Loading video...

Video Failed to Load

Go Home

Test-time learning has a catch that scaling alone does not solve: the state a system writes to in order to learn is also the record a user must inspect to audit it. Optimize it only for learning, and auditability usually degrades. Blog: Part I treats the learning state and...

rei_labs's profile picture

Rei

16,092 subscribers

37,866 views • 1 month ago •via X (Twitter)

LIVE

Anya Rossi• Live Now

Private livecam show

0 Comments

No comments available

Comments from the original post will appear here

Related Videos

A couple of months later, in a very similar situation, but this time, Balde is calm enough to not only see the pass but, more importantly, execute it properly and not rush it. The boy is learning.
0:15

A couple of months later, in a very similar situation, but this time, Balde is calm enough to not only see the pass but, more importantly, execute it properly and not rush it. The boy is learning.

sm

759,377 views • 3 years ago

The ultimate blackpill is learning that Gavin Newsom and the CA State Park employees would rather see you burn alive than to see a firefighter take a shovel to this plant on Topanga State Park during a wildfire. Not hyperbole.
0:17

The ultimate blackpill is learning that Gavin Newsom and the CA State Park employees would rather see you burn alive than to see a firefighter take a shovel to this plant on Topanga State Park during a wildfire. Not hyperbole.

Mann Made Cinema

1,200,015 views • 9 months ago

This is the best AI agent-first notes app I've found. It's called . It has the potential to be a productized version of Karpathy's "LLM Wiki" knowledge bases. Here's what I did: 1. Imported "Learning, Fast and Slow" a Continual Learning paper 2. Asked OpenKnowledge to create a visual explainer 3. Read the explainer and had Claude explain to me 4. Created a new section of my own understanding 5. Saved the durable version in my Obsidian vault because it's markdown Free and open-source. Absolutely incredible learning tool.
0:14

This is the best AI agent-first notes app I've found. It's called . It has the potential to be a productized version of Karpathy's "LLM Wiki" knowledge bases. Here's what I did: 1. Imported "Learning, Fast and Slow" a Continual Learning paper 2. Asked OpenKnowledge to create a visual explainer 3. Read the explainer and had Claude explain to me 4. Created a new section of my own understanding 5. Saved the durable version in my Obsidian vault because it's markdown Free and open-source. Absolutely incredible learning tool.

Dan McAteer

17,326 views • 2 months ago

This is Milo. He has the reflexes of a sloth and the composure of a statue. Learning catch is going to take a minute. 12/10
0:10

This is Milo. He has the reflexes of a sloth and the composure of a statue. Learning catch is going to take a minute. 12/10

WeRateDogs

311,136 views • 7 months ago

learning how to sing, play guitar & harmonica, as well as learning the body language and stage presence of a huge rock icon all in the span of only 6 months is incredibly impressive and i will get my client this nom even if i have to do it myself
0:18

learning how to sing, play guitar & harmonica, as well as learning the body language and stage presence of a huge rock icon all in the span of only 6 months is incredibly impressive and i will get my client this nom even if i have to do it myself

lee ⭐️

215,363 views • 8 months ago

Pitt has a 🔥 Student Section. They are in the same tier as Purdue and K State.
0:18

Pitt has a 🔥 Student Section. They are in the same tier as Purdue and K State.

Big Game Boomer

724,155 views • 3 years ago

This figure from HIL-SERL is one of the clearest visualisations of how RL learns differently from imitation learning. The difference comes down to this: imitation learning treats each (state, action) pair as independent. A correction at timestep 20 teaches nothing about timestep 19 or 21. RL propagates reward backward through time. One successful insertion updates the value estimate of every state along the trajectory. So RL builds a full map of "which states lead to success"; imitation learning just memorizes individual snapshots. Setup: a robot inserting a RAM stick into a motherboard slot. Each dot is an end-effector position (Y = lateral, Z = height). Starting position is randomized. Left to right = training progressing. Top row (RL): the policy builds a funnel. Broad at the top, narrowing into the target. It systematically fills in the state space, learning which paths lead to success from many different starting positions. Bottom row (imitation learning / HG-DAgger, same human data): sparse, diffuse, no funnel. The policy only learns near states the human demonstrated. Both have access to the same data, including human corrections, but a completely different structure emerges.
0:14

This figure from HIL-SERL is one of the clearest visualisations of how RL learns differently from imitation learning. The difference comes down to this: imitation learning treats each (state, action) pair as independent. A correction at timestep 20 teaches nothing about timestep 19 or 21. RL propagates reward backward through time. One successful insertion updates the value estimate of every state along the trajectory. So RL builds a full map of "which states lead to success"; imitation learning just memorizes individual snapshots. Setup: a robot inserting a RAM stick into a motherboard slot. Each dot is an end-effector position (Y = lateral, Z = height). Starting position is randomized. Left to right = training progressing. Top row (RL): the policy builds a funnel. Broad at the top, narrowing into the target. It systematically fills in the state space, learning which paths lead to success from many different starting positions. Bottom row (imitation learning / HG-DAgger, same human data): sparse, diffuse, no funnel. The policy only learns near states the human demonstrated. Both have access to the same data, including human corrections, but a completely different structure emerges.

