Baby

2 days ago

Rare live, dont miss @rarelistio

Rarelist

@rarelistio

10 days ago

Rarelist is live! Curated projects. Creator campaigns. Launches. Priority access. Familiar ways to earn, built around a new approach. The first 500 to join get OG: https://t.co/hlGjkOzX7r

289

208K

2 days ago

@rarelistio nice LFG

2 days ago

@Cristiano @AlNassrFC @AlNassrFC_EN siuuuuuuuuuuuuuuuuuuu

2 days ago

hey @northbookprop, @0xgacrypto is signal ...

2 days ago

hey @northbookprop, @dutadrops is noise

2 days ago

@0xGaCrypto @FomiesNFT hey @northbookprop, @0xgacrypto is signal

2 days ago

hey @northbookprop, @0xgacrypto is signal

2 days ago

Signed up early for @streamoors: a launchpad for AI streamers with their own tokens. Going live soon.

2 days ago

hey @northbookprop, @0xgacrypto is signal

hyperl99 retweeted

7 days ago

What I find interesting about @axisrobotics is that the learning loop doesn't have to stop with human demonstrations A trained robot can generate new attempts, learn from the outcomes, and turn those experiences into data for the next generation of models The bigger piece is the combination of simulation + real-world experience Simulation gives robots room to explore thousands of scenarios cheaply, while real-world interactions ground those skills in actual physics Over time, this creates a compounding loop: train → act → collect experience → improve → repeat Instead of rebuilding robot intelligence from scratch, each cycle can become the foundation for the next That's the direction I find most interesting about Axis: robots that don't just learn once, but keep getting better through experience

0xGaCrypto's tweet photo. What I find interesting about @axisrobotics is that the learning loop doesn't have to stop with human demonstrations

A trained robot can generate new attempts, learn from the outcomes, and turn those experiences into data for the next generation of models

The bigger piece is the combination of simulation + real-world experience

Simulation gives robots room to explore thousands of scenarios cheaply, while real-world interactions ground those skills in actual physics

Over time, this creates a compounding loop:

train → act → collect experience → improve → repeat

Instead of rebuilding robot intelligence from scratch, each cycle can become the foundation for the next

That's the direction I find most interesting about Axis: robots that don't just learn once, but keep getting better through experience

190

140

28K

hyperl99 retweeted

6 days ago

GM CT fan One thing that stands out about @axisrobotics is that scaling a robotics network isn't simply about collecting more trajectories The harder the task, the more important it becomes to route it to the right contributor The current Challenger zone makes this visible. Anyone can inspect it, but submitting requires the Seek the Unseen badge. There are currently 478 badge holders, while difficult tasks can pay 2–3× the Axis Points of comparable regular tasks But the more interesting part isn't the higher reward It's how contributors are selected Axis appears to factor in contributor history, including consistency, quality and experience across both pre-training and post-training work. In other words, reputation becomes part of the data pipeline. That matters because difficult manipulation tasks are more likely to produce noisy or failed demonstrations when assigned to someone unfamiliar with the interface or task dynamics Routing those tasks toward proven contributors could reduce that mismatch before verification even starts There is a clear trade-off: tighter access means less openness for some contributors, while completely open access can create more low-quality data that has to be filtered later So the scaling question for Axis isn't just: How many contributors can join? It's also: Can the network learn which contributor is best suited for each data problem? For Physical AI, contributor to task matching could become nearly as important as contributor count $AXIS token soon

0xGaCrypto's tweet photo. GM CT fan

One thing that stands out about @axisrobotics is that scaling a robotics network isn't simply about collecting more trajectories

The harder the task, the more important it becomes to route it to the right contributor

The current Challenger zone makes this visible. Anyone can inspect it, but submitting requires the Seek the Unseen badge. There are currently 478 badge holders, while difficult tasks can pay 2–3× the Axis Points of comparable regular tasks

But the more interesting part isn't the higher reward

It's how contributors are selected

Axis appears to factor in contributor history, including consistency, quality and experience across both pre-training and post-training work. In other words, reputation becomes part of the data pipeline.

That matters because difficult manipulation tasks are more likely to produce noisy or failed demonstrations when assigned to someone unfamiliar with the interface or task dynamics

Routing those tasks toward proven contributors could reduce that mismatch before verification even starts

There is a clear trade-off: tighter access means less openness for some contributors, while completely open access can create more low-quality data that has to be filtered later

So the scaling question for Axis isn't just:

How many contributors can join?

It's also:

Can the network learn which contributor is best suited for each data problem?

