WhitneyLand

Great work. How many GPU hours to train?

How rare is this?

G4 storms are ~100 per solar cycle (~11 years).

So roughly 9 G4 events/year on average.

Left this on his blog but it’s awaiting moderation:

It would be helpful to have more clearly targeted and titrated criticism, because you’ve mentioned press releases, a sciam article, the paper, and Sabine all without differentiation.

I hope it’s clear enough the paper itself is legit and doesn’t seem to make any inappropriate claims. Beyond that, the PRs seem to be the real offenders here, the sciam article less so (could be argued that’s healthy popsci), and I’m not sure what comment you’re making about Sabine. The title of her video may be click baity but the content itself seems to appropriately demarcate string theory from the paper.

1. I tried a couple OpenAI models under a paid account with no issue:

“In U.S. presidential elections, you’re eligible to vote if you meet all of these…” goes on to list all criteria.

2. No issue found with Gemini or Claude either.

3. I tried to search for this issue online as you suggested and haven’t been able to find anything.

Not seeing any evidence this is currently a real issue.

That doesn’t sound right. What model treats this as a controversial question?

"who is eligible to vote in US presidential elections"

PNG as a format is not lossy, it uses deflate.

What you’re referring to is pngquant which uses dithering/reduces colors to allow the PNG to compress to a smaller size.

So the “loss” is happening independent of the format.

Yes, this was exactly my point.

Yeah, I’d be wary of hiring the developer even if he’s in the right.

“It's really not that hard”

I’m going to guess your job does not involve much UX design?

Why not use these as a benchmark for LLM ability to make breakthrough discoveries?

For example prompt the 1913 model to try and “Invent a new theory of gravity that doesn’t conflict with special relativity”

Would it be able to eventually get to GR? If not, could finding out why not illuminate important weaknesses.

It can be difficult to explain to management why in certain scenarios AI can seem to work coding miracles, but this still doesn’t mean it’s going always speed up development 10x especially for an established code base.

Tangible examples like this seem like a useful way to show some of the limitations.

Shouldn’t there be a lot of skepticism here?

All the problems they claim to have solved are on are the Internet and they explicitly say they crawled them. They do not mention doing any benchmark decontamination or excluding 2024/2025 competition problems from training.

IIRC correctly OpenAI/Google did not have access to the 2025 problems before testing their experimental math models.

Where did you read that? It’s not true.

- Originally she would have got 8 mil if the transaction closed

- She had to give that up due to bad press

- 10% would have been 2.5B so no way was that ever happening

- Her big score was the golden parachute tied to leaving the company. You could say the merger accelerated her exit indirectly, but there was no commission on the deal.

I think we may not be disagreeing much…

Agree fMRI doesn’t tell us for sure what information is in electrical activity. I only mean to imply it’s suggestive.

But we still can get a lot of information from electrodes. Do words count as thoughts? I’d say so.

Definitely don’t think electrical signals are all there is to thought.

They’re definitely part of it, but I think time will show unlike with computer parts, it’s hard to make clean separation of responsibilities in the brain.

Exactly, if everything looked too good to be true and there was no transparency or hint of a business model it’s actually less attractive for some who value predictability.

You say you don’t believe something is true and then say you don’t know, but I’ll disagree with “electrical signals don’t “define” (encode) thoughts.

To be clear, of course it’s true that our thoughts are more than just electrical activity. The brain is a system. However, it seems clear that thoughts are at least partially encoded in electrical activity.

What you mentioned those startups will find fruitless, that’s already been done for years in a research setting. It may not be a successful business model, but it’s already been demonstrated.

There are fMRI studies and electrical measurement studies. You could argue fMRI decoding of images is not electrical activity which is true, but a bunch or work shows they are strongly correlated.

For electrical activity alone we’re already decoding information like words, so it’s hard to claim electrical activity doesn’t define thoughts.

Maybe you mean to say, doesn’t define all the content of our thoughts which is a much different claim.

>>It has rendered Maya nearly obsolete

Hadn’t heard that. How many AAA vfx studios have left Maya for Blender?

>>benchmarks are meaningless

No they’re not. Maybe you mean to say they don’t tell the whole story or have their limitations, which has always been the case.

>>my fairly basic python benchmark

I suspect your definition of “basic” may not be consensus. Gpt-5 thinking is a strong model for basic coding and it’d be interesting to see a simple python task it reliably fails at.

For some people having a child fundamentally changes their definition of love and their belief in how it’s possible to relate to other human beings.

This is mostly disconnected from true/false values. The facts haven’t really changed. Yet it can be so powerful and rewarding few would trade it regardless of the risks/pain/hardships that can come with it.

Experiences can be profound and change perspectives in a way that’s so rewarding and wholistic it’s really impossible to describe.

Comparing such experiences to math doesn’t really work, apples and oranges.

For point to point sure, but if you want to connect multiple machines in an actual fabric you’ll need some kind of network interop.

The Asus clustering speed is not limited to p2p.

Made the correction to 80Gb/sec thank you.

W.r.t ip, the fastest I’m aware of is 25Gb/s via TB5 adapters like from Sonnet.

GX10 vs MacBook Pro M4 Max:

- Price: $3k / $5k

- Memory: same (128GB)

- Memory bandwidth: ~273GB/s / 546GB/sec

- SSD: same (1 TB)

- GPU advantage: ~5x-10x depending on memory bottleneck

- Network: same 10Gbe (via TB)

- Direct cluster: 200Gb / 80Gb

- Portable: No / Yes

- Free Mac included: No / Yes

- Free monitor: No / Yes

- Linux out of the box: Yes / No

- CUDA Dev environment: Yes : No

Simple explanation, they used AI as a voice for their writing instead of using it as a tool for writing in their own voice.

LLMs are good to proofread, check your tone, generate ideas, etc.

Letting them take over your connection with an audience or be a substitute for gut checks or taste is not helping anyone.

As the old saying goes the graveyards are full of irreplaceable men.

But even if we grant that only one person could deliver a solution, it wouldn’t change the fact that you’re giving up on certain things to get it.

Having a solid product that solves a problem well can be orthogonal to how well a codebase lends itself to readability, learning curve, and efficiently ramping up new developers on a project.

Just because you succeed at one says nothing about other practical and important metrics.

It looks cool, nice project.

Recommend taking a few minutes to make the web page work on mobile.

For example on iPhone the actual webgpu/ray tracing seems to work fine but html formatting is shoving things over to be barely visible.

tldr: Seeing what happens internally in an LLM lets you reconstruct the original prompt exactly.

Maybe not surprising if you logged all internal activity, but it can be done from only a single snapshot of hidden activations from the standard forward pass.

Merely disagreeing with a guy like Rovelli on physics feels like hubris. :)

But agree with you in this case. Animals perceive the flow of time because we have memory and prediction abilities. This gives us a psychological arrow that aligns with the thermodynamic arrow.

This is what’s meant by no objective reality as alluded to in the article.

https://en.wikipedia.org/wiki/Relational_quantum_mechanics

The claim is not that objects don’t exist, just that they don’t have observer independent properties.

>>no objective physical reality is…the best conclusion that one can draw given strict adherence to what the mathematical formalism of QM provides…

Can’t get on board with that. Relational QM/no objective reality is just one viewpoint, and it’s worth noting it’s not consensus.

To clarify he’s not claiming objects don’t exist, just that they don’t have observer independent properties.

It is fun to try and wrap your head around what no objective reality would mean. To grasp what we already have shown to be true about time being relative, the examples around simultaneity are a great wtf demonstration: https://en.wikipedia.org/wiki/Relativity_of_simultaneity