Mugur Marculescu

Graph weaver at Magnitude Labs. Prev Google PM on GRPC, API Console, VR/AR.

[Mugur Marculescu @mugurm]

New paper: Labels as Computational Primitives.

The claim: a neuron's label isn't a comment. It's a compressed behavioral spec implying the unit's decay, threshold, bias, and connectivity.

Which means a working neural network can be compiled from language instead of trained.

https://t.co/q0AwrulaoA

[Mugur Marculescu @mugurm]

@ylecun @firesidealpha yes, but state transitions exist as well and so does computational irreducibility therefore unpredictability remains. 🤔

[Mugur Marculescu @mugurm]

@auradeluxe turns out you can fit the internet on a cd rom drive

[Mugur Marculescu @mugurm]

@GaryMarcus time to get some new executives

[Mugur Marculescu @mugurm]

I’ll just say it… computation requires an observer.

[Mugur Marculescu @mugurm]

@NoahChrein That is what’s happening:

[Mugur Marculescu @mugurm]

Today the conversation in AI is shifting from loops to graphs... The next two hops in that journey are computational irreducibility and then consciousness.

[Mugur Marculescu @mugurm]

@Plinz Here’s the question:

[Mugur Marculescu @mugurm]

you cannot have a medium-less process, so then what would you say is the medium of (pure) computation?

[Mugur Marculescu @mugurm]

@SubjectsPro best guess is a function boundary… which means the temporal window enough to self-replicate. and hierarchical nesting is at the very least allowed.

[Mugur Marculescu @mugurm]

@anilkseth @blaiseaguera @TheEconomist thought experiment:

[Mugur Marculescu @mugurm]

you cannot have a medium-less process, so then what would you say is the medium of (pure) computation?

[Mugur Marculescu @mugurm]

you cannot have a medium-less process, so then what would you say is the medium of (pure) computation?

[Mugur Marculescu @mugurm]

@blaiseaguera @TheEconomist I think so too. I imagine a mechanism like this:

[Mugur Marculescu @mugurm]

From uniformity to complexity simply through recursive polarization.

[Mugur Marculescu @mugurm]

@fchollet surprise = prediction Δ outcome

[Mugur Marculescu @mugurm]

@Plinz @skdh yes!

[Mugur Marculescu @mugurm]

@magicseth Grounding is relative... it's turtles all the way down to the level you care about... and both AI and humans live on earth (for now)... so it's all good if you define a practical reference frame.

[Mugur Marculescu @mugurm]

@trq212 can you also watermark me as the author of my content?

[Mugur Marculescu @mugurm]

@alexwg ha. made me think of my reaction to dynamic workflows.

[Mugur Marculescu @mugurm]

tokens are just FLOPs disguised as a billing mechanism, which are joules disguised as FLOPs. dynamic workflows are the correct direction but the substrate is not.

[Mugur Marculescu @mugurm]

@sserrano44 i feel the same way

[Mugur Marculescu @mugurm]

@allTheYud They were all on the same wavelength because they all likely shared the starting conditions.

[Mugur Marculescu @mugurm]

Just because you can create LLMs to be fully deterministic doesn’t mean you can bypass computational irreducibility during inference. Which means no amount of validation can make you fully trust output on an unbound time scale.

[Mugur Marculescu @mugurm]

LLMs are neurosymbolic but they are unlabeled. It doesn’t have to stay that way.

[Mugur Marculescu @mugurm]

New paper: Labels as Computational Primitives.

The claim: a neuron's label isn't a comment. It's a compressed behavioral spec implying the unit's decay, threshold, bias, and connectivity.

Which means a working neural network can be compiled from language instead of trained.

https://t.co/q0AwrulaoA

[Mugur Marculescu @mugurm]

@dileeplearning True, but I think it's even deeper:
https://t.co/oLmZtp0zPu

[Mugur Marculescu @mugurm]

New paper: Labels as Computational Primitives.

The claim: a neuron's label isn't a comment. It's a compressed behavioral spec implying the unit's decay, threshold, bias, and connectivity.

Which means a working neural network can be compiled from language instead of trained.

https://t.co/q0AwrulaoA

[Mugur Marculescu @mugurm]

graphs = networks, networks = graphs so neural networks are neural graphs, just not as flexible as we normally would assume when we would use the term graph.

[Mugur Marculescu @mugurm]

@phillip_isola @gabriberton same.

[Mugur Marculescu @mugurm]

New paper: Labels as Computational Primitives.

The claim: a neuron's label isn't a comment. It's a compressed behavioral spec implying the unit's decay, threshold, bias, and connectivity.

Which means a working neural network can be compiled from language instead of trained.

https://t.co/q0AwrulaoA