flx

AI tools, engineer,crypto plays & degen experiments

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

this stanford professor co-founded an $11 Billions robotics startup at just 34

Chelsea Finn teaches CS224R, Stanford's deep reinforcement learning course, and co-founded Physical Intelligence

the course covers:

→ imitation learning
→ policy gradients
→ Q-learning
→ offline reinforcement learning
→ teaching robots new skills from demonstrations

the key idea for robot learning is how to teach a robot what to do from demonstrations — and why even small mistakes can add up until the robot fails

start with this lecture: RL for Robots

key timestamps:

04:54 - why robots still need human help
16:54 - why long robot tasks are hard
22:19 - forward-backward reinforcement learning
28:06 - teaching robots to reset themselves
42:36 - learning to repeat tasks on their own
52:34 - learning and adapting without a reset

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

https://t.co/wlXiRyfTce

[flx @flx2255]

50 robot demos is the tutorial

50 good demos is a dataset

LeRobot literally tells beginners to record ~50 demonstrations

so i would use the 50 to answer 5 questions

not just to hit the quota

1. HOW CONSISTENT IS THE OPERATOR?

record the same task 10 times

look at:

grasp position
approach angle
speed
number of corrections
trajectory length

if every demonstration solves the task differently

you aren't collecting one behavior

you're collecting 10 behaviors

2. HOW MUCH OF THE WORLD DOES THE DATA COVER?

don't put the cube in the same place 50 times

change:

starting position
object orientation
distance from the robot
background
lighting

otherwise the policy can memorize the scene instead of learning the task

this is one of the biggest problems in imitation learning

the model learns from the states you showed it

then makes a small mistake

that mistake puts it into a state it never saw

then the next prediction is worse

then worse again

one small error compounds

3. LABEL THE FAILURES

don't delete every bad demo

split them into:

success
failure
suboptimal
recovery

a failed grasp followed by a clean recovery can contain useful information

a collision followed by someone manually resetting the robot probably doesn't

the difference matters

4. DON'T RANDOMLY SPLIT THE DATA

this is a sneaky one

if you randomly split frames from the same trajectory

your train and test sets can contain almost identical states

you think the policy generalized

when it actually saw almost the same scene during training

split by episode

better:

train → different episodes

test → completely held-out starts

5. TEST THE THING YOU DIDN'T TRAIN ON

this is where the 50 demos finally become useful

train on:

cube at positions A–E

then test on:

position F

train with normal lighting

test with different lighting

train with one object orientation

test with another

measure:

success rate
failure mode
recovery rate
number of retries

because this is the question that matters

not:

"did my robot learn the task?"

but:

"what happens when the world changes?"

research on imitation learning makes the same point from a different angle

data quality isn't just about having more examples

it's about whether the dataset keeps the learned policy close to states it can actually handle :chatgpt-content-reference{index="2"}

that's why i'd rather see:

50 demos
→ 37 clean
→ 8 suboptimal
→ 5 failures

with every failure explained

than:

500 demos
→ 500 recordings
→ no idea why the robot fails

the dataset is not the 50 videos

the dataset is the decisions you make about them

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

3 sensors → 9 environmental signals → 93.1% validation accuracy

and the LLM never sees the raw sensor stream

this is how Atmosis turns an Arduino into an environmental intelligence system

the hardware is surprisingly simple

→ Arduino UNO Q
→ Waveshare X6 sensor
→ Waveshare dust sensor
→ DPS310 pressure sensor
→ WS2812B RGB strip

but the UNO Q is doing something interesting

it has 2 completely different processors

the STM32U585 handles the real-time layer

the Qualcomm QRB2210 runs Linux and handles the heavy stuff

so the architecture is basically:

sensors
→ STM32
→ 1-second data frames
→ Linux
→ Edge AI
→ Gemini
→ dashboard / Telegram

the STM32 continuously reads 9 parameters:

IAQ
TVOC
HCHO
CO
temperature
humidity
PM2.5
pressure
altitude

and it doesn't just dump numbers into an AI

the firmware first calculates an air-quality score

each pollutant gets a sub-score from 100 → 0

the final score is the lowest sub-score

so one dangerous pollutant can't get hidden behind 8 good readings

the board also handles the safety layer locally

PM2.5 alarm → 150 µg/m³
CO alarm → 9 ppm

and those alarms keep working even when Linux or the internet goes down

the data is sent from the microcontroller once per second

every 30 seconds the board also sends a health report

the ML part is where it gets really interesting

the model doesn't look at one sensor

it looks at all 9 simultaneously

each training window:

