Nacho Mellado @uavster Twitter profile

Pinned Tweet

Nacho Mellado

@uavster

4 months

Mapping speech to actions in real time. Everything is processed on the robot. Sound on!

58

170

1K

Last Seen Profiles

@from_sakai2022

@wcmk_

@winnyokstate

@qwertyuiazxcvbn

@rehema_kiteto

@discoverLA

@osushivip

@GinaLazola71165

@bokeplokalmalam

@El_Hormiguero

@forest_9119

@evligiz60125833

@vegetable6__6

@DrNotavel

@ciaooow

@wolfkingvillian

@sweetweety613

@Andre_Firstlove

@shmh_bh

@uri_noona

@thaboarr

@GunkelNic

@PayPayP78026068

@Tahseen9001

@acdattt

@JeseniaWhyte

@angular

@kagurabachi_x

@bokeplokalmalam

@takahashiS35349

@1Taif17

@bokeplokalmalam

@ahmedmona111111

@haley_huard

@NASAJPL

@softiethefilm

Nacho Mellado

@uavster

3 years

@roydanroy @jmsl

0

54

971

Nacho Mellado

@uavster

4 years

@neilsardesai @nheagy Let me check mine...

3

1

548

Nacho Mellado

@uavster

8 months

Not to steal @karpathy ’s spotlight (more like me tripping on a wandering photon from the sun, in this case), but I was going to post that same post. Friday is my official last day. I’m leaving to focus on personal projects that I managed to fund 🙏

Andrej Karpathy

@karpathy

8 months

Hi everyone yes, I left OpenAI yesterday. First of all nothing "happened" and it’s not a result of any particular event, issue or drama (but please keep the conspiracy theories coming as they are highly entertaining :)). Actually, being at OpenAI over the last ~year has been

2K

1K

23K

10

228

Nacho Mellado

@uavster

1 year

Folks, I have been busy building my own robot, and I finally have a first iteration to show. I'll use it to experiment with perception algorithms for home applications (stay tuned for posts). Without further ado, I give you... *drumroll*... HF1!

11

132

Nacho Mellado

@uavster

1 year

@GaryMarcus Beware of “model collapse”

Nicolas Papernot

@NicolasPapernot

1 year

In work led by @iliaishacked we ask what happens as we train new generative models on data that is in part generated by previous models. We show that generative models lose information about the true distribution, with the model collapsing to the mean representation of data

13

143

533

4

10

129

Nacho Mellado

@uavster

3 years

@ValaAfshar @YvesMulkers

4

118

Nacho Mellado

@uavster

3 years

@Grady_Booch We should rephrase terms like “sprint”. Nobody has won a marathon by sprinting all the time.

6

24

93

Nacho Mellado

@uavster

3 years

How it started; how it’s going

2

4

75

Nacho Mellado

@uavster

5 months

@Its_Garix @MKBHD -People pay for solutions, not tech: "AI" is a tech term, they talked about their product -They want to stay different: "AI" has been overused, from A/C to toasters -They keep their distance from the hype: they don't want others to influence their customers' expectations

1

4

71

Nacho Mellado

@uavster

6 years

@lirondos Driven by some kind of accelerated evolution due to environmental pressure (a.k.a. “chancla”), Homines sapientes are developing abilities to avoid “lo fregao”.

1

19

66

Nacho Mellado

@uavster

4 years

1

20

63

Nacho Mellado

@uavster

1 year

Testing gestures and expression. Yea or nay? #Robotics

6

12

63

Nacho Mellado

@uavster

4 years

@Burgunbeerd @neilsardesai @nheagy What? I want my $25M back!

0

54

Nacho Mellado

@uavster

1 year

@pikuma The infamous chicken lock

1

0

55

Nacho Mellado

@uavster

10 months

Happy 2024! The 486 DX4 turns 30 years old in March. It's the CPU that got me into real-time graphics and the #demoscene . That’s why I target it in almost all my #retroprogramming projects. To celebrate, I've coded a blur effect in true color optimized for the DX4! Code in 🧵

3

8

50

Nacho Mellado

@uavster

9 months

Imagine 1994 pixelated games with cinematic effects. I coded a real-time shallow depth of field effect on a 30-year-old computer. 🧵 #retrogames #dosbox #486versary #demoscene

1

6

49

Nacho Mellado

@uavster

10 months

The Everyday Robots team (and yours truly) worked very hard to make these robots wipe tables, sort trash, and other cool stuff. It’s great to see the robots are still in the best hands. Go, @GoogleDeepMind ! 🦾

Sergey Levine

@svlevine

10 months

These robots are coordinated by an LLM that decides what tasks to try, where to try it, how to obey a "constitution" with some "Asimov-inspired" laws. Check out more about AutoRT here: And in the blog post: Here is a thread 👇

14

173

766

2

6

44

Nacho Mellado

@uavster

2 years

@GaryMarcus 🧐

3

2

39

Nacho Mellado

@uavster

1 year

Everyday Robots was a vibrant project that taught me so much. Such a cool team of humans and robots! #NewProfiePic

0

3

32

Nacho Mellado

@uavster

5 months

@ENDESGA Your father and I would get along well.

