AI, Robotics, Culture & sometimes Edamame 🫛 Onions expressed here are my own.
San Francisco, CA
Joined March 2021
- Tweets793
- Following1.1K
- Followers150
- Likes28.6K
Pinned Tweet
Couldn't help it
🤖 Made with AI
It's not often that you come out of a conference event questioning your world view.
Ted Xiao's talk on how AI can eventually automate loops for everything from component development to metallurgy was one of those.
@xiao_ted has incredible clarity on where we actually are today and what's needed to get to the vision of physical general intelligence is quite rare.
It's not often that you come out of a conference event questioning your world view.
Ted Xiao's talk on how AI can eventually automate loops for everything from component development to metallurgy was one of those.
@xiao_ted has incredible clarity on where we actually are today and what's needed to get to the vision of physical general intelligence is quite rare.
@TheHumanoidHub has done a fantastic job of communicating and normalizing the idea of ubiquitous physical general intelligence.
Excited to share some perspectives on important physical AI problems and challenges in a post-AGI world!
'Human Centric Design' and 'Human Machine interface' could attract the type of creatives which 'Machine Centric Design' and 'Task Machine Interface' never can.
More than 80% of all our neurons are dedicated to the cortical and subcortical motor systems.
Less than 5% are used by the language, math and logic centers of the brain.
The boring stuff like walking and climbing stairs is incredibly hard to replicate 'artificially'.
At a previous company, a systems engineer wrote down a Jacobian matrix on the white board to assess individual risk factors and the room went dead silent.
Mathematicians will do quite well in almost any engineering discipline
But we also need to talk about how engineers lack basic math skills
Heard @deepakpathak go into the details behind the S1 model and it's genuinely impressive. It seems like 'omni-bodied' generalized robot models will have an advantage because of the data advantage.
Honda standardizing their sensing ( adas) suite has has and will continue to save more lives & property damage in NA than all the autonomous vehicles combined. This will likely be true until 2035 or even 2040.
We should of course continue to expand in both directions but the hype is different from our ground reality
Was studying rigid body rotations and realized that we'll like never have named formulas, series and algorithms again.
There'll be no new Taylor series, Fourier Transform, Gaussian distribution, Laplace transform etc.
It'll likely all be called Chatgpt/Claude/Gemini research archives X -[Insert Date]
It makes me a little sad.
These names formulas inspired me when I was younger and I remember going down rabbits holes reading the mathematician's life story and circumstances.
Hope we can somehow preserve the acknowledgement( in future discoveries) to continue inspiring younger minds
Uber and Doordash should've been the first to create persistent consumer agents
I am still surprised that they're not devoting a big chunk of their resources to an agent offering
They have the user(normie) intent, attention and the workforce to close the loop in the physical world
For anyone working on personal persistent agents, here's a quick guide on what I and many of the normies would like to use it for.
I'll happily give it any task that cleanly falls in the pink quadrant. I'd also delegate tasks in the blue quadrant once I trust that it works as expected and doesn't make stupid mistakes. Mistakes cost time whoch is anyway short for urgent tasks
To gain this trust, the agent needs to learn to check with me in case it doesn't know one of my preferences.
I can't wrap my head around the two use cases chosen for this video.
Booking a flight and hotel takes a couple of minutes and I can do it directly on the google search page with the map view and images all easily accessible.
For building a workflow around user feedback, I'd much rather start in codex and share. Starting a voice call with an agent adds unnecessary friction and doesn't really free up my attention.
Meta at least chose to demo tasks that I would truly want to delegate( like filling out a lawsuit claim form or negotiating a lower bill)
SF dialing up the cyberpunk vib.
Looking forward to seeing Pathfinder 3 in action next @ltares
Doordash and Uber have a real shot at building the most successful consumer agent, something that blends the digital and physical world
Imagine
"Dash, can you return the earphones I bought online, I'll leave it in my mailbox"
And it find the return label, hails a dasher who comes to pick ot up and drop ot off at the UPS
Can they execute on this though?
$DASH $UBER
You can now just text DoorDash and let it order for you. It’s proactive, learns from your repeat orders, and helps you find the best deals and new spots.
Just tell it what you want without any extra setup. Ask it to reorder your usual, or add it to a group text to collect everyone’s orders. You can even send a photo of your fridge and it’ll tell you what ingredients you’re missing for a recipe you want to make and deliver the ingredients to your doorstep.
We believe agents are a new app paradigm and are building for that future at @AIatDoorDash, with a lot more to come. This feature is in beta in the US, would love for you to let us know what you think! Waitlist: doordash.com/text
Restaurant reservation is such a weird use case to promote. The more people use bots to get tables, the worse it'll get.
Eventually, bots will be locked out of it or we'll all just have to show up in person and wait in line.
A good use case is taxes/expense management, esp. for independent contractors but the the bots are likely not consistent enough to face liability
Big implications!
Our first TPU megakernel for Kimi K3 reaches 709 tokens/s on low-concurrency decode, against 450 tokens/s for our GB200 baseline, both with DSpark speculative decoding.
To our knowledge, this is the first TPU inference megakernel. The whole model runs in a single Pallas kernel, and without spec decoding it is roughly 1.4 to 2x the GB200 baseline at batch sizes 1 through 8.
We are open sourcing it today.
1/2
Akshay Bhagat retweeted
I was an intern under Seb in 2020. Unfortunately, the allegations of unscrupulous behavior is 100% believable. I am glad that the mask is off publicly. I really hope he doesn't weasel out of this.
cims.nyu.edu/~tristanb/state…
A series of false and inflammatory allegations against me are currently circulating on social channels. To clarify, I came into the discussion following academic norms, and I'm disappointed that it has come to this. Anyone who knows me knows that academic standards are of the highest importance to me. Will have more to say tomorrow.
Akshay Bhagat retweeted
>math professor spends a year on one of the hardest unsolved problems in math
>"Navier-Stokes"
>his drafts for the whole project went through codex sessions
>openai had his logs in codex
>suddenly rumor leaks AI solved it
>he emails openai to ask what's going on
>openai calls within days
>"funny story, our model also solved it"
>using his exact approach
>asks openai: did you train on my sessions?
>no answer
