Mathapelo mercy @MercyMathapelo

Mpumalanga trichartd Joined February 2018

Tweets

11
Followers

10
Following

55
Likes

20

Mathapelo mercy @MercyMathapelo

20 hours ago

An AI chatbot being the explanation here is just wild

Hoodlum 🇺🇸 @NotHoodlum

2 days ago

See? I knew she’d have an excuse that made sense. Gargoyle wasn’t bribing people. It was her AI chatbot. 🥴 wsj.com/politics/polic…

59 72 296 7K 7

0 0 2 3 0

View Details

Mathapelo mercy @MercyMathapelo

2 days ago

What an incredible life she lived. Fresh bread with butter is hard to beat

Colette Harrington @sweetcarolinatv

3 days ago

It’s been 12 years since my Grandmother passed away. She was a former Miss Michigan. Her heart was in sewing, baking, and being homemaker. She baked bread for the week on Friday. Nothing like fresh baked bread and butter. She raised 14 kids and lived her entire life on a farm.

14 7 126 2K 0

0 0 0 7 0

View Details

Mathapelo mercy @MercyMathapelo

3 days ago

Does anyone let AI fix a bug without reading the whole diff first?

0 0 0 7 0

View Details

Mathapelo mercy @MercyMathapelo

4 days ago

Walked past this fence and had to stop. Somebody was VERY serious about keeping people out

0 0 0 9 0

View Details

Aliba @Aliba_79

6 days ago

✅ @axisrobotics and the Grammar of Embodied Intelligence : When Data Becomes a Living Language Most discussions around Axis focus on it as a data engine solving Physical AI’s data bottleneck . That is accurate , but incomplete . The real distinction is not the sheer volume of trajectories collected . It is how those trajectories are transformed into a structured form that a model can read , identify gaps in , and actively request more of over time Consider the difference between a pile of isolated words and a working grammar . Language models advanced once text was organized into learnable structures . Axis is attempting something similar for physical action . Its unified schema does more than store joint states and end-effector commands . It standardizes 4D hand-pose sequences alongside subtask-level temporal alignment . A simple action such as “pick up the watering can and water the plant” stops being a raw motion sequence and becomes a coherent sentence with a subject , a verb , and a resulting state The model no longer learns from fragmented video . It learns from a language of action that has syntax The more interesting mechanism is model-conditioned data generation . Instead of humans independently deciding the next task , a trained policy surfaces its own weaknesses through failure mining and policy-gap analysis . The system then constructs new tasks specifically to address those gaps . This is no longer one-way data collection . It is the model posing questions that the community answers with corrective trajectories . Human-gated DAgger in the post-training stage makes this concrete: the policy is allowed to act , and a human intervenes only when it fails . Those interventions become the highest-value data because they sit precisely at the edge of the model’s current capability A less-discussed technical detail is the refinement pipeline . Smoothing raw teleoperation trajectories reduces mean acceleration by 63.9% and mean jerk by 80.8% . The goal is not merely cleaner individual demos . It is a distribution that remains usable even when single trajectories are noisy . Axis demonstrated this with continual pretraining on π0.5 : success on LIBERO-Plus rose from 83.9% to 88.8% , with the largest gains under sensor noise and camera perturbations—the exact conditions where distribution-level cleanliness matters most Combined with on-chain provenance on Base (each trajectory can be signed and linked to a Data ID) , Axis is building a form of attributable embodied memory . Every contribution is not just a score; it is a sentence in a shared language of action that can be traced and later used by the model to improve itself . This is the angle still underexplored in most discussions : not simply “collect more data,” but construct a protocol for how embodied skill is expressed , refined , and compounded . Failures become targeted requests for the next sentence rather than discarded noise If Physical AI ultimately requires something equivalent to an internet of physical interactions , Axis is testing an early grammar for that internet—not by recording everything indiscriminately , but by letting the model identify which sentences are missing and asking the community to write them with purpose . That is the actual compounding mechanism

248 2 219 4K 1

View Details

Mathapelo mercy @MercyMathapelo

5 days ago

A half-billion parameter model turning a phone video into a 3D room is pretty wild

Md Ismail Šojal 🕷️ @0x0SojalSec

6 days ago

a tiny AI 0.5B just turned a normal phone video into a fully labeled 3D room run phone. SpatialLM that understands walls, doors, furniture, and everything layout. it trained as a standard multimodal LLM. Most vision models see pixels, SpatialLM sees space. Open models :

27 204 2K 67K 2K

0 0 0 8 0

View Details

Mathapelo mercy @MercyMathapelo

7 days ago

What do you guys make with chicken thighs when you don't feel like grilling?

0 0 0 5 0

View Details

Mathapelo mercy @MercyMathapelo

a week ago

@BullTheoryio Burry calling a bubble is gonna get people arguing that's for sure

0 0 0 106 0

View Details

Mathapelo mercy @MercyMathapelo

a week ago

Banana bread and chocolate chips is a pretty easy sell for me

Recipes by Clare @recipesbyclare

a week ago

Banana Bread Cookies Ingredients: - 3 large ripe bananas - 1 cup granulated sugar - 1 stick (1/2 cup) unsalted butter, softened - 2 eggs - 1 teaspoon vanilla extract - 2 cups all-purpose flour - 1 teaspoon baking soda - 1/4 teaspoon salt - 1 cup semisweet chocolate chips

1 16 103 4K 77

0 0 1 7 0

View Details

Mathapelo mercy @MercyMathapelo

a week ago

That track is getting swallowed by the woods, that's a hell of a cleanup

NCC-1701-A Starship 🏴󠁧󠁢󠁳󠁣󠁴󠁿🚄 @1701aStarship

a week ago

Well another working day at PG and back into tidying up the trackwork and getting started on the thick vegetation near the crossing gate. Not for the faint hearted.

4 7 126 4K 1

0 0 1 8 0

View Details

Mathapelo mercy @MercyMathapelo

a week ago

Getting reliable real-world data for robots at scale is gonna be tough

Tazik smerti @andawey12

a week ago

Physical AI is currently receiving a lot of attention, especially in areas related to models and robotics. However, another important component is needed for these systems to work effectively: high-quality data from the real world. This is exactly the direction Vangrid is working