that’s exactly my question. why can’t he depend on C API and just consider it is memory safe and all.

i am dealing with ros these days for calibrating my custom hardware. it really is gROSs.

Thankfully, I have fully isolated ROS parts into their own dockerfiles.

“one in 4 eventually drown”… does this mean it’s a recent evolutionary trick and selection hasn’t settled yet.

It might just be because they have to die somehow. Probably 1 in 4 humans' hearts eventually stop working too.

> The dunk-and-spin maneuver is risky, Yarger notes, as roughly one in every four dragonflies eventually drowns.

It's not that risky, with about a 99.99% success rate. But in their lifespan of a few weeks, they will still be doing thousands of them...

Selection is a balance. It is hard to judge what is being selected for when you only look at one factor (drowning).

Aging versus cancer is an example of balance: one theory is that age related diseases are a side effect of selection forces against cancer: programmed cell death after X reproductions (telomeres) are a general anti-cancer defence but the cell-death has aging effects. Also beware that selection is for successful reproduction: death after reproduction is not so relevant to evolution.

if they really want to bet on robotics, I want them to release a $10 variant of jetson board.

For 10 USD you get an ESP32S3 board, which can do basic computer vision tasks. For example using OpenMV or emlearn-micropython. For 15-20 USD you can get a board that includes an OV2640 camera. Examples would be XIAO ESP32S3 Sense, LilyGo T Camera S3 or "ESP32-S3-CAM" board from misc manufacturers.

yes that’s what i am working on these days but there is a need for a generally available neural chip (see google’s coral as one attempt). in my tests, esp32s3 is very very slow for any model with conv2d involved.

i just want a tiiiny gpu for $10 so i can run smaller models at higher speed than possible with xtensa/rp2040 having limited simd support etc.

Are you utilizing the SIMD and acceleration instructions in the S3? What kind of performance are you seeing?

Neural accelerators are coming into MCUs. The just released STM32N6 is probably among the best. Alif with the U55/U85 has been out for a little while. Maxim MAX78000 has a CNN accelerator out for a couple of years. More will come in the next few years - though not from Nvidia any time soon.

I'd love to hear more about your experience with Coral. Sounds like that'd be a good fit for a tiny GPU to run models with conv2d?

A few weeks ago they have reduced the price of the Orin Nano development kit from $500 to $250, while also increasing a few of the performance limits that cripple it in comparison with the more expensive Orin models.

Previously it was far too overpriced for most uses (except for someone developing a certified automotive device), but at the new price and performance it has become competitive with the existing alternatives in the same $150 to $300 price range, which are based on Intel, AMD, MediaTek, Qualcomm or Rockchip CPUs.

When they reduced the price of the dev kit, they priced it below the low volume sales price of the cheapest Orin Nano 4GB module. Presumably the module prices go down when you buy in bulk but for small volumes it was (is still?) cheaper to buy the dev kit and throw away the carrier than to just buy the module. Granted the dev kits went out of stock pretty quick.

damn you are right!

i wish this didn’t use mpu6050 imu, which is obsolete and unavailable it seems.

but i guess they used it due to existing code/drivers widely available for it and esp32.

Agreed. I'm in the market for an imu and thought "ah the MPU 6050, I've heard of this one a lot even recently" and it's obsolete. This is typical of consumer to adafruit/sparkfun/aliexpress levels where they have countless old stock of cheapy proto boards to buy from, but if you're designing a whole new thing from scratch, that's inexcusable.

i see it as just accelerating normalizing cheap drones everywhere.

just upload it to jlcpcb and get it delivered with components soldered.

this is amazing. on similar note, I have spent last few months trying to fit visual inertial odometry into esp32. Combining that with this would be insane (and so cheap!)

I've had similar thoughts and have been working on firmwares for the esp32. My contact info is in my profile. Hollar at me and lets compare notes.

i don’t see your contact info.

My mistake. Updated.

What do you mean by this? Do you mean using an onboard camera for navigation and/or orientation? I’m very interested.

kinda how we do it. language is just an io interface(but also neural obv) on top of our reasoning engine.

It’s not just a protocol buffer for concepts though (weak wharf Sapir, lakoff’s ubiquitous metaphors). Language itself is also a concept layer and plasticity and concept development is bidirectional. But (I’m not very versed in the language here re ‘latent space’) I would imagine the forward pass through layers converges towards near-token-matches before output, so you have very similar reason to token/language reasoning even in latent/conceptual reasoning? Like the neurons that nearly only respond to a single token for ex.