Pinned Tweet
let's see some crazier bimanual demos yeah?
0
1
8
1,180
we will be doing lots of validation testing and private events/demos over the coming year

history will be made in this room
0
0
6
244
using number formats you've never heard of in visual spaces you've only dreamt of
0
0
2
210
we've made rapid, intense upgrades to our multimodal bayesian model

we wanted to see how our latest model J2.3 compares to Nvidia's FLORA - the results are yet again so stunning

rather than benchmaxx, we will retest the 2 straggler regions when J2.4 comes out!
We built a robotics data engine with Astra-level performance at 125x lower cost. And we are open-sourcing the weights today! Introducing Flora, a suite of open models to parse the physical world and turn raw footage into rich training signals.

Flora already powers our in-house 00K hours of videos a day on 500 GPUs. We created it to scale the frontier for robot learning. Now you own the weights too.

Flora gives you:

- High-quality, timestamped captions to curate and filter human egocentric and robot data;
- Dense progress rewards for quality control and RL post-training;
- Rubrics scoring for real-world benchmarking

Try it today, links in thread:
1
1
9
1,613
Replying to @solidSF
so what's the 65K? we figured a way to parameterize the model with fixed point!

it's perhaps a more convenient measure to understand our model size vs firepower for this new class of models, for this benchmark it was particularly relevant
2
0
2
154
Replying to @solidSF
ELI5:

this is a gamified video challenge where a premier model from openai underperforms a specialized model made by nvidia

and then i show my model on bayesian predictive arch at a tiny fraction of the nvidia model outperforming both of them
0
0
0
38
Replying to @solidSF
we love nvidia and will continue to benchmark ourselves against their models/suites, and others as well

the results of the comparison are being shared for clarity and illustration of the absurd step change in capability this architecture affords
1
0
1
111
Replying to @solidSF
additional suites are rolling out soon that leverage the full range of J2.3 improvements, and J2.4 is always on the way

prices will rise to meet demand and balance our compute operations, so snag a grandfather price while you can :)
0
0
0
69
J2.1 is tuned up for the IDE finally, discovered lots of issues with windows/linux and ended up improving things across the board

2.2 will be out soon after a new series of h2h training, but if you were waiting for a real agentic interaction with no coherency bugs, try now :)
0
0
2
324
one-shot video2cad + video2world

superswarming for the fun of it, not posting files
0
0
6
549
windows, mac, linux :)

all updated to run J2.1

have fun!

solidsf.com/download
0
0
6
338
she is only getting cooler at expressing herself
3
0
8
656
Replying to @solidSF
sick
0
0
0
138
we are moving from being p chill to being cryogenically ice cold

100% designed, drawn, analyzed, animated, and produced by Jesse 1.4

all i asked for was a CO2 canister-style cold discharge fitting to make a freeze tip beyond a gas expansion

as a career thermalfluids guy, WOW
1
0
1
469
awesome
1
0
2
232
Replying to @solidSF
the hilbert order of her brain is insane
0
0
0
128
we should make all the benchmarking companies figure out how to benchmark proof writing skills so we can finally join high sciences with ground level engineers
0
0
0
324
we applaud all ai/si leaders in trying to get this right
0
0
11
677
this new era will save so many lives and help yall make so much sense of your own life

we love yall
0
1
4
418
we are constructing a superintelligence benchmark

good luck to all competitors
0
0
6
510
pure luck actually idk prolly fluke or somethin in the scoring engine idek
1
0
5
429