Nacho Mellado

@uavster

9 months

Imagine 1994 pixelated games with cinematic effects. I coded a real-time shallow depth of field effect on a 30-year-old computer. 🧵 #retrogames #dosbox #486versary #demoscene

1

6

49

0

30

Nacho Mellado

@uavster

3 years

One particularly unsettling piece of @lexfridman ’s interview with Avi Loeb: people live every day as if they were going to live forever.

4

29

Nacho Mellado

@uavster

2 years

Our fellow Americans, from the bottom of our hearts, thank you for welcoming our family.

2

1

32

Nacho Mellado

@uavster

4 months

I spent a couple of weeks optimizing the latency of transformer models on an NVIDIA Jetson Nano. Hopefully, I'll be able to show what they do soon. Here are few things I learned:

2

24

Nacho Mellado

@uavster

4 months

@JeffCaldwell_OG Thank you! I have plans to open source. Stay tuned 😊

1

0

26

Nacho Mellado

@uavster

1 year

Yesterday was my last day at Google after a 4-year wild ride of learning, growing and making general-purpose robots. I am so grateful. Will miss working with such a stellar team of roboticists (and friends), but also can’t help being thrilled by what’s coming next!

0

24

Nacho Mellado

@uavster

1 year

@Grady_Booch

0

2

23

Nacho Mellado

@uavster

2 years

@EricDraven1989 @NxtFukingLevel Correct. I’m being a tad bitter, but the target market seems educated first-time parents. It predates on inexperience and the natural inclination to believe that our babies will be special.

1

0

21

Nacho Mellado

@uavster

10 years

@microsiervos Muy interesante. Aquí explico cómo hacerlo con la cámara del móvil, código fuente mediante: http://t.co/sZi2ihHCsy

0

7

22

Nacho Mellado

@uavster

1 year

@awesomekling Hey, I managed to sit at my computer 24/7 and turn bald without creating an entire OS 😅 Sometimes one just needs to build something beautiful. Not everyone will understand, but those of us who do get inspired to build too and continue the cycle. Keep going!

0

22

Nacho Mellado

@uavster

3 months

Terrible advice. Please don’t do this in production code without learning C first and thoroughly understanding and checking the generated code.

3

1

20

Nacho Mellado

@uavster

4 months

New capability alert! HF1 can now transcribe speech in real time.

2

22

Nacho Mellado

@uavster

1 year

I still have to make some chroma adjustments, but here's HF1 streaming what it sees in stereo to my laptop! Under the hood, a simple encoder-decoder GStreamer pipeline with frame composition of the two stereo frames.

2

3

20

Nacho Mellado

@uavster

8 months

I’ll be posting more updates as I figure things out. First thing is turning this space into a lab.

3

0

19

Nacho Mellado

@uavster

3 years

@SchmidhuberAI - Gradient descent you say?

0

1

19

Nacho Mellado

@uavster

2 years

@ylecun What prevents anyone with access to the model from arming antivaxxers et al with a plethora of “scientific” papers? I’m sure this was well-intentioned, but what the world desperately needs is automated pseudo-science discrimination, not generation.

2

0

17

Nacho Mellado

@uavster

6 years

Billy Corgan killing it with the @SmashingPumpkin last Monday in Oakland.

1

3

16

Nacho Mellado

@uavster

1 year

It’s done. We gave back our house, sold our car, got rid of most of our possessions, and fit the rest in a rental car. We’ll drive north for twelve days; then, move back to Spain. We already miss our friends, and look forward to seeing our familia y amigos.

1

0

18

Nacho Mellado

@uavster

1 year

Packing list: clothes, toothbrush, deodorant, calibration pattern…

0

18

Nacho Mellado

@uavster

4 months

@Ysqander Thanks! It has an Arduino-based board for low-level control. ML models run on a Jetson Nano from NVIDIA.

1

17

Nacho Mellado

@uavster

2 years

@nearcyan False. Ann Wilson seemed pretty stoked wit it.

