“They thought the Icelandic dating market was going to be bigger than AI.” - @EricEncord
When Encord went through YC during COVID, most of the batch was building fintech, crypto and remote work tools. Encord made it to the final round with one fund for its seed, and the fund t think AI was a big enough market. Its next investment was an Icelandic dating company.
#LongHorizon
When Encord went through YC during COVID, most of the batch was building fintech, crypto and remote work tools. Encord made it to the final round with one fund for its seed, and the fund t think AI was a big enough market. Its next investment was an Icelandic dating company.
#LongHorizon
0
1
3
231
Today, Sebastian Thrun (@SebastianThrun) announced a $130M seed for Dulo, his new startup building foundation models for hardware design.
Last week at Long Horizon, he shared his take on how hardware gets designed today:
“The tool I hate the most by a large margin is CAD.”
Last week at Long Horizon, he shared his take on how hardware gets designed today:
“The tool I hate the most by a large margin is CAD.”
0
0
3
258
Daniel Burrows, Head of Data Platform at @Mind_Robotics, on instrumenting a live automotive plant for robot learning at Long Horizon.
1
1
6
283
Encord retweeted
At @encord_team's Long Horizon, our co-founder and CTO @Ben_Burchfiel shared his take on 2026: general-purpose robot manipulation is having its “GPT moment”. Progress is accelerating, and the field is moving beyond demos toward real commercial value.
Thanks for having us!
Thanks for having us!
0
3
5
382
Encord retweeted
End-to-end deep learning is not only the best approach for driving models, but also for safety, world and reward models.
I was back in San Francisco last week to share more about our vision at the #LongHorizon summit on Embodied AI.
I was back in San Francisco last week to share more about our vision at the #LongHorizon summit on Embodied AI.
1
4
42
2,334
Eric Landau (@EricEncord) and Ulrik Stig Hansen (@ulrikstighansen) with Nicolas Dessaigne (@dessaigne) of YC on building the data layer for Physical AI.
0
1
10
405
“I think robotics is going to go through a hype cycle just like autonomy did.” - @alexgkendall, CEO of @wayve_ai
Alex’s argument is that the route to a general-purpose robotics company runs through self-driving. Driving is where you can deploy millions of vehicles doing useful a and revenue from that is what funds the move into everything else.
Alex’s argument is that the route to a general-purpose robotics company runs through self-driving. Driving is where you can deploy millions of vehicles doing useful a and revenue from that is what funds the move into everything else.
0
2
22
1,811
“I feel like there’s a VLA crisis right now.” - @Ken_Goldberg
Ken’s argument is that the field should build bottom-up: ship lots of machines that work, earn revenue, and expand their capabilities over time until they generalize toward a humanoid. Going general-purpose from day Vision-Language-Action models have hit a plateau.
Ken’s argument is that the field should build bottom-up: ship lots of machines that work, earn revenue, and expand their capabilities over time until they generalize toward a humanoid. Going general-purpose from day Vision-Language-Action models have hit a plateau.
1
0
12
574
Capture what cameras can't. Introducing EMG-1, the largest open dataset pairing wrist muscle signals with egocentric video – opening up new possibilities for studying object interaction at scale.
1
0
5
202
In our initial experiments, combining EMG with video reduced fingertip tracking error by 25% against a camera-only reference. It also reduces finger articulation error by 35% during tracking gaps.
Beyond hand movement, EMG can also indicate object contact and grip force.
Beyond hand movement, EMG can also indicate object contact and grip force.
1
0
0
33
We're releasing EMG-1 free of charge so researchers can explore new applications for muscle signals in robot learning.
Explore the dataset and get access: encord.com/data-labs/intr…
Explore the dataset and get access: encord.com/data-labs/intr…
0
0
0
24
“For humans, a backflip is very hard, but for robots, it’s very easy” - Deepak Pathak, CEO of @SkildAI
@deepakpathak's argument is that free-space moves like dancing and backflips are the easy part, because the robot can track every joint. Climbing arbitrary stairs is far ot has to see and react to the world around it.
@deepakpathak's argument is that free-space moves like dancing and backflips are the easy part, because the robot can track every joint. Climbing arbitrary stairs is far ot has to see and react to the world around it.
1
0
5
252
“We don’t need to have our prime engineers understand the code” - Alberto Rodriguez, @BostonDynamics.
@_albertorod_ lays out that instead of hand-engineering behavior trees, you can now tell Atlas how to handle exceptions in natural language. Anyone on the Hyundai factory floor from failure data and deploy it.
@_albertorod_ lays out that instead of hand-engineering behavior trees, you can now tell Atlas how to handle exceptions in natural language. Anyone on the Hyundai factory floor from failure data and deploy it.
0
0
9
409
"What if you didn't have to go through classical compute limits at all?" - @PimDeWitte
Pim’s argument is that the optimal approach is to learn directly from the data the universe produces continuously, rather than relying on simulators. On top of that, the training runs are er LLMs.
Pim’s argument is that the optimal approach is to learn directly from the data the universe produces continuously, rather than relying on simulators. On top of that, the training runs are er LLMs.
0
6
33
20,941
Introducing Encord-REACH: an end-to-end pipeline for real-world egocentric hand tracking.
State-of-the-art 3D hand-pose accuracy, near-perfect wearer/bystander tracking and 30–70% lower error on industrial gloves.
See it in action ↓
State-of-the-art 3D hand-pose accuracy, near-perfect wearer/bystander tracking and 30–70% lower error on industrial gloves.
See it in action ↓
1
2
9
441
Finally, we rounded out our approach by building a modular and flexible pipeline which can accommodate multimodal information flows while providing consistent annotations.
On TACO, adding depth cuts mean joint error by 28% and p90 error by 41%. Visual odometry and head IMUs also cking.
On TACO, adding depth cuts mean joint error by 28% and p90 error by 41%. Visual odometry and head IMUs also cking.
1
0
0
26
The recent surge in egocentric data collection has created a real need for accurate and robust hand tracking techniques.
Encord-REACH addresses these real-world challenges in an end-to-end pipeline. It is now in production with leading robotics companies.
Read our article 'Hands at Work' for more: encord.com/data-labs/hand…
Encord-REACH addresses these real-world challenges in an end-to-end pipeline. It is now in production with leading robotics companies.
Read our article 'Hands at Work' for more: encord.com/data-labs/hand…
0
0
0
29