Daniel_Van_Zant

Lumen Labs | Robotics/Hardware Engineer, Simulation Engineer | San Francisco, CA | ONSITE | Full-time | $150k–200k + equity | https://lumenresearch.co

Howdy! Lumen Labs is building the cognitive layer for physical AI. Nearly every major robotics effort today teaches machines by showing them thousands of hours of human teleoperation. We think that paradigm has a ceiling, and we’re working on nature-inspired architectures for unstructured environments (construction sites, mining, pipelines, battlefield). Pre-seed, backed by top VC funds and operators from robotics/AI companies. Today we run wheeled navigation, we’re dipping our toes into drones, and we’re eyeing weirder platforms like home inspection.

We’re a small team hiring two engineers who will work closely with each other and with our model engineer:

ROBOTICS / HARDWARE ENGINEER: Own the hardware and infrastructure that keeps our research platforms running: the ROS2 (Humble) stack across onboard and offboard compute, sensor integration and calibration (LiDAR, depth, IMU, encoders), the safety pipeline during autonomous runs, and getting trained policies onto real hardware. Will be hands-on with real mobile robots (UART, encoders, PWM, IMU, LiDAR), strong Python + embedded, comfortable reading code you didn’t write, and can debug hardware and firmware at the same time. Full posting: https://desert-bearskin-26d.notion.site/Robotics-Engineer-Ro...

SIMULATION ENGINEER: Own “the gauntlet,” our test library for putting new architectures through their paces: a Gymnasium-style interface on both sides (environment and model) so anything built to spec is compatible with everything else. Mostly MuJoCo and Gazebo today, moving toward Isaac Sim, and some other more bespoke sims. You will be building task simulators for training and testing and keeping them running (GPS-denied drone nav, construction sites, boat cleaning, forklifts, abstract control), including physics, honest sensor models, digital twins of our real platforms, domain randomization, and the sim compute. Dungeon Master. If you have built or substantially rebuilt a simulator of some kind (robotics/RL is closest, but a physics-based game, scientific sim, or flight sim mod counts) this would be a good role for you. Comfortable with the Gymnasium/dm_control way of thinking, can pick up GPU-vectorized sim (MJX, Isaac Lab, JAX) quickly, and test-driven. Full posting: https://desert-bearskin-26d.notion.site/Simulation-Engineer-...

Email hi@lumenresearch.co, mention HN and which role, and tell us about a project you’ve built. If you can link to a trained policy running on real hardware (robotics) or a simulator you built or fixed and something you trained inside it (sim), you go to the top of the pile.

Very cool. Super excited about stuff like this getting rid of software-based paywalls on hardware.

I'm curious has anyone tried hooking up abliterated large language models to systems like this for physical pen-testing? I know there has been lots of hype lately on the capabilities of large language models at finding and exploiting bugs in online software. I know very little about physical pen-testing but would large language models also open up vulnerabilities in the kinds of things this interfaces with or do digital door locks and such have higher levels of security?

Super cool research! I think this is still a formally verifiable domain but different than the kind of formal verification I was thinking. Maybe if LLMs continue to be great at this human science turns into making “fuzzy domains” much more formally verifiable (although I don’t think this is possible in some areas).

To add to this I think there are “reductionist sciences” and “systems sciences”. Reductionist sciences are ones, like physics where there are lots of copies of a thing with near identical properties at different scales (the periodic table would be considered an incredibly short list and that seems to be one of the longest list of “different things” in physics).

Systems sciences are ones where each scale has lots of things with unique properties. If you think of “biological things between 2-4m” the amount of unique objects with unique properties and unique interactions is absolutely mind boggling. These sciences generally do not lend themselves to clean highly parsimonious models. The best you can do is coarse-grain a thing as a system (some very roughly defined set of interactions between different things) with certain behavior while accepting that you are losing enormous amounts of predictive power.

I suspect AI will remain worse at systems sciences for a long time because the “possible space” of models for a particular system is much larger than in the reductionist sciences and navigating it takes a human taste and instinct that LLMs don’t seem to have (yet).

Lots of progress recently in formally verifiable domains. Really curious if this is a product of the machine learning folks behind most large language models being much more likely to have a background in math, physics, etc. Or if it is that large language models have some fundamental limitations in the more "scruffy" sciences like biology.

My guess is that in formally verifiable domains you can use RL-esque processes to get lots of good high-quality training data with reward signals (does it verify?), but the same does not apply to "scruffier" work. I would also imagine that the data volume of high-quality scientific work in "scruffy" domains is too low for the large language models to have a good manifold for "high-quality biology research". Would love to be proven wrong if anyone knows of anyone doing great scientific work with large language models in non-verifiable domains though.

