←
Jacob Van Meter79 posts
Jacob Van Meter avatar

Jacob Van Meter

@jacobvm04

robotics/ai research | unc dropout | prev @plasticlabs

◷ Joined April 2023▧ 16 Photos and videos
Jacob Van Meter avatar
Jacob Van Meter@jacobvm04·Nov 15

We're excited to release the first open-source VIO to beat Aria on LaMAria! Industry level SLAM for everybody will power tons of applications, can't wait to see what people build :) So how did we beat Aria's SLAM? (1/4)

Post attachment

when LaMAria was introduced a year ago, Meta's closed-source Aria SLAM was far ahead of academic systems. one year later, we've surpassed it in 3/5 categories: medium, long, and low light. here's what closing that gap looks like:

0:17
ayush paul avatar
⟳ Jacob Van Meter retweeted
ayush paul@hyusapx·Nov 16

a camera gives you 2d pictures, not 3d measurements. how do you figure out where you are from that? turns out SLAM (simultaneous localization and mapping) is one of the hardest problems in robotics. applications range from AR glasses to drones and autonomous cars! 1/5

We're excited to release the first open-source VIO to beat Aria on LaMAria! Industry level SLAM for everybody will power tons of applications, can't wait to see what people build :) So how did we beat Aria's SLAM? (1/4)

Post attachment
Jacob Van Meter avatar
Jacob Van Meter@jacobvm04·Nov 15

Modeling camera calibration drift is increasingly important as trajectories grow longer. We find that the ratio between the error of the constant calibration system and the time-varying camera calibration system strongly increases with trajectory length. (3/4)

Post attachment
Jacob Van Meter avatar
Jacob Van Meter@jacobvm04·Nov 15

The core of SLAM Dunk is a conventional visual-inertial bundle adjustment pipeline. By just adding time-varying camera calibration and gyroscope intrinsics calibration to our BA, we beat Aria on 3/5 categories. (2/4)

Post attachment
Shubham Patil avatar
⟳ Jacob Van Meter retweeted
Shubham Patil@shubhampatilsd·Nov 14

how well can fable 5.2 and gpt-6 astra act as air traffic controllers? in light of recent atc closures, i wondered how efficiently the most chaotic, high-consequence tasks could be carried out by frontier models. results below :)

0:05

JUST IN: 🇺🇸 FAA issues nationwide ground stop for all flights. All flight departures have been halted.

ayush paul avatar
⟳ Jacob Van Meter retweeted
ayush paul@hyusapx·Nov 10

fable 5.2 is surprisingly good at spatial reasoning! here's it scanning and bagging groceries across multiple pick and place steps. seems like the labs are beginning to scale physical post-training. excited to see where this goes.

0:32

Fable 5.2 shows substantially improved skills in controlling physical robots, even on tasks it has never explicitly been trained to do. On RoboDojo, a benchmark with a simulated two-arm robot, it outperformed other models and most AI systems built to control robots.

0:23
Jacob Van Meter avatar
Jacob Van Meter@jacobvm04·Nov 7

SLAM Dunk is sota on uzh-fpv too 👀

0:23
Shubham Patil avatar
⟳ Jacob Van Meter retweeted
Shubham Patil@shubhampatilsd·Nov 6

i got gpt-2 running purely in css! all 124m parameters inferencing at 15 sec/tok, with zero lines of js. i also used a custom font as a BPE tokenizer, and variable fonts for matmuls :)

0:19
ayush paul avatar
⟳ Jacob Van Meter retweeted
ayush paul@hyusapx·Nov 5

saw that @meetgranola was trying to compete with @WisprFlow on dictation so i ran a quick test. overall formatting and quality seems better on wispr!

0:49
Jacob Van Meter avatar
Jacob Van Meter@jacobvm04·Nov 4

dm for early access. write up coming soon!

when LaMAria was introduced a year ago, Meta's closed-source Aria SLAM was far ahead of academic systems. one year later, we've surpassed it in 3/5 categories: medium, long, and low light. here's what closing that gap looks like:

0:17
Jacob Van Meter avatar
Jacob Van Meter@jacobvm04·Nov 3

when LaMAria was introduced a year ago, Meta's closed-source Aria SLAM was far ahead of academic systems. one year later, we've surpassed it in 3/5 categories: medium, long, and low light. here's what closing that gap looks like:

0:17

wow @jacobvm04 and i just took first place on one of robotics’ hardest problems. machines knowing where they are in the world. slam dunk.

Post attachment
Qiang.W avatar
⟳ Jacob Van Meter retweeted
Qiang.W@supertoday·Nov 2

I just checked the current LaMAria leaderboard. The gap to Aria's SLAM is much smaller. All methods in the sceenshot except Aria's SLAM are posted within one month. Wild. #SLAM #VIO #LaMAria

Post attachment
Ben C avatar
⟳ Jacob Van Meter retweeted
Ben C@BdcauntBen·Oct 26

Jev harness for robot navigation

2:50

Building a new harness to control robots with TypeSafe AI Jev and Cloudflare Clairic. Prompt "Go to the kitchen and find the stove"

0:33
Jacob Van Meter avatar
Jacob Van Meter@jacobvm04·Oct 15

faith is all you need

Post attachment

interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works already-clean.goodfire.com/