Engineer building everything in public Background: CGI | GameDev | Fullstack | ML | Web3 Goal: $2k MRR | Current: $150 MRR
Ukraine
- Tweets 631
- Following 200
- Followers 1,197
- Likes 331
soldered and tested
first run in the car failed :) spent a while debugging...
turned out i plugged the HAT into the wrong row of the sticks header, heh
now it runs smooth, ~2k frames/s
0 dropped frames in 5 min
driving will show how numbers behave on the road
previous setup was too complex
esp32 + CAN + dc-dc, and the stickS3 read it over wifi
so I designed a HAT for @M5Stack StickS3
it powers the stick from the car's 12V and converts CAN signals for the stick
i believe it's way more handy than a separate dongle
case is printed
time to solder and test
4
12
230
33,090
everyone is training fruit flies to do some crazy stuff, but what if we don’t train one at all and put part of its connectome on an ESP32-S3... lets try
here is what I was able to fit on it:
- 48,311 neurons
- 9.46M connections
no training, and it runs fully on the chip
spider approaches then escape neurons activate and the fly jumps
it works! fly connectome with simplified neuron dynamics
the bars on top show escape neuron activity for each side
the simulated spiders distance and direction stimulate the fly’s looming-detector neurons.
activity travels through the connections, then I read 8 escape neurons
when one side crosses a threshold, it triggers a jump away from that side (not always perfect)
I coded the world, sensory input and jump rules
but the connectome produces the neural response
and seeing it in action on this tiny board is sooo cool!
1
34
4,305
previous setup was too complex
esp32 + CAN + dc-dc, and the stickS3 read it over wifi
so I designed a HAT for @M5Stack StickS3
it powers the stick from the car's 12V and converts CAN signals for the stick
i believe it's way more handy than a separate dongle
case is printed
time to solder and test
connect @M5Stack StickS3 to esp32 to decode tesla CAN bus.
this is the first test, definitely a lot of room for improvement.
in dbc insane amount of data that can be rendered on device.
but probably everything that you want to know is power consumption/charging and bms data.
kudos to @steveruizok
3
13
219
64,340
Jev by @typesafeai in action!
20.7k YouTube comments classified in 2m 27s for just $0.20
- p50 319ms, p95 556ms per comment
- 140 comments/sec
- sentiment + emotion + intent + spam/toxic, each with confidence
mostly Apple WWDC videos
and results are kinda funny:
43% negative
#1 intent is criticism (5.1k)
my classification rules are probably not ideal...
but you can tweak them and rerun the whole batch for 20 cents! LFG!!
After co-inventing ChatGPT, I kept asking myself: why have superhuman chat models not led to AGI?
I’ve spent the last 2 years in stealth building a new way to train models (RLCD), and a new type of frontier AI model that we are releasing today: Jev
• 20-200x faster
• 40-400x cheaper (w/ output tokens free)
• Frontier composable intelligence optimized for decisions
AFAICT the shortest path to AI-based economic revolution
2
3
599
this is so much fun, remember Crimsonland?
lets take Zig version from @banteg Crimsonland rebuild and put it on an iPad mini 1st gen from 2012 !
A5 / 512 MB RAM / iOS 9.3.5
~90k lines of Zig, not rewritten, ~100 lines of iOS patches
Zig -> C -> Xcode clang for 32-bit ARM
custom raylib iOS backend, OpenGL ES 2
touch sticks
sound works
60 fps, ~60 MB RAM
don't throw away your old apple devices!
this is an iPad mini 1st gen from 2012!
it can't run anything, modern apps dropped iOS 9 long ago.
so what do you do with this hardware + Claude... heh
once you get ssh access, you can build any app you want.
mine is now a native Home Assistant wall panel.
(HA app doesn't support iOS 9, and the web dashboard breaks in old Safari)
and the best part:
you can grab one on ebay for ~$30-50,
same price as a popular esp32 IoT board :)
4
2
19
6,737
don't throw away your old apple devices!
this is an iPad mini 1st gen from 2012!
it can't run anything, modern apps dropped iOS 9 long ago.
so what do you do with this hardware + Claude... heh
once you get ssh access, you can build any app you want.
mine is now a native Home Assistant wall panel.
(HA app doesn't support iOS 9, and the web dashboard breaks in old Safari)
and the best part:
you can grab one on ebay for ~$30-50,
same price as a popular esp32 IoT board :)
82
315
4,556
734,231
fly brain results are cool but how did it learn?
I tried to show learning episodes in this simple game:
a fly holds a joystick and tilts a platform to roll a ball into the hole.
like those old tabletop labyrinth toys.
166,700 neurons feeding a trained joystick controller.
levels are:
- straight path
- corner
- switchbacks
- full labyrinth.
goal: +5
wall hit: -0.1
checkpoint: +0.15
small penalties for taking too long and jerky commands.
you can watch recorded training attempts, rewards and neural activity together.
hits a wall, overshoots, tries again.
went from rolling a ball in a straight line to finishing the whole labyrinth. rly cool to see how it got there !
everyone is training fly brains to do different things
i wanted to look inside before starting my own experiments
so here is FLY ATLAS
139,662 neurons in 3D from the MaleCNS dataset
click a neuron to see its strongest incoming connections
switch regions on/off, zoom in, look around.
dots are actual cell-body positions
lines show connections
its incredibly cool seeing how much fits in a fly
link 👇
2
1
33
2,723
What were you doing yesterday?
I'm teaching Fly to play ping-pong.
How crazy is that? lol
I trained a fly connectome model and made it play ping pong against itself !!
Incredible, imagine this scene in your kitchen, lol!
91 returns in 100 seconds!
each fly runs the MaleCNS graph:
166,700 neurons
25.6 million connections
same trained weights, separate neural states
they control their flight height and swing a racket attached to a front leg.
you can watch both brains + the nerve cord light up as they play. the flashes come from the model's activity.
1
20
1,988
I trained a fly connectome model and made it play ping pong against itself !!
Incredible, imagine this scene in your kitchen, lol!
91 returns in 100 seconds!
each fly runs the MaleCNS graph:
166,700 neurons
25.6 million connections
same trained weights, separate neural states
they control their flight height and swing a racket attached to a front leg.
you can watch both brains + the nerve cord light up as they play. the flashes come from the model's activity.
everyone is training fly brains to do different things
i wanted to look inside before starting my own experiments
so here is FLY ATLAS
139,662 neurons in 3D from the MaleCNS dataset
click a neuron to see its strongest incoming connections
switch regions on/off, zoom in, look around.
dots are actual cell-body positions
lines show connections
its incredibly cool seeing how much fits in a fly
link 👇
7
11
132
15,310
everyone is training fly brains to do different things
i wanted to look inside before starting my own experiments
so here is FLY ATLAS
139,662 neurons in 3D from the MaleCNS dataset
click a neuron to see its strongest incoming connections
switch regions on/off, zoom in, look around.
dots are actual cell-body positions
lines show connections
its incredibly cool seeing how much fits in a fly
link 👇
2
9
116
19,528