Dominique Paul

24,433 views • 7 months ago

the BAD choreo is becoming so viral bc of the people learning it as a trend AND I LOVE THIS
0:17

the BAD choreo is becoming so viral bc of the people learning it as a trend AND I LOVE THIS

ell

30,566 views • 2 months ago

Payment infrastructure is still fragmented. Open Payments is the open standard designed to change that. Open Payments Learning Month is 30 days of hands-on learning for developers and fintech builders. Learn the API flow, explore the test environment, and explore ideas with the Open Payments team. Free. Self-paced. Starts June 1.
0:14

Payment infrastructure is still fragmented. Open Payments is the open standard designed to change that. Open Payments Learning Month is 30 days of hands-on learning for developers and fintech builders. Learn the API flow, explore the test environment, and explore ideas with the Open Payments team. Free. Self-paced. Starts June 1.

Interledger Foundation

525,972 views • 4 months ago

Code mode is not a one-shot build. If the result does not fully match your product goal, submit a Bubble Up. Bubble Engine will test, evaluate, and iteratively optimize the underlying Coding SOP, moving refinement from the user to the system. State the goal. Build with Bubble Code. Improve with Bubble Up.
0:10

Code mode is not a one-shot build. If the result does not fully match your product goal, submit a Bubble Up. Bubble Engine will test, evaluate, and iteratively optimize the underlying Coding SOP, moving refinement from the user to the system. State the goal. Build with Bubble Code. Improve with Bubble Up.

DAPPOS

28,693 views • 2 months ago

Last year’s National Audit Office report found that 700,000 children were learning in unsuitable school buildings that required major repairs 🚧 It’s unacceptable that across the country, our children are learning in schools with broken boilers, mouldy walls, crumbling concrete, and flooding classrooms 🏚️ This #generalelection, make sure that every school is properly resourced, and each child has the right conditions to learn 📚 Join the campaign now ➡️
0:10

Last year’s National Audit Office report found that 700,000 children were learning in unsuitable school buildings that required major repairs 🚧 It’s unacceptable that across the country, our children are learning in schools with broken boilers, mouldy walls, crumbling concrete, and flooding classrooms 🏚️ This #generalelection, make sure that every school is properly resourced, and each child has the right conditions to learn 📚 Join the campaign now ➡️

Stop School Cuts

10,318 views • 2 years ago

In Super Paper Mario, it is possible for an enemy to die and not die at the same time, entering a glitched state of superposition. Attacking it again is not recommended; the game cannot handle the enemy's paradoxical state and crashes immediately.
0:17

In Super Paper Mario, it is possible for an enemy to die and not die at the same time, entering a glitched state of superposition. Attacking it again is not recommended; the game cannot handle the enemy's paradoxical state and crashes immediately.

Supper Mario Broth

36,992 views • 10 days ago

We removed 93.25% of the connections in our Un-0 image model, fully expecting to pay for it in quality. But it got better. FID 7.15 on ImageNet 64x64, roughly 1.9 ahead of the dense baseline at matched size. Same family of model, a fraction of the couplings, a better score. Here is why that is not as strange as it sounds. Un-0 is a coupled oscillator model, and in the dense version every oscillator talks to every other one. That sounds like a strength, but it means the whole system can fall into catastrophic synchronization: everything locks into step, gradients go flat, and learning stalls. Sparser connectivity leaves room for coherent and incoherent activity to coexist. The dynamics stay alive, and the model keeps learning. Connectivity turns out to be a control knob, not a dial you turn up until it stops. Learn more here:
0:12

We removed 93.25% of the connections in our Un-0 image model, fully expecting to pay for it in quality. But it got better. FID 7.15 on ImageNet 64x64, roughly 1.9 ahead of the dense baseline at matched size. Same family of model, a fraction of the couplings, a better score. Here is why that is not as strange as it sounds. Un-0 is a coupled oscillator model, and in the dense version every oscillator talks to every other one. That sounds like a strength, but it means the whole system can fall into catastrophic synchronization: everything locks into step, gradients go flat, and learning stalls. Sparser connectivity leaves room for coherent and incoherent activity to coexist. The dynamics stay alive, and the model keeps learning. Connectivity turns out to be a control knob, not a dial you turn up until it stops. Learn more here:

Unconventional AI

36,508 views • 1 month ago

Can we make Transformers better and more efficient for robot learning? Excited to introduce Body Transformer (BoT), an architecture that leverages robot embodiment in the attention mechanism, by treating it as a graph of sensors and actuators.
0:17

Can we make Transformers better and more efficient for robot learning? Excited to introduce Body Transformer (BoT), an architecture that leverages robot embodiment in the attention mechanism, by treating it as a graph of sensors and actuators.

Carlo Sferrazza

60,714 views • 2 years ago

Lagos is learning from Aba development.... An Ibo man mocks Lagos state for not catching up to the Gigantic development and investment in Aba after seeing 4 buildings for the first time in Aba
0:16

Lagos is learning from Aba development.... An Ibo man mocks Lagos state for not catching up to the Gigantic development and investment in Aba after seeing 4 buildings for the first time in Aba

Nononsense

43,406 views • 2 months ago