Lumen Labs | Robotics / Hardware Engineer | San Francisco, CA | ONSITE | Full-time | $130k–200k + equity | https://lumenresearch.co

I'm Daniel, co-founder. Lumen Labs is building the cognitive layer for physical AI. Nearly every major robotics effort today teaches machines by showing them thousands of hours of human teleoperation. We think that paradigm has a ceiling, and we're working on nature-inspired architectures for unstructured environments (construction sites, mining, pipelines, battlefield). Pre-seed, backed by top VC funds and operators from robotics/AI companies.

We're a team of two and hiring teammate #3 to own the hardware and infrastructure that keeps our research platforms running: the ROS2 (Humble) stack across onboard and offboard compute, sensor integration and calibration (LiDAR, depth, IMU, encoders), the safety pipeline during autonomous runs, and getting trained policies onto real hardware. Today that's wheeled navigation; we're dipping our toes into drones and eyeing weirder platforms like pipeline inspection. Where we're at: zero-shot sim2real transfer with a policy that runs at 100Hz on an ESP32 and took under a day to train.

You: hands-on with real mobile robots (UART, encoders, PWM, IMU, LiDAR), strong Python + embedded, comfortable reading code you didn't write, and can debug hardware and firmware at the same time. If "full-stack" to you means firmware, a multimeter, and a ROS node in the same afternoon, you'll fit right in. Visa/relocation support for the right person.

Full posting: https://desert-bearskin-26d.notion.site/Robotics-Engineer-Ro...

Email hi@lumenresearch.co, mention HN, and tell us about a project you've built. If you can link to a time you got a trained policy running on real hardware, you go to the top of the pile.

Lumen Labs | Simulation/RL Integration Engineer | REMOTE (US/Canada), then ONSITE SF from fall 2026 | $130k–200k + early equity | Visa OK

Lumen (https://lumenresearch.ai) is building the cognitive layer for physical AI. Most robotics efforts today learn from thousands of hours of human teleoperation; we think that has a ceiling. White paper: http://lumenresearch.co/paper

Team of two ex-founders — you'd be #3. Pre-seed, runway well into 2027, raising a seed later this year.

We test many model architectures against many sims, and hooking them together is our biggest pain point. We use Gymnasium and Shimmy as the plumbing and upstream useful pieces as open-source contributions. You'd build per-sim adapters to a shared interface contract, bridge single-env vs. GPU-vectorized and sync vs. async paradigms, and own our eval harness.

Looking for: strong Python, Gymnasium (especially the vector API), hands-on experience with Isaac, MuJoCo, or Gazebo, and a track record of getting models running in sims where they didn't before.

Email hi@lumenresearch.co about you and your projects — link one where you got a model working in a sim it wasn't built for and you'll jump the pile.

Full posting: https://desert-bearskin-26d.notion.site/Simulation-Engineer-...

Honestly having an AI agent that interviews a interviewees agent sounds like a great "first filter" for certain tech jobs if you do it right. As in "here are the api specs, build an agent that can receive questions and reply with information based on your Resume." Would be vibe-codeable by anyone with skills in an hour. I remember seeing a company a while back that switched to only accepting Resumes through a weirdly formatted API and they said it cut down on irrelevant spam for software jobs immensely.

As someone that enjoys carpentry, I don't think I would be entirely upset about this way of taking the question haha.

I have dealt with this issue as well before. If folks need something more in depth I will use a LLM + some massaging of my own to create a supporting document. Here is an example of a very disorganized conversation and the supporting document I made with it: https://www.danielvanzant.com/p/what-does-the-structure-of-l... It has clear definitions of the key terms. Timestamps for the important moments, and links to external resources to learn more about any of the topics.

I agree with the author, but I would also say there is something above goals and constraints. Values. A set of things that, when comparing multiple options, make the choice clear. An example of some values I frequently use is "What will give me the most enjoyment the furthest into the future? "What will result in the world being a better place?" "What will make me become someone who resembles Jesus more?" They are different from constraints as they don't knock out any options by default. Instead, they make triaging when there are many different things I could be doing much easier, and circumvent my messy intuition which is based on hormones, hunger, weather, etc.

I think values, goals, and constraints are all valuable, but it's a hierarchy. We should create constraints that help us become more aligned with our values. We should create shorter-term goals that make it easy to stay within our constraints.

To support both my point and the authors, here is Benjamin Franklin's "Thirteen Virtues," which seem to be a mix of constraints and values (zero goals): https://fs.blog/the-thirteen-virtues/

I fight to record any presentations I do as often as possible. When I am asked for the slides I send the full recording instead as the way to manage this exact issue.

