Phillip Haeusler

builder. Staff ML @ Canva. thinking about video. innovation, adventure, community

[Phillip Haeusler @PhillipHaeusler]

@willccbb What still matters are data structures and interfaces.

I spend a great deal of time with AI co-designing the right data formats and APIs.

The implementation is irrelevant. Code generation is the new compiler.

[Phillip Haeusler @PhillipHaeusler]

@giffmana parallel rollouts?

[Phillip Haeusler @PhillipHaeusler]

I’d want to ban social media if I was this bad at it too

[Liberal Party @LiberalAus]

Vote Liberal for cost of living relief

[Phillip Haeusler @PhillipHaeusler]

Humanoids literally run a half marathon in China

AFR this morning - “China’s first robot marathon runners trip, emit smoke, fall apart”

Headline of a loser

[Phillip Haeusler @PhillipHaeusler]

This has direct parallels to the "critical batch size" when training transformers

At a certain point (sooner than you think), your direction is good enough that taking a step is more effective than deriving any more certainty.

[Phillip Haeusler @PhillipHaeusler]

Progress comes from balancing direction (gradients) and execution speed (step frequency).

In practice, for the most effective innovation-focused teams I’ve seen, this balance tends to favor uncomfortably noisy direction and uncomfortably rapid steps.

[John Carmack @ID_AA_Carmack]

I have never seen it expressed exactly like that, but I wholeheartedly endorse it:

Feedback beats planning.

My plea at Meta was “No grand plans, follow the gradient of user value”.

[Phillip Haeusler @PhillipHaeusler]

“So, is the singularity here?”
…
“It’s hard to tell”

[Phillip Haeusler @PhillipHaeusler]

CoT is a mechanism for search. It gives the general scratch-pad for exploring ideas.

[Dimitris Papailiopoulos @DimitrisPapail]

The goal of a CoT is not to be accurate, but to help the model arrive at a correct solution. And to the extent that this happens they are successful.

The fact that CoTs are readable is a happy accident of grounding reasoning to natural language, but that may soon go away.

[Phillip Haeusler @PhillipHaeusler]

@jhk_______ the future can be bright

[Phillip Haeusler @PhillipHaeusler]

Had fun following this

The winner trained a single-step model

Turns out the task is 1:1 - the condition fully defines the image

[Simo Ryu @cloneofsimo]

You don't have to DM me. Just Join this channel:
https://t.co/OX1u1wlZdt
$1k in cash
$2k in fal credit.
+ lot of fun in the way!!

[Phillip Haeusler @PhillipHaeusler]

@TXhunyuan neat! reminds me of “an image is worth 32 tokens”

[Phillip Haeusler @PhillipHaeusler]

I get the feeling household robots will be a way larger jump in quality of life than we expect

[Phillip Haeusler @PhillipHaeusler]

@vikhyatk yep

[Phillip Haeusler @PhillipHaeusler]

https://t.co/1h83I7MiXC

[Phillip Haeusler @PhillipHaeusler]

I found muon fascinating & explored ways to parameterize orthogonal linear layers in torch

[Phillip Haeusler @PhillipHaeusler]

~50% fewer params, better stability

[Phillip Haeusler @PhillipHaeusler]

@jhk_______ Absolutely

[Phillip Haeusler @PhillipHaeusler]

@jam3scampbell yeah

AI will change what it means to be human

[Phillip Haeusler @PhillipHaeusler]

@PaulineHansonOz good idea

[Phillip Haeusler @PhillipHaeusler]

In the long run, you really don't want to forget how to build things. That's why tariffs matter. Remember when we made cars?

[Phillip Haeusler @PhillipHaeusler]

Australia's largest glass manufacturer, Oceania Glass, has collapsed - citing impossible competition from cheap Chinese imports.

This comes during a housing crisis.