This is critical. The previous framework tried to regulate every model trained with higher than 10^23 FLOPs. It was absurd then, no reason to think 10^26 is now the magic number.

It's always *just* a bit higher than where we are today. Imagine if we had done this!!

What's telling about the AI regulation push is that nobody, not a single soul, has a credible defense of the 10^26 FLOPs threshold, only 'its slightly higher than today' or 'its only a few extra $m'

Just imagine if we regulated other things this way, pollution, or malware.

These models have no understanding of what they're reading and saying -- with correspondingly unreliable, sometimes comical results. Yet, they're optimized to sound as knowledgeable and authoritative as possible. Not a good combination.

Random close packing of binary hard spheres predicts the stability of atomic nuclei: arxiv.org/abs/2405.11268

Statically Typed Functional Programming with Python 3.12

wickstrom.tech/2024-05-23-sta…

There is only one way to fix the California anti-AI bill:

Shift the burden of a crime back to the person or people committing a crime.

What this bill effectively does is put model makers and trainers on the hook for life, facing fines and even jail time, for any possible misuse

This article presents a very cool idea. How do you design networks that could generalize to data living in dimensions it was not trained on? In equivariant networks it might be feasible by using the idea of representation stability. arxiv.org/abs/2306.06327

Stuff like this is more dangerous and likely to happen than someone using an LLM to build out a basement bioweapon lab.

Actual AI safety is this type of issue. Not fantasies about imminent superintelligence

It feels like this sort of thing is a major issue waiting to happen. 10 year olds use Google all the time. What happens if a child accidentally poisons themselves?

The biggest danger from GenAI is that many people will believe it is actually intelligent and start trusting it too much, while in reality, it is a giant interpolative database with no advanced reasoning skills or common sense.

Important misconception worth clearing up:

'Senate passed our bill' != law

This only becomes law if the CA Assembly votes for it in August

So, the next 2 months are critical for little tech and open source devs to get our shit together

Sucks to be here. But not too late

If I wanted to make a Unitree robot in the USA with zero imported goods, what would be the supply chain bottleneck?

This is really exciting. I've been writing about the dangers of China's navy buildup for a bit (deepforest.substack.com/p/focusing-on-…, deepforest.substack.com/p/how-to-build…, deepforest.substack.com/p/a-welcome-to…). We need more shiptech in silicon valley!

An atrocious bill where sentators failed to listen to their constitutients, and have singled handedly worked to crush the very source of California's rise to prominence as the center of technical excellence.

Imagine if we'd created an arbitrary line in the sand for 4K of RAM or

This is really disappointing. Regulatory capture by bad actors successfully executed

Good article. Has anyone studied what would happen in a serious crisis with China if the CCP just turns off all exports to the US? We saw a bit of this with masks in the pandemic. Most US policy is focused on high tech (semiconductors and EV), but what about base essentials?

Check out our new result on quantum simulation of LGTs! arxiv.org/abs/2405.10416 We show exponential improvement in fermion color, leading to up to 25 orders of magnitude improvement in computational volume. This brings LGT simulation to competitive ranges with quantum chemistry.

The #USNavy Can’t Build Ships - FP

Decades of deindustrialization and downsizing have left America without enough shipyards to build and maintain a fleet.

Deindustrialization, downsizing and an obsession with military presence in the Middle East have left the United States

As an off the cuff thought, maybe we as a community need to move away from static benchmark datasets to benchmark distributions. That is, the benchmark is a piece of code that randomly samples a datapoint and label from a distribution (assuming automatic labeling is available).