I'm glad someone else had the same thought. I have been wondering what their "secret sauce" is for a while given how their model doesn't degrade for long-context nearly as much as other LLMs that are otherwise competitive. It could also just be that they used longer-context training data than anyone else though.

Is there any effort to organize scientific literature like this? I know journals often generate tags for papers but those can often be quite poor and restricted to the field The journal is in. I would happily join a volunteer effort to create tags and do some tag-wrangling for scientific literature in my research area.

Have been searching for a deep research tool that I can hook up to both my personal notes (in Obsidian) and the web and this looks like this has those capabilities. Now the only piece left is to figure out a way to export the deep research outputs back into my Obsidian somehow.

I see Cohere, is there any support for in-line citations like you can get with their first party API?

Being able to control how many tokens are spent on thinking is a game-changer. I've been building fairly complex, efficient, systems with many LLMs. Despite the advantages, reasoning models have been a no-go due to how variable the cost is, and how hard that makes it to calculate a final per-query cost for the customer. Being able to say "I know this model can always solve this problem in this many thinking tokens" and thus limiting the cost for that component is huge.

Thanks for the explanation this was very helpful. You've given me a whole new list of stuff to Google. The quasipotential/comittor functions especially seem quite interesting although I'm having a bit of trouble finding good resources on them.

This out of everything, convvinced me. The more I get the "full picture" the more I appreciate what a wondrous thing computers are. I've learned all the way down to Forth/C and from the bottom up to programming FPGAs with Verilog so Assembly may be just what I need to finally close that last gap.

This looks to be very cool will check it out. Wild to see it on a Mario Kart Wii Site, but I guess modders/hackers are one of the groups of people who still need to work with assembly frequently.

Is stochastic calculus something that requires a computer to stimulate many possible unfolding of events, or is there a more elegant mathematical way to solve for some of the important final outputs and probability distributions if you know the distribution of dW? This is an awesome article. I've seen stochastic calculus before but this is the first time I really felt like I started to grok it.

I'm curious from anyone who has done it. Is there any "pleasure" to be had in learning or implementing assembly (like there is for LISP or RISC-V) or is it something you learn and implement because you want to do something else (like learning COBOL if you need to work with certain kinds of systems). It has always piqued my interest but I don't have a good reason in my day-to-day job to get into it. Wondering if it is worth committing some time to for the fun of it.

I think there is a large pool of near minimum-wage white collar workers who wouldn't care about that difference when it comes to executing on their jobs. These are the folks who are already using VBScript, AutoHotKey, Excel wizardry, etc. to automate large parts of their job regardless of any risks and will continue to use these new tools for similar purposes.

I agree that at the business contract level, it's more about sales and regulations than tech. But in my experience working close to minimum wage white-collar jobs, about 1 in 4 of my coworkers had automated most of their job with some unholy combination of VBScript, Excel wizardry, AutoHotKey, Selenium, and just a bit of basic Python sprinkled in; IT, security, and privacy concerns notwithstanding. Some were even dedicated enough to pay small amounts out-of-pocket for certain tools.

I'd bet that until we get the risks whittled down enough for larger organizations to adopt this on a wide scale, the biggest user group for AI automation tools will be at the level of individual workers who are eager to streamline their own tasks and aren't paid enough to care about those same risks.

What made you decide to build for the NES specifically at opposed to building for a fantasy console with similar restrictions but more usability like the Pico-8?

Yeah! I am a computational neuroscientist and I was trying to understand this paper: https://arxiv.org/abs/1906.01678 so I had added the paper to my personal Gnosi and was asking questions about it. I had added this paper https://pubmed.ncbi.nlm.nih.gov/34512508/ to my Gnosi a few months prior after reading through it and completely forgotten about it. As I was asking questions Gnosi started making connections to this earlier paper and helping me understand the new paper in the context of this older paper I had already read and understood, as well as giving me some ideas on using thermodynamic neural networks as a toy model for how Parkinson's can be created by poor management of entropy.

Understanding something new more easily and making new connections and coming up with new ideas is exactly why I have worked hard on maintaining a zettlkasten for the past 6 years, but Gnosi allowed me to make the same kinds of connections much more frictionlessly.

We have similar success stories from our alpha users which include folks interested in anthropology, complex systems, UAP studies, physics, philosophy understanding new tech for startups, etc.

This looks cool. Would be willing to pay for it if there was a RSS option.

Gnosi: https://www.gnosi.ai/ . Y You can auto-generate a constantly updated encyclopedia from a set of documents and conversations with an AI about those documents. Trying to make the process of having a Zettlkasten or personal notetaking system frictionless.