0

16

Nacho Mellado

@uavster

3 years

29% of CPU on a 486 DX4-100. Got to optimize this Scream Tracker module I made in 1998 to leave as many cycles as possible for real-time visuals. Could set playback to 22kHz; possibly use fewer channels by premixing sample pairs. Twitter, please treat audio well this time.

2

16

Nacho Mellado

@uavster

10 months

If you see this, post a robot.

Gammitin (Ben) 💾

@Gammitin

10 months

If you see this, post a robot.

11

2

46

2

3

16

Nacho Mellado

@uavster

1 year

@ylecun I’d like to try it on a Jetson Nano. Here’s my robot showing the kind of interaction I’m looking for:

Nacho Mellado

@uavster

1 year

Testing gestures and expression. Yea or nay? #Robotics

6

12

63

1

14

Nacho Mellado

@uavster

2 years

@ashleymcnamara @TamsinLM It’s an improvised drum kit. The driver likes to play when stuck in traffic. Each can is filled with a different amount of clams, which gives it a unique pitch when hit with a chopstick.

0

15

Nacho Mellado

@uavster

2 months

Apple openly looking for roboticists. Guess it’s happening?

6

0

14

Nacho Mellado

@uavster

1 year

I can now calibrate my robot's cameras wirelessly. No need to connect a monitor to the Jetson. Added camera streaming to NVIDIA's capture program with GStreamer's python API. Just because I couldn't take my monitors moving out of the US. Who would think constraints can be good?

3

14

Nacho Mellado

@uavster

2 years

@joseph_h_garvin @ylecun Our ability to spot fraud is limited by intelligence, expertise, biases, attention, time, and other factors out of our control. Making bad actors more efficient in disguising misinformation is worse for everybody, not just conspiracy theorists.

0

11

Nacho Mellado

@uavster

9 months

The effect above runs at ~11 fps at 320x200 32 bpp on an emulated 486 DX4 100 Mhz on DOSBox. I know, it's borderline real time for today's standards, but it would have been considered a decent frame rate back then. Making things run in true color was very hard at the time.

1

0

13

Nacho Mellado

@uavster

2 years

Homegrown

0

1

5

Nacho Mellado

@uavster

3 years

@lukeisamazing Please show some respect. Not only was he vital for the success of the Apollo mission, but he became the drummer and singer of Genesis, and left us anthems like "In the Air Tonight" and "Against All Odds" in his solo career.

0

12

Nacho Mellado

@uavster

8 months

Working @PickNikRobotics has been great. If I could go back in time, I would totally encourage my younger self to take that path. Supercharging perception in MoveIt Pro was fun! Consider joining their incredible team if you’re looking for a new gig.

MoveIt Pro Runtime & Developer Platform

Rapid development for commercial-grade, unstructured robotic arm applications.

picknik.ai

1

0

13

Nacho Mellado

@uavster

7 months

I hacked a 28-year-old C++ compiler to support exceptions! That’s two years before exceptions were even part of the first C++ standard. 🧵 #retrocomputing #retrocoding #DOS

2

1

12

Nacho Mellado

@uavster

2 months

It’s not a bug, it’s a feature.

1

0

12

Nacho Mellado

@uavster

2 years

@TheRetroRoomRoo

1

0

11

Nacho Mellado

@uavster

3 months

Dear Community, You’re getting your hands on this robot project, one way or another. I’m exploring funding and ways to deliver, while developing more features and teaching materials. It will take time, but we will learn together. This is our robot!

1

11

Nacho Mellado

@uavster

2 months

Solved the unintentional backward wheelie. The speed controller now applies a smooth ramp to the target speed. This also prevents wheels from slipping, which improves localization and trajectory accuracy.

2

0

11

Nacho Mellado

@uavster

9 months

Dear startups, you’re not Google. Your problems don’t have the same scale yet. That extra complexity will cost you $$ and cognitive space that won’t be used to develop the solutions your customers will pay you for.

Vlad Mihalcea

@vlad_mihalcea

9 months

Software development is the art of replicating the atypical architecture choices made by Google, Facebook, or Netflix into trivial projects with modest requirements.

5

118

692

1

2

11

Nacho Mellado

@uavster

3 years

Just baked a #chiptune to test the new panning support in my DOS audio library. What kind of game would it go well with?

1

2

10

Nacho Mellado

@uavster

3 months

Distributed computing is a huge opportunity for home robots. While power budget may be limited on the robot itself, latency-tolerant high-level intelligence can run on other home network nodes: charging station, laptops, desktops… private data doesn’t need to go anywhere else.

