dbredesen

ML, data, truth seeker, world traveler. Informing people through data and combating the spread of misinformation. Always open to evidence-based debate.

[dbredesen @dbredesen]

@jun_song Is the max RAM expected to be 512? 1024?

[dbredesen @dbredesen]

It’s more than just a brute force robotic workforce.

Compound physical AI with recursive self improvement of materials, hardware, logistics, and processes.

Mark my words, we will literally see skylines transform in front of our eyes in the not-too-distant future.

[David @DavidSHolz]

5 million humanoid robots working 24/7 can build Manhattan in ~6 months. now just imagine what the world looks like when we have 10 billion of them by 2045. now imagine the year 2100.

[dbredesen @dbredesen]

@ihtanboga The biggest limitation of MJ’s scanner is that ultrasound cannot penetrate bone or air; so no diagnostics on lungs, brain, marrow, parts of the GI tract.

I am a huge fan of more frequent, inexpensive diagnostics but people need to be careful branding this as an MRI replacement.

[dbredesen @dbredesen]

@HealthcareLdr @midjourney Traditional ultrasound cannot image inside bones or air-filled lungs, so unless this somehow gets around those limitations, the claims that this will replace MRI or CT seem overstated.

[dbredesen @dbredesen]

@bruhlencia9a @PolymarketMoney Huge missed opportunity if they don’t cast Matthew Rhys in the biopic.

[dbredesen @dbredesen]

@engineers_feed Transistors belong on this list, close to the top.

[dbredesen @dbredesen]

@chakravartiiin Elite job title

[dbredesen @dbredesen]

@cremieuxrecueil @moreisdifferent Can you post some of the CoT?

[dbredesen @dbredesen]

@joinshiftX shift

[dbredesen @dbredesen]

@AURELIANbuilds @Scobleizer @karpathy It finally works!

[dbredesen @dbredesen]

@KaiXCreator DeepMind, Isomorphic, ElevenLabs, AMI Labs (Yann LeCun’s new company), etc.

[dbredesen @dbredesen]

@ecommerceshares Demis Hassabis and DeepMind would like to have a word…

[dbredesen @dbredesen]

@gavinrbrown1 @alkimiadev If you don’t embrace it, you’re gonna face an uphill battle.

[dbredesen @dbredesen]

@idavidrein @threadreaderapp unroll

[dbredesen @dbredesen]

@haridigresses @viiskb Small, clean codebase. Built orchestration on a fresh context so the complex model code and notebooks wouldn’t bias the context.

5.5 low just phoned it in. Didn’t extract string constants. Ignored common patterns.

I’d have a talk with an intern if they did this implementation.

[dbredesen @dbredesen]

@haridigresses @viiskb Used 5.4 high/xhigh to build a very capable time series correlation model.

Then used 5.5 low/med to build a simple orchestrator.

1. It literally built a _debug() method that just checked debug=True then print()s a string
2. It ignored my request to comment the code

Amateur.

[dbredesen @dbredesen]

@haridigresses @viiskb I’ve been using 5.5 on Low, and it has missed completely amateur things. I’ve never had to do so much damage control with a coding agent.

I wonder why people are getting such vastly different results…

[dbredesen @dbredesen]

@nntaleb I was looking for ReLU as an answer 🙃

But yeah it’s (D)

[dbredesen @dbredesen]

@celestepoasts @vibecoding_era @amilabs is poised to change that

[dbredesen @dbredesen]

TIL that Claude is a sanctimonious Bayesian 👀