Soham Mehra

Co-Founder & CTO @morphlabs_ai • Building non-invasive BCI for prosthetics • NeuroAI • AI/ML

[François Chollet @fchollet]

The hardest problems are rarely solved by adding more complexity to the solution -- they are solved by reframing the question until a simpler, clearer answer reveals itself.

[Andrej Karpathy @karpathy]

This is the the quote I've been citing a lot recently.

[Soham Mehra @sohammehra_]

It's cute when some random people you just met think that their feedback is going to deter you from building the company you've been working on for 1.5 years

[Soham Mehra @sohammehra_]

@haider1 Paper: https://t.co/Hh15tKtIo4

[Soham Mehra @sohammehra_]

@haider1 There's a field of NeuroAI focused on how to create the neural datasets that might be able to better train models on metacognition, emotional intelligence, and take steps towards continual learning. Definitely an interesting problem

[Soham Mehra @sohammehra_]

@AndrewLampinen The idea of modeling positive forward transfer seems like it would involve an extension on prospective learning, but backwards transfer feels like it would be a different problem: revising existing knowledge. Not sure if KV cache and document retrieval get us there

[Soham Mehra @sohammehra_]

Anyone else headed to #IntefaceNeuro hosted by @gt_neuro and @RiceNeuro ? Looking forward to showcasing some work and meeting others there

[Soham Mehra @sohammehra_]

@sama I wonder what Nvidia makes for every additional user using Codex.

More people use codex -> more inference compute demand -> OpenAI purchases more GPUs

How much does the additional flow of income offset the spend on tokens?

[Soham Mehra @sohammehra_]

@Sachin_and_Adam @WoodingtonBen @Elise__Jenkins @coherenceneuro BCIs for cancer treatment is a use case most people don't even know exists yet. The fact that they're already at human trials with SOMA-1 says a lot about how fast the field is moving. Excited to see where this goes!

[Soham Mehra @sohammehra_]

@threeaus @AmandaAskell It says something about the state of the industry, when every lab has a full time philosopher

[Soham Mehra @sohammehra_]

@anne_churchland Even within humans this gets complicated fast. Building EEG motor decoders, we see enough connectivity variability between subjects that generalization is a real challenge. Architectures are getting closer, but figuring out what's conserved vs individual is part of the problem

[Soham Mehra @sohammehra_]

Motor decoding accuracy climbing 1-2% a day from surgical changes to our graph learning pipeline. How you define electrode connectivity matters as much as the architecture itself. Also a bonus when those graph structures start revealing something interpretable about the subject.

[Soham Mehra @sohammehra_]

@aakashgupta It's interesting given how much praise was given for Thinking Machine Labs's talent retention when Zuck first started poaching. Does everyone really have a price, or is there something else going on behind the scenes?

[Soham Mehra @sohammehra_]

microtutor now has a knowledge graph to show what you've learned in a course, and how the topics your learning relate to one another

Coming soon: A knowledge graph that shows how everything you're learning relates and a TA that works through interactive exercises with you

https://t.co/RF2Hknl85I

[Soham Mehra @sohammehra_]

@intuitiveml Very cool to see this. It comes down to designing a system to compress the bottlenecks that aren't dependent on AI so that you can iterate much quicker. Curious how deeptech / hardware companies are adapting systems like this to compress their iteration cycles

[Soham Mehra @sohammehra_]

@s_y_chung @natmesanash @_TheTransmitter Great read! Interesting that different regions show qualitatively different geometric strategies. Is that a property of the region itself or of where it sits in the processing hierarchy?

[Soham Mehra @sohammehra_]

Hitting metrics is nice but there’s a beauty to the people you’re building technology for getting excited about what they can do with it

[Soham Mehra @sohammehra_]

You would never be able to tell from just the content whether these were gen z college students or some of the most famous ppl in ML

[Soham Mehra @sohammehra_]

Seeing the difference between X freaking about about Mythos compared to what's happening on Linkedin makes you realize how little so many people still know about these model

[Harry Stebbings @HarryStebbings]

Why building deep tech out of the echo chamber of the Valley is a massive advantage.

"Being outside the Valley means you are less plugged into trends, gossip and the latest hype.

But it allows you to think more deeply and be more original in your approach.

That is a big advantage in deep tech, where long-term thinking matters more than following fads." @demishassabis

To what extent does being in the valley help vs harm your ability to think differently but innovatively @PalmerLuckey @garrytan @t_blom @paulg @rabois