Alex Cheema - e/acc

@ac_crypto

3 months

2 MacBooks is all you need. Llama 3.1 405B running distributed across 2 MacBooks using @exolabs_ home AI cluster

210

728

5K

0

10

Nacho Mellado

@uavster

3 years

@estoyausente @ShionKaz Todos: es la C. Yo: no, es F!

0

10

Nacho Mellado

@uavster

1 year

It’s my first day at @PickNikRobotics . PickNik collaborates with NASA on space robots, contributes to ROS with a world-class robotic manipulation platform, and creates products that dozens of companies rely on to program their robots and create value. This is going to be fun!

0

10

Nacho Mellado

@uavster

6 years

Recovered my very first computer. Helped my 4-year-old write his first program: print “no”. #MSX #RetroComputing #dadlife

3

0

10

Nacho Mellado

@uavster

10 months

Robotics is VERY hard, and this is the right mindset. New Year’s resolution: “talk to self like this”.

Sequoia Alexander 🏳️‍⚧️

@TLAlexander

10 months

@scottmonaghan @JRBowling @madgrizzle @RoboFoundry yeah happens about once per day I am like "making robots like this has been my dream since I was 11 years old and now I am just out here really doin it"

0

5

1

2

10

Nacho Mellado

@uavster

2 years

Large language models, grounded on robot capabilities, help robots solve complex tasks expressed in natural language. @mkalakrishnan explains how Google Research and our team collaborated to achieve that. Robotics and AI are, more than ever, a team sport.

1

0

10

Nacho Mellado

@uavster

1 month

Impromptu user study with the robot today. Son's friends came to the house and asked to see it. They ended up playing with the robot for 1 hour and still talking about it!🤯 Took notes of what's working/what's missing. Finished product will still take time, but I'm thrilled.

1

0

10

Nacho Mellado

@uavster

2 months

These are two independent speed controllers (one for each wheel) keeping both average speed of 0.4 m/s across 2.5 meters. The robot oscillates a bit, but the trajectory is fairly straight. I think I’ll try to use it as a building block in the trajectory controller.

1

0

10

Nacho Mellado

@uavster

3 years

@docmilanfar I love it. Also, couldn’t help it 🙂

0

10

Nacho Mellado

@uavster

2 years

My 8yo and I had a lot of fun with our first game programming hackathon in Scratch. He learned the basics of control flow, variables, events and concurrency. And, if you ask me, I think the result is... eggcelent!

1

0

10

Nacho Mellado

@uavster

3 years

Hello, world! We’re building a new type of robot that can learn by itself, to help anyone with (almost) anything.

0

9

Nacho Mellado

@uavster

2 years

@gayle_writing @lucygrxce_15 @UMFL2002 @hannahmsays I don’t hate you at all, dear stranger. For us, it was 3.5… …years. And they still wake up sporadically so, if you write a book, I’ll gladly purchase a copy.

2

0

9

Nacho Mellado

@uavster

2 years

@fchollet Some things may be easier to learn later in life. I think learning potential is multifactorial. Motivation, focus and mental stamina can develop later in life, or benefit from incentives that intensify later. I feel more capable to learn some things now than in my twenties.

0

9

Nacho Mellado

@uavster

3 years

@TamsinLM Product life cycles were much longer back then. Also, many of us were kids with tight budgets. I had to work for a year or so to save the money to replace my 286 with a 486. Pentiums were simply out of reach.

0

8

Nacho Mellado

@uavster

2 years

Procedural mesh generation like it's 1996 🧵

2

1

9

Nacho Mellado

@uavster

6 months

@hive_echo @mattshumer_ Here’s an alternative explanation to emergent abilities worth taking into account:

Are Emergent Abilities of Large Language Models a Mirage?

Recent work claims that large language models display emergent abilities, abilities not present in smaller-scale models that are present in larger-scale models. What makes emergent abilities...

arxiv.org

2

8

Nacho Mellado

@uavster

3 months

That’s what I’m here for! ❤️

Clinton M

@CooKeMonter

3 months

Inspired by Nacho 👇 finally built my first prototype, not as sophisticated, an old raspberry pi, with a motor controller on a cheap RC car. But I got it running and do things. Still lots of work to do before I finish, GPS, mics (for voice control), internal power etc

2

6

1

0

9

Nacho Mellado

@uavster

2 months

founder mode

0

1

9

Nacho Mellado

@uavster

6 years

Yesterday was my last day at Apple. Working with the greatest minds in the industry has been a real pleasure. What an exciting 4-year ride!

