That reminds me of a joke.

A museum guide is talking to a group about the dinosaur fossils on exhibit.
“This one,” he says, “Is 6 million and 2 years old.”
“Wow,” says a patron, “How do you know the age so accurately?”
“Well,” says the guide, “It was 6 million years old when I started here 2 years ago.”

From Wikipedia: this is only a 1-sigma result compared to theory using lattice calculations. It would have been 5.1-sigma if the calculation method had not been improved.
Many calculations in the standard model are mathematically intractable with current methods, so improving approximate solutions is not trivial and not surprising that we’ve found improvements.

This seems like more of an achievement for the Barbie brand than for the individual director.

Apparently Inflection AI have bought 22,000 H100 GPUs. The H100 has approximately 4x the compute for transformers as the A100. GPT4 is rumored to be 10x larger than GPT3. GPT3 takes approximately 34 days to train on 1024 A100 GPUs.

So with 22,000*4/1024=85.9375x more compute, they could easily do 10x GPT4 size in 1-2 months. Getting to 100x the size would be feasible but likely they’re banking on the claimed speedup of 3x from FlashAttention-2, which would result in about 6 months of training.

It’s crazy that these scales and timelines seem plausible.

This is an essay about the Barbie brand and its relationship to feminism and capitalism through history and the modern day. The Barbie movie is discussed but it’s not the primary focus.

NGC 1277 is unusual among galaxies because it has had little interaction with other surrounding galaxies.

I wonder if interactions between galaxies somehow converts regular matter to dark matter.

Oh certainly, that series took quite a risk on writing style and it’s quite divisive.
If you enjoy fantasy, you could try her other series as an alternative. The Inheritance Trilogy is a more standard writing style.

I almost put The Fifth Season down after the first chapter, I remember thinking: “This author has a chip on their shoulder”. I’m glad I persevered though, and I definitely recommend the series to people as it is quite different. I’d suggest giving it another shot.

I might try jumping in again on season 2, thanks.

Claude 2 would have a much better chance at this because of the longer context window.
Though there are plenty of alternate/theorised/critiqued endings for Game of Thrones online, so current chatbots should have a better shot at doing a good job vs other writers who haven’t finished their series in over a decade.

As a counterpoint to other comments here, I didn’t like Babylon 5. I gave up in the first season on the episode about religions, where each alien race shows a single religion but then humanity shows an enormous number of them.

Showing planets in sci fi as homogenous is a common trope, but such a simplistic take. This resonated poorly with me as I felt the aliens all behaved exactly like humans as well, to the point where you have stand-ins for Jehovah’s witnesses. That episode cemented for me the feeling I had when watching. Babylon 5 is racist against aliens.

This looks amazing, if true. The paper is claiming state of the art across literally every metric. Even in their ablation study the model outperforms all others.

I’m a bit suspicious that they don’t extend their perplexity numbers to the 13B model, or provide the hyper parameters, but they reference it in text and in their scaling table.

Code will be released in a week https://github.com/microsoft/unilm/tree/master/retnet

Why do you say they have no representation? There are a lot of specific bodies operating in the government, advisory and otherwise, with the sole focus of indigenous affairs. And of course, currently, indigenous Australians are over represented in terms of parliamentarian race (more than 4% if parliamentarians are of indigenous descent).

While in general, I’d agree, look at the damage a single false paper on vaccination had. There were a lot of follow up studies showing that the paper is wrong, and yet we still have an antivax movement going on.

Clearly, scientists need to be able to publish without fear of reprisal. But to have no recourse when damage is done by a person acting in bad faith is also a problem.

Though I’d argue we have the same issue with the media, where they need to be able to operate freely, but are able to cause a lot of harm.

Perhaps there could be some set of rules which absolve scientists of legal liability. And hopefully those rules are what would ordinarily be followed anyway, and this be no burden to your average researcher.

From the study:

The tasks demanded clear, persuasive, relatively generic writing, which are arguably ChatGPT’s central strengths. They did not require context-specific knowledge or precise factual accuracy.

And:

We required short tasks that could be explicitly described for and performed by a range of anonymous workers online

The graphs also show greater improvement for the lowest performers than for the high performers.

Definitely an encouraging result, but in line with anecdotes that, currently, LLMs are only useful for genetic and low complexity tasks, and are most helpful for low performers.

See this comment on another thread about this for some more details.

Taking 89.3% men from your source at face value, and selecting 12 people at random, that gives a 12.2% chance (1 in 8) that the company of that size would be all male.
Add in network effects, risk tolerance for startups, and the hiring practices of larger companies, and that number likely gets even larger.

What’s the p-value for a news story? Unless this is some trend from other companies run by Musk, there doesn’t seem to be anything newsworthy here.

So, taking the average bicep volume as 1000cm3, this muscle could: exert 1 tonne of force, contact 8% (1.6cm for a 20cm long bicep), and require 400kV and must be above 29 degrees Celcius.

Maybe someone with access to the paper can double check the math and get the conversion efficiency from electrical to mechanical.

I expect there’s a good trade-off to be made to lower the force but increase the contraction and lower the voltage. Possibly some kind of ratcheting mechanism with tiny cells could be used to overcome the crazy high voltage requirement.

GPT-4 was fine-tuned on English and Chinese instruction examples only (source). There’s clearly some western bias in the historic events, but it would have been interesting to also discuss if there was a bias towards Chinese events as well. And, if so, what other languages or prompts may elicit that bias.
As an example, could you get the model to have an English bias with “I’m from America…” and a Chinese bias with “I’m from China…” even when using English?

DALL-E was the first development which shocked me. AlphaGo was very impressive on a technical level, and much earlier than anticipated, but it didn’t feel different.
GANs existed, but they never seemed to have the creativity, nor understanding of prompts, which was demonstrated by DALL-E. Of all things, the image of an avocado-themed chair is still baked into my mind. I remember being gobsmacked by the imagery, and when I’d recovered from that, just how “simple” the step from what we had before to DALL-E was.
The other thing which surprised me was the step from image diffusion models to 3D and video. We certainly haven’t gotten anywhere near the quality in those domains yet, but they felt so far from the image domain that we’d need some major revolution in the way we approached the problem. The thing which surprised me the most was just how fast the transition from images to video happened.