3 seconds
5 Hz
1 second stride

the current dataset produced 141 training windows

those became 135 input features

then a small neural network:

135 inputs
→ 20 neurons
→ 10 neurons
→ 3 output classes

30 training cycles
0.0005 learning rate

result:

93.1% validation accuracy

the classes are:

clean air
indoor pollution
poor ventilation

and the model runs locally on the UNO Q as a quantized int8 C++ library

then comes Gemini

but this part is deliberately separated from the ML model

Gemini doesn't decide whether the air is dangerous

it gets the already-processed environmental state

and turns it into something a human can understand

for example:

"PM2.5 is rising while ventilation is poor"

→ open a window
→ run the purifier
→ avoid the room until conditions improve

and it can send the result directly to Telegram

there's even a local web dashboard showing:

live readings
air score
detected patterns
trends
system status
AI recommendations

the entire project is documented as 36 build steps

from designing the enclosure in Fusion 360

to 3D printing

to sensor assembly

to data collection

to training

to deploying the model

to running the dashboard

what i like most is the architecture

Sense
→ Understand
→ Interpret
→ Respond

the microcontroller handles what must be reliable

the Linux processor handles what needs compute

the ML model recognizes patterns

the LLM explains them

instead of asking:

"what is the PM2.5 value?"

the system asks:

"what is changing
what does the pattern mean
and does something need to happen?"

that's a much more interesting way to build an AI device

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

i put this together in 30 minutes and have already made $10k on Polymarket in the weather market

this started as a tiny $ESP32 satellite globe

and i realized it could become a weather signal for Polymarket

the hardware is ridiculously simple

→ Adafruit QT Py ESP32-S2
→ 240×240 ST7789 TFT
→ 128×64 OLED
→ microSD
→ 1 button

that's it

the ESP32 checks the NSMC satellite availability API every 15 minutes

when a new frame appears:

→ downloads the global infrared image
→ saves the original PNG to microSD
→ decodes it locally
→ maps it onto a rotating 3D Earth
→ calculates the real day/night terminator
→ stores the frame in the archive

that's:

96 frames per day
672 frames per week
2,880 frames in 30 days

and the archive is actually replayable

1 click → 24 hours
2 clicks → 7 days
3 clicks → 30 days

so this isn't just a pretty weather display

it's a tiny local satellite database

the same ESP32 also tracks the ISS

it downloads the ISS TLE from CelesTrak once every 24 hours

then AioP13 propagates the orbit locally

the globe shows:

60 minutes before the ISS
+
60 minutes after it

and the OLED can show:

azimuth
elevation
pass predictions
visibility

even the image decoder has an interesting optimization

PNGdec's default buffer can corrupt some satellite images

so the project raises the allocation to 2,624 buffered pixels

large retained buffers are pushed into PSRAM

so the ESP32 still has enough internal memory for Wi-Fi and TLS

all of this is running on a tiny embedded board

and here's where it gets interesting

take those 2,880 satellite frames

add:

cloud coverage
cloud movement
cloud-top temperature
historical observations
weather forecasts

now turn the data into probabilities

21°C → 18%
22°C → 51%
23°C → 31%

then compare that against prediction-market prices

for example, Polymarket's London temperature market had:

22°C → 48%
23°C → 26%
21°C → 25%

with $16.8K+ traded

and the market resolves using the highest NOAA temperature reading from London City Airport

so the satellite isn't the oracle

it's another information source

the entire pipeline becomes:

satellite
→ ESP32
→ image archive
→ feature extraction
→ weather model
→ probability
→ market price

30 minutes of hardware

15 minutes per new satellite frame

30 days of local history

and potentially $10K on the other side of the signal

this is the kind of embedded project i'd actually build

not another LED blinking demo

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

8 robotics accounts that show what actually broke, what they changed, and what finally worked

some build humanoids
some train robot models
some work on controls
some build the hardware itself

here are the accounts i would keep on my feed:

1: JIM FAN - @DrJimFan

NVIDIA robotics researcher

posts actual robot experiments, papers and demos

recently showed a robot model running with 8,000 timesteps of context and work on robots improving themselves