0

9

Nacho Mellado

@uavster

4 years

My wife asked me to call her to find her phone. Me: Hey Siri, call my wife. Siri: Which wife? *shows three names from my contacts* Wife: 🤨

1

0

9

Nacho Mellado

@uavster

3 years

@nitrillo and I coded a 3D engine + 3DS loader back in the 90s, but never attempted procedural generation, AFAICR. Well, I'm fixing some of the code to do just that! Here's a first test on a 486. #oldschool #demoscene

1

3

9

Nacho Mellado

@uavster

5 years

This is what I’ve been up to lately. Our stellar team of humans and robots is on a mission.

The Team at X

@Theteamatx

5 years

Introducing the Everyday Robot Project: teaching robots how to learn so that one day they’ll be able to help us in everyday life.

1

21

44

0

1

9

Nacho Mellado

@uavster

2 months

I just implemented a tracking controller by Kanayama et al (paper in 🧵).

1

0

10

Nacho Mellado

@uavster

3 years

@thespite I don’t know what I’m doing, and I’m sequencing this with dang Clips on my phone, but here’s a concept 😅

0

8

Nacho Mellado

@uavster

3 years

@danidonovan When you laugh and no one made a joke

1

0

7

Nacho Mellado

@uavster

10 months

The idea crossed my mind weeks ago, and I couldn't help but try. Code's here!

1

0

7

Nacho Mellado

@uavster

2 months

I am making my robot follow arbitrary trajectories with a new controller. Its feedforward path takes waypoints and outputs anticipated tangential and angular velocities, while its feedback loop corrects errors. Here, a simulated red robot tracks a trajectory, feedforward only.

2

0

8

Nacho Mellado

@uavster

2 months

@hardmaru @SchmidhuberAI @MLStreetTalk

0

9

Nacho Mellado

@uavster

3 years

@tunguz Most def! SOTA is running a CNN on a screenshot of the spreadsheet.

0

8

Nacho Mellado

@uavster

8 months

Wife and I successfully moved in to our 12th home together. More homes than years we’ve been married, and growing exponentially. Every human should expect us living at their place sometime in the next 97 years.

0

8

Nacho Mellado

@uavster

2 years

@castillo__io Maybe complement/augment, but not fully replace. Where will it learn from as new languages and problems arise? LLMs are not free from distribution shift.

1

0

8

Nacho Mellado

@uavster

4 months

Here’s what one of those transformers does.

Nacho Mellado

@uavster

4 months

New capability alert! HF1 can now transcribe speech in real time.

2

22

0

8

Nacho Mellado

@uavster

2 years

Going for my daily run to reflect on and appreciate the important things in life. Good day to do it.

0

8

Nacho Mellado

@uavster

2 years

@GaryMarcus

0

1

8

Nacho Mellado

@uavster

9 months

But how does it blur/deblur so much in real time? Images are convolved with a kernel of up to 17x17 pixels. That's 289 multiplications/additions per pixel and channel, which would totally obliterate such an old CPU! I found a way.

1

0

8

Nacho Mellado

@uavster

3 years

DOS audio library by yours truly playing the sound track of the most influential PC demo of all time. Just a few effects missing for full S3M support! #demoscene #retrocomputing #chiptune

1

3

8

Nacho Mellado

@uavster

5 months

@BackTo_1970 @flowidealism @slatestarcodex I’m an engineer, my wife is a literature teacher. She beats me big time at helping my 4th grader understand math. Knowledge is a just fraction of being a good educator.

2

0

7

Nacho Mellado

@uavster

3 months

Imagine what would be like to depend this much on systems than can’t be introspected to fix a problem like this quickly. #AI

WIRED

@WIRED

3 months

NEWS: A software update from cybersecurity company Crowdstrike appears to have inadvertently disrupted Microsoft IT systems globally.

38

296

505

0

1

8

Nacho Mellado

@uavster

4 years

@sethrosen I support that, but ML Bootcamps have to be called Jax In The Box.

0

8

Nacho Mellado

@uavster

4 months

@Brymir7 I have not, but thanks for suggesting! My onboard compute and memory is pretty limited and I have to run other things in parallel so I try to keep the software stack thin. I believe that would thicken it? Here’s one transcription test. Made some improvements since.

Nacho Mellado

@uavster

4 months

New capability alert! HF1 can now transcribe speech in real time.

2

22

1

0

4

Nacho Mellado

@uavster

2 months

@chris_j_paxton @DanraeP An affordable companion robot for the home that does not share your data.

Nacho Mellado

@uavster

4 months

Mapping speech to actions in real time. Everything is processed on the robot. Sound on!

58

170

1K

2

7