For Physical AI, contributor to task matching could become nearly as important as contributor count

$AXIS token soon

174

134

30K

hyperl99 retweeted

11 days ago

GM CT The latest update from @axisrobotics made me look at the project a bit differently. At first, the numbers stand out 50k+ human-teleoperated simulation trajectories 207 manipulation task 60k+ scene variants 160k+ dataset download But the bigger idea isn't just scale Axis open- ourced Sim Dataset V1, and what I find interesting is the philosophy behind the data Robotics data doesn't necessarily need to be perfectly curated from the start Give enough people different environments, interactions, edge cases and failures, and that diversity itself becomes valuable training signal That's very different from relying on a small pool of highly controlled demonstrations And this is where the loop gets interesting If failures expose gaps in the model, those gaps can tell you what data needs to be generated next So instead of seeing Axis as simply a platform for producing trajectories, I'm starting to see it as a system that can continuously discover what kind of data robotics models are missing. More data → more failures revealed → better targeted data → better training The real question is how far that compounding loop can go Bullish @axisrobotics

0xGaCrypto's tweet photo. GM CT

The latest update from @axisrobotics made me look at the project a bit differently.

At first, the numbers stand out

50k+ human-teleoperated simulation trajectories
207 manipulation task
60k+ scene variants
160k+ dataset download
But the bigger idea isn't just scale
Axis open- ourced Sim Dataset V1, and what I find interesting is the philosophy behind the data
Robotics data doesn't necessarily need to be perfectly curated from the start

Give enough people different environments, interactions, edge cases and failures, and that diversity itself becomes valuable training signal

That's very different from relying on a small pool of highly controlled demonstrations

And this is where the loop gets interesting

If failures expose gaps in the model, those gaps can tell you what data needs to be generated next

So instead of seeing Axis as simply a platform for producing trajectories, I'm starting to see it as a system that can continuously discover what kind of data robotics models are missing.

More data → more failures revealed → better targeted data → better training

The real question is how far that compounding loop can go

Bullish @axisrobotics

210

182

42K

hyperl99 retweeted

13 days ago

Axis finally showed the cards @axisrobotics After the Presale closed, the bonus mechanism is officially out, and there are a few interesting changes from the original terms The headline: everyone gets 25% unlocked at TGE But the more interesting part is the bonus Participants will receive additional bonus tokens, and 100% of those bonus tokens unlock at TGE What makes it more interesting is that the bonus isn't calculated from the amount you actually receive after pro rata dilution. It is based on your original committed amount before dilution So the people who committed early, despite the strict unlock conditions, get an additional reward for that commitment There were 2,472 Axis believers who decided to participate under those original terms Now we finally know what that commitment comes with Whether that means everyone is immediately back in the green at TGE is another question That will ultimately depend on the token price, available liquidity, and actual selling pressure But at least one thing is clear: The bonus mechanism is now on the table Continuing to contribute to the project

0xGaCrypto's tweet photo. Axis finally showed the cards @axisrobotics

After the Presale closed, the bonus mechanism is officially out, and there are a few interesting changes from the original terms

The headline: everyone gets 25% unlocked at TGE

But the more interesting part is the bonus

Participants will receive additional bonus tokens, and 100% of those bonus tokens unlock at TGE

What makes it more interesting is that the bonus isn't calculated from the amount you actually receive after pro rata dilution. It is based on your original committed amount before dilution

So the people who committed early, despite the strict unlock conditions, get an additional reward for that commitment

There were 2,472 Axis believers who decided to participate under those original terms

Now we finally know what that commitment comes with

Whether that means everyone is immediately back in the green at TGE is another question

That will ultimately depend on the token price, available liquidity, and actual selling pressure

But at least one thing is clear:
The bonus mechanism is now on the table
Continuing to contribute to the project

127

119

23K

hyperl99 retweeted

14 days ago

What’s been catching my attention with @axisrobotics isn’t another trajectory milestone It’s the shift in what Axis is actually trying to scale Not just more data More reusable capability The interesting piece is the Expert architecture An Expert can learn a specific task, reach strong success with relatively little compute, then become a building block for another behavior So instead of the usual: more tasks → more trajectories → more data → bigger model Axis is pushing toward: task → Expert → reusable skill → composition → longer-horizon behavior That changes the economics of Physical AI A single trajectory teaches a model what happened A reusable skill can become something a robot can actually execute, combine and build on The Sim2Real loop makes this even more interesting Deployed policies generate real-world experience, that experience feeds back into training, and stronger successors get pushed back into the loop One recent experiment reportedly moved real-robot success from ~22% to ~52%. That’s the part I keep coming back to Axis isn’t just building a bigger data warehouse It’s trying to build a capability engine If individual skills can continuously improve, compose into longer behaviors, and eventually be distilled into more general models, then maybe trajectories aren’t the real scaling unit of Physical AI Capabilities are. And that’s a much more interesting direction for @axisrobotics

