So while the author's points are completely true and valid, an executive will say "True, but Claude will get smarter faster than these problems and in 3 years it'll fix everything" and there's absolutely nothing you can say or do in response to this.

* Here I'm using the alternate definition of elite - someone with money, power, position, or privilege - and not the conventional "barista with hair colour".

That was definitely true in the hand-crafted code era, but I've found all the agentic-type things to be basically the same? Even if you're fairly involved in the code, you're still just mostly reading diffs and editing the odd line, the kind of basic work that's the same across all modern editors.

They do have a small circle of trusted people who they like (like the 1%, lol), but if you're not in, you're just trash that they haven't gotten around to cleaning out yet.

ICE cars come with a variety of add-ons and schemes to improve efficiency: fuel injectors, ECUs, braking energy capture systems (aka hybrid), small batteries for short trips that no one plugs in (aka plug in hybrids), etc.

An LLM/agent can take 100 tokens worth of spec and turn it into 1000 tokens worth of code. We all understand this well enough in the most basic case: you type a one line comment and the LLM will give you the correct 10 line function. But this scales - 10 line comment into 100 line function, 100 line spec into a 1000 line object etc. Depending on the work, you could get more or less than 10x compression. The disagreements are all largely "what what level of resolution should devs be operating", my opinion is that the current sweet spot it roughly 1000 tokens of spec (~750 word doc/3 pages) producing 10000 tokens (~1000loc)

This compression allows your spec to be much denser and more readable than raw code. It can be shared, understood and modified by non-dev partners more easily, gives you flexibility into how it's implemented etc.

https://rodneybrooks.com/why-todays-humanoids-wont-learn-dex...

I saw him post this article on his Bluesky saying that they're the first ones he's seen that are close to cracking this issue (he's an investor/adviser).

LLMs are really good at the sort of tasks that have been missing from robotics: understanding, reasoning, planning etc, so we'll likely see much more use of them in various robotics applications. I guess the main question right now is:

- who sends in the various fine-motor commands. The answer most labs/researchers have is "a smaller diffusion model", so the LLM acts as a planner, then a smaller faster diffusion model controls the actual motors. I suspect in many cases you can get away with the equivalent of a tool call - the LLM simply calls out a particular subroutine, like "go forward 1m" or "tilt camera right"

- what do you do about memory? All the models are either purely reactive or take a very small slice of history and use that as part of the input, so they all need some type of memory/state management system to actually allow them to work on a task for more than a little while. It's not clear to me whether this will be standardized and become part of models themselves, or everyone will just do their own thing.

Comments URL: https://news.ycombinator.com/item?id=47770124

Points: 113

# Comments: 164

I've been using AI heavily to do this, so everything is in ROS2 since it's "standard" and AIs have pretty good training for it. I can see how it's annoying and suboptimal if you're writing manually and after a more integrated system, but it's been pretty good for getting up and running because it's "standard" and kinda plug and play. I see why you'd want to rewrite it for production, the endless processes and nodes and startup processes can get annoying

One of the more useful things I've done so far is actually not robotics related directly, it's a Godot based "game" with a ROS bridge that lets me drive the robot from Foxglove, which I will eventlly get a vlm based agent to drive. Seems much easier and faster than Issac Sim for getting started with.

So first, why is robotaxi not happening? Well, it is just not on the scale they need it to. In addition to Waymo, many other companies are working on driverless systems and they're all making good progress. nVidia has a full stack competitors based on VLA models, as do Chinese manufacturers etc. In short, we'll see capable self driving systems from probably a dozen companies world wide. Tesla has some advantage here, but real world acceptance and scaling are slow, trust in Tesla is low and most of their cars are physically incapable of running the size of models L4 systems will require. My 2022 Model Y has roughly the same processing power as my Rovers Jetson Orin Nano and that thing isn't driving itself.

The humanoid part is even worse, Tesla has absolutely no structural advantages there - they're no better than Chinese companies at making the hardware and they're no better than the AI Labs at developing the models required. Also, again low trust. Tons of people will absolutely not let a tesla bot in their house. There's no reason to think they'll be wildly successful, and the actual capabilities and economic usefulness of humanoids are still some years away.