clear edge here because of the data coming from drivers. Lidars help with perception (like sun glare or seeing in darkness), but add _a lot_ of complexity with sensor fusion and calibration.

@kd_at_x @Tesla @robotaxi What would Lidar add? Seriously, more sensors means more to process before making a decision, which is time and cost. Tesla has focused on decision making instead for their Robotaxi.

There will be no better year to make things than 2026. Let's build America.

2025 was the year when self-driving tech and operations made a substantial leap forward. Waymo started driving commercially on highways — a major bet that shows their technology scales. Tesla released FSD v14 that carried us all the way from Mountain View to SF and back at night without a single intervention. We (Avride) launched commercial service in Dallas with Uber. The technology is here to stay and will only get better and bigger. This is great news for robotics and autonomous vehicles. Robotics is notoriously hard due to the chaotic nature of the real world — this is why robots started in warehouses where chaos is minimal. Self-driving comes next. Roads are still somewhat structured, but the long tail of rare events remains unsolved. But we make progress. The next frontier is general robotics. The breakthrough won't come overnight, but we are moving in the right direction. What a time to be alive!

ODS.ai wins 2018 Data Science Bowl competition. Congrats! twitter.com/NvidiaAI/statu…

Our paper on 3D Cell Nuclear Morphology was accepted to Computer Vision for Microscopy Image Analysis Workshop at #CVPR2018! Updated pre-print is on bioRxiv: doi.org/10.1101/208207. Full dataset & code coming soon. Looking forward to presenting in Salt Lake City! #cvmi18 #cvpr18

@ingdir поздравляю, Макс! :)

@yaroslav Из серии "Only in SF"