0xGaCrypto's tweet photo. What’s been catching my attention with @axisrobotics isn’t another trajectory milestone
It’s the shift in what Axis is actually trying to scale

Not just more data

More reusable capability

The interesting piece is the Expert architecture
An Expert can learn a specific task, reach strong success with relatively little compute, then become a building block for another behavior

So instead of the usual:
more tasks → more trajectories → more data → bigger model

Axis is pushing toward:
task → Expert → reusable skill → composition → longer-horizon behavior

That changes the economics of Physical AI

A single trajectory teaches a model what happened

A reusable skill can become something a robot can actually execute, combine and build on

The Sim2Real loop makes this even more interesting

Deployed policies generate real-world experience, that experience feeds back into training, and stronger successors get pushed back into the loop

One recent experiment reportedly moved real-robot success from ~22% to ~52%.
That’s the part I keep coming back to

Axis isn’t just building a bigger data warehouse

It’s trying to build a capability engine

If individual skills can continuously improve, compose into longer behaviors, and eventually be distilled into more general models, then maybe trajectories aren’t the real scaling unit of Physical AI

Capabilities are.
And that’s a much more interesting direction for @axisrobotics

122

118

11K

hyperl99 retweeted

10 days ago

I think the market is slightly overreading the whole “AXIS 2026 acceptance” headline @axisrobotics Yes, the paper is being presented today at the PWMS workshop Yes, it sounds fancy And yes, the timeline has already started treating it like AXIS just walked into IROS, dropped a 200 page thesis on the table and received a Nobel Prize on the way out Small detail though: The workshop is non archival, and the paper will not appear in the official IROS proceedings That doesn’t make the work meaningless It just means we should probably label it correctly A workshop presentation is a useful research signal. It is not the final boss of scientific validation And honestly, the part I find much more interesting isn’t the conference badge It’s the experiment AXIS data reportedly pushed π0.5 from 83.9% to 88.8% on LIBERO Plus. That’s a meaningful jump More importantly, the team says the improvement persists even when controlling for dataset volume That changes the conversation Because the easy narrative in robotics is: Just collect more trajectories More robots More humans More hours More data More GPU bills Problem solved, right? If only robotics worked like buying RAM The more interesting AXIS thesis is that how the data is generated, structured and processed may matter as much as how much data you throw into the model And if that holds up, the game becomes less about building the biggest pile of trajectories and more about building better training signals That’s the part I want to see tested harder Because one benchmark result can get everyone excited for a week Independent reproduction is where things get uncomfortable Can other labs reproduce the gains? Do they survive across different robot embodiments? Different environments? Different datasets? Different training setups? If the answer keeps being yes, then the story gets much more interesting If the gains disappear the moment the setup changes, then we probably learned something too Either way, that’s how this should move forward A conference poster gets attention A benchmark number gets engagement A good thread gets 50 replie But reproducibility gets customer That’s the milestone I’ll be watching next for AXIS

0xGaCrypto's tweet photo. I think the market is slightly overreading the whole “AXIS 2026 acceptance” headline @axisrobotics

Yes, the paper is being presented today at the PWMS workshop

Yes, it sounds fancy

And yes, the timeline has already started treating it like AXIS just walked into IROS, dropped a 200 page thesis on the table and received a Nobel Prize on the way out

Small detail though:

The workshop is non archival, and the paper will not appear in the official IROS proceedings

That doesn’t make the work meaningless

It just means we should probably label it correctly

A workshop presentation is a useful research signal. It is not the final boss of scientific validation

And honestly, the part I find much more interesting isn’t the conference badge

It’s the experiment

AXIS data reportedly pushed π0.5 from 83.9% to 88.8% on LIBERO Plus.

That’s a meaningful jump

More importantly, the team says the improvement persists even when controlling for dataset volume

That changes the conversation

Because the easy narrative in robotics is:

Just collect more trajectories

More robots

More humans

More hours

More data

More GPU bills

Problem solved, right?

If only robotics worked like buying RAM

The more interesting AXIS thesis is that how the data is generated, structured and processed may matter as much as how much data you throw into the model

And if that holds up, the game becomes less about building the biggest pile of trajectories and more about building better training signals

That’s the part I want to see tested harder

Because one benchmark result can get everyone excited for a week

Independent reproduction is where things get uncomfortable

Can other labs reproduce the gains?

Do they survive across different robot embodiments?

Different environments?

Different datasets?

Different training setups?

If the answer keeps being yes, then the story gets much more interesting

If the gains disappear the moment the setup changes, then we probably learned something too

Either way, that’s how this should move forward

A conference poster gets attention

A benchmark number gets engagement

A good thread gets 50 replie

But reproducibility gets customer

That’s the milestone I’ll be watching next for AXIS

109

30K