2: CHELSEA FINN - @chelseabfinn

co-founder at Physical Intelligence

one of the best people to follow for robot learning, RL and general-purpose robots

her recent talk covered the reliability problem, learning from failures and robots working autonomously for hours

3: SERGEY LEVINE - @svlevine

one of the key names in modern robot learning

posts about imitation learning, RL and why methods like action chunking actually work

4: PIETER ABBEEL - @pabbeel

robot learning + AI

great account for papers, new robotics systems and where research is heading

5: RUSS TEDRAKE - @russ_tedrake

MIT robotics

controls, locomotion, manipulation and the engineering side of robotics

6: KAROL HAUSMAN - @karolhausman

Physical Intelligence

robot learning, foundation models and general-purpose robotics

7: SHURAN SONG - @shuransong

robotics + manipulation

worth following for work on dexterity, robot learning and real hardware

8: HUY HA - @haqhuy

robotics researcher working on simulation, control and robot design

often posts the actual experiments, not just the paper title

the pattern i noticed:

the useful accounts don't just announce new models

they show:

what failed
what changed
what worked
and what they learned

that's exactly what you want in your feed if you're trying to become better at robotics

not more "10 courses you need to take"

more real robots
more experiments
more failures
more numbers

save this list before the algorithm buries it

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

THIS MIT ROBOTICS LECTURE IS 20 YEARS OLD

and some of the engineering lessons still sound like they were written for SpaceX today

MIT's 2005 lecture "EVA and Robotics on the Shuttle" breaks down how astronauts and robotic systems had to work together in one of the most complex machines ever built

interesting parts to watch:

[00:00] why EVA wasn't even part of the original Shuttle plan

[15:00] how the payload bay doors were designed with redundancy and manual EVA backup

[30:00] why engineers had to train for mechanical failures instead of assuming the system would always work

[45:00] the systems-engineering problem behind EVA and why changing one part of the spacecraft affects everything else

[60:00] how robotics changes what astronauts can actually do in space

the hardware is old

the engineering loop isn't

design
test
operate
fail
fix
repeat

that's exactly the mindset i used to build my 12-month roadmap for becoming job-ready in aerospace and robotics

MIT lecture ↓

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

I MADE A TIER LIST OF 13 COMPANIES YOU CAN JOIN IF YOU WANT TO BUILD ROCKETS, HUMANOID ROBOTS OR AUTONOMOUS SYSTEMS

not based on valuation

not based on hype

I ranked them by the engineering work, number of openings, scale and access to real hardware

S TIER

the companies where you can work directly on some of the hardest problems in robotics and aerospace

- SPACEX

$140K+ starting salary for many engineering roles

13,000+ employees

rockets, spacecraft, propulsion, GNC and manufacturing

you can work on Starship, Falcon, Dragon or Starlink

- TESLA

134,785 employees worldwide

Optimus roles list up to $330K base salary + stock and cash awards

you can work on mechanical systems, thermal design, simulation, controls and AI

- FIGURE

$1B+ raised

200+ employees

humanoid robots + AI

the goal is to build the intelligence that makes the robot useful

- 1X

87+ open roles

$100M+ raised in its latest funding rounds

AI, robot learning, robotic controls, hardware, manufacturing and robot operations

- APPTRONIK

$935M raised in its Series A

Apollo humanoid + robotics AI

the company is now moving from prototypes toward commercial deployment

- BOSTON DYNAMICS

30+ years of robotics development

Atlas + Spot

current openings include humanoid robotics software, Atlas data operations, simulation and mechanical engineering

A TIER

companies doing cutting-edge work, but with a more specific focus

- ROCKET LAB

8+ launches in 2025

$436M+ revenue in 2025

rockets, satellites, propulsion, software and launch systems

- PHYSICAL INTELLIGENCE

$400M raised in its Series A

robot learning and foundation models for physical AI

the goal is to make one AI system work across many different robots and tasks

- ANDURIL

$2.5B raised in its latest funding round

10,000+ employees

autonomous systems, aircraft, software, sensors and defense hardware

B TIER

smaller or more specialized teams where the work is still very hands-on

- AGILITY ROBOTICS

$400M+ raised

60,000 sq ft Physical AI facility in Fremont

Digit humanoid + large-scale robot deployment

- RELATIVITY SPACE

$1.3B+ raised

reusable rockets + large-scale additive manufacturing

- SKILD AI

62 open roles

$300M raised in its latest funding round

general-purpose robot intelligence across AI/ML, robotics, software, hardware, operations and data

C TIER

strong engineering teams, but less direct overlap with humanoids and robotics

- SIERRA SPACE

$1.7B+ raised

2,000+ employees

spacecraft, satellites and mission systems

one current embedded software role lists $156K–$215K base salary

and here's what surprised me

you don't need to start at the S tier

you don't even need to start as a research engineer

>mechanical engineering
>controls
>software
>manufacturing
>testing
>simulation
>data
>technician work

all of them can become an entry point

the goal isn't to collect certificates

it's to build something real enough that one of these companies can see what you can do

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

I FOUND 84 OPEN ENGINEERING JOBS THAT MATCH THE SKILLS IN THIS ROADMAP

12 months of work
3 serious projects
84 jobs at the end of it

so i worked backwards:

what do these companies actually ask for?

Python?
C++?
CAD?
controls?
robotics?
hands-on experience?

here's what stood out:

1: 3 projects beat 27 small ones

the goal isn't another 20 certificates

it's 3 serious engineering projects

build → test → break → fix → repeat

2: the projects need numbers

built a PID controller" means little

settling time: 0.42s" gives an interviewer something to ask about

what failed?
why?
what changed?

3: hands-on work matters

SpaceX says strong candidates often have significant contributions to hands-on extracurricular projects

robotics
rocketry
CubeSats
UAVs

embedded systems

own a subsystem and make it work.

4: the interview starts with the project

technical screens

project tests

programming tests

then questions about your technical achievements and failures

cheat-codes:

1. build 3 deep projects
2. measure everything
3. document failures
4. own one subsystem
5. make every project something you can defend

main insight:

you don't need 84 different skill sets

you need the engineering fundamentals that keep showing up:

build

test

break

fix

repeat

and one more thing

by month 12, you should have 3 projects you can defend — not 20 certificates you can't

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

THE MODELS KEEP GETTING BETTER.

Gemini 4 Argon hits:

91.9% on Vibe Code Bench
88.8% on LABBench
99.7% on GraphWalks

coding, science, math, long context — the models are getting very good at working with information.

but there's a whole category missing from this table:

real robots.

because robot learning has a different bottleneck.

you can have the best model in the world

but if the demonstrations are bad, the model learns bad behavior

someone still has to collect the data

check it
fix it
repeat the task

and measure what actually worked

and you don't need a $10,000 lab to start doing it

an SO-101 leader + follower setup costs about $230

that's enough to start collecting real robot data, testing policies and building a portfolio

the industry is already paying people to do this:

$25/hr at Physical Intelligence
up to ~$48/hr for some Tesla Optimus data collection roles

35 open roles at PI

only 3 have "research" in the title

the rest are engineers, technicians, operators, data, infra and ops

better models still need better data

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

How to get your first job in robot learning, without a PhD:

i pulled every open role at Physical Intelligence, one of the best-funded robot AI companies in the world

35 openings

only 3 have "research" in the title

the other 32: engineers, technicians, robot operators, data, infra, ops

so the door into robot learning isn't the research team. it's everything around it

here's my workflow for walking through it:

1: pick your door by what you already are

- good hands, no degree → robot operator. PI pays $25/hour + benefits and lists no degree. you teleoperate robot arms: folding clothes, sorting cups, opening jars
- mechanic or electrician → robotics technician. PI asks for an associate degree or "equivalent hands-on experience"
- software engineer → data. PI is hiring for data quality, data systems and robot interfaces right now
- student → internships. PI's mechatronics intern post says students below the usual level can still be considered with "exceptional projects, GitHubs, portfolios"

2: learn what the data is for

in Stanford's AA203 lecture 16, at 0:50, the lecturer says imitation learning is "capped in terms of performance" by the expert it copies

read that again

the person giving the demonstrations is the ceiling of the model

an operator who knows why a demo is bad is worth more than one who just hits the quota

that knowledge is free: AA203 lecture 15 and CS224R lecture 2

3: build a data portfolio, not a model portfolio

everyone shows a trained policy

almost nobody shows the data behind it

- record 50 demos of one task on an SO-101 arm with LeRobot
- write a rubric: what makes a demo bad — hesitation, collisions, a different grasp every time
- train one policy on clean demos, one on messy demos
- publish both success rates and the dataset on the Hugging Face hub

that one project speaks the language of the data team, the operator team and the research team at the same time

4: keep learning at night

PI's operator shifts run morning, evening and overnight

a paid shift inside a robot learning company + 2 hours a day of the free Stanford path

you learn the theory at night and see it fail on real robots by day

cheat-codes to stand out:

1. read the job post like a syllabus

PI's operator post lists the exact tasks: bagging groceries, sorting cups and plates, opening jars, folding clothes, installing light bulbs

pick one
do it on your own arm
put the video in your application

2. your $229.88 arm is the interview

PI's operators "lead robot movements with your arms (the robot mirrors your actions)"

that's exactly what an SO-101 leader + follower pair does

$229.88 in parts, official list

3. apply to the data team, not the research team

3 research openings vs data quality, data systems, robot interfaces, integration, deployments

do the math

4. Tesla pays for data too

Optimus data collection operators have been listed at up to about $48/hour in Palo Alto

they wear a motion-capture suit and a VR headset while collecting data for the robots

5. talk in numbers

success rate

success rate after one change in the scene

demos needed

that's the language robot learning teams actually use

main insight:

PI's own job post says it plainly: "Data collection is the fuel that drives our mission"

my read:

robot learning isn't only bottlenecked by people who can design models

it's also bottlenecked by people who can produce, check and fix the data those models learn from

and one more thing

if imitation learning is capped by the expert, then the best operators aren't doing the boring job

they're setting the ceiling for the whole model...

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

While all of Twitter is talking about Dior and Taco Tuesday - here’s a practical 12-month plan to build a SpaceX-level portfolio

Not certificates
Not “another course”
Build. Test. Break. Fix. Repeat.

A bookmark ≠ action

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

THIS GUY BUILT A ROBOT AND ENDED UP AT SPACEX

Joshua Mora Sanchez recently joined @SpaceX as a Test Reliability Engineer

Look at what he was building before getting there ↓

6-wheel rover
real hardware
robotics
testing
iteration

This is what a SpaceX-ready portfolio actually looks like

So I mapped out the path:
Python → C/C++ → electronics → embedded → controls → ROS 2 → autonomous robotics → AI

12 months
3 serious projects
No certificate collecting

I put the entire roadmap here ↓

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

ANTHROPIC LEAKED A FILE SHOWING HOW TO TRAIN YOURSELF FOR SPACEX-LEVEL ENGINEERING

12 months

Not 4 years of collecting certificates

Python → C++ → electronics → controls → ROS 2 → autonomous robots → computer vision → AI

The roadmap tells you exactly what to build:

$0 simulations
$20–30 multimeter
ESP32
closed-loop motor
2-axis robot
autonomous rover
robot learning project

And the final portfolio isn't:

“passionate engineer with 17 certificates”

It's:

3 real systems
test data
failure analysis
GitHub history
measurable results

Build. Test. Break. Fix. Repeat.

I turned the entire roadmap into a 60-step guide ↓

[flx @flx2255]

https://t.co/uTCYo3YJQj

[flx @flx2255]

Bitget-linked funds are moving through one central hub.

Bitget-linked wallets
0xe07Bd590E1198666230932BAd5db3DbBE21e7d57
Uniswap / Universal Router
Mayan Finance
Celer
new wallets

$8.3M BNB
$1.5M USDT
$770K USDT
$400K USDT
$200K+ USDT

this is what laundering can look like onchain:

aggregate the funds
swap assets
bridge across chains
split into fresh wallets

the goal is simple:
make one traceable flow look like hundreds of unrelated transactions.

[flx @flx2255]

bitget wallets just moved ~$170m in under an hour

not to another labeled exchange wallet
to one fresh address
then that address started splitting it

10k eth
another 10k eth
$19.7m usdt0
$12.9m usdc
3k xaut
12.7k bnb
hot wallet in
cold wallet in
new wallets out

the interesting part isn’t the headline
it’s the sequence

labeled bitget addresses feed a hub
the hub fans out to fresh wallets
stables start turning into eth

that’s usually the dump phase, not a routine reshuffle

bitget hasn’t confirmed anything
people are already saying withdrawals are stuck

if this is what it looks like, eth eats the first hit
not because the chain broke.

because stolen size and exchange fear both hit the same exit.

[flx @flx2255]

@XYZCryptos @variational_io the interesting part is that the first account being down $7K doesn't necessarily mean the strategy failed

if the second account farmed enough points to offset that loss, the real variable is the implied $ value of each point