To me it's obvious that anything related to caching has little to do with the model itself. And it's nice to know from the title that you could only be affected if you used this model.

It took me until "We have remediated the issue" to realize Cloudflare is talking about an issue they have in their infrastructure, up until that point they're not making that clear at all, and seems to point towards something else, unless you of course already are aware of what exact names Cloudflare uses for their products...

Isn't whole CS basically pure math?

No, there's a lot of stuff where formal proofs don't make sense or are highly prohibitive.

The flagship case is AI itself, where we rely on benchmarks, not proofs (how could one prove "is good at seducing people").

Something that could have qualified as a "CS discovery" would have been, for example, a SAT or SMT solver that dwarfs in performance any existing one.

Something that could have qualified as a "physics discovery" would have been a new particle. Anyway, something that would come out of experimental data.

I understand its hyperbole but are there people seriously trying to benchmark how good ai is at seducing people? lol

Relevant XKCD: https://xkcd.com/435/

And all illustrations have the AI look (they're not AI images, but AI-generated SVG or Powerpoint or something)

You'd only get plain unstyled rendering with browser support. And at that level you don't need any complexity from Hugo or Obsidian - you only need a very simple script to convert md to plain html. No config or anything to maintain.

> You'd only get plain unstyled rendering with browser support

What I imagined is browsers rendering the content with headings, quotes, etc. Browsers could have some default styles to render MD files. Those styles could be configurable the same way “Reading mode” is usually configurable.

The plain unstyled rendering of strings of text is, of course, available even now. That's why I'm saying this needs a browser-level support.

> you only need a very simple script to convert md to plain html

I get, that it feels simple compared to other software. Even more, MD was designed specifically to be simple!

But if you think about it, this is still too much complexity to just share some text.

With native browser support, you could run zero commands to build a website.

I think this is mostly important for regular users, not developers. Imagine if creating a Word document would require the same level of complexity as building a web page: I doubt that the format would be so popular as it is nowadays.

I think a better comparison is website builders, and people still paid web devs.

It is just a band-aid.

We noticed that some equations only make sense if mass is greater than what we'd expect from what we can see. So for now we pretend that there's this dark matter that we can't see, and with it all math so far checks out.

But we don't know if that missing mass exists at all, or if some formulas are wrong. So it's definitely a "hack", but it works for now and we don't have anything better. And we do expect to find an explanation at some point.

I think its standing is a bit better than that. We have several independent lines of evidence like patterns in the CMB that aren't just "hmm, we need m to be bigger in our model of a galaxy." And we have a couple of energy scales compatible with standard particle physics where the dark matter might be hiding.

TFA mentions a possible speculative explanation, that perhaps we can see some galaxies that are very low on dark matter.

I don't think the idea is to give agents raw data, but for it to run many queries. Humans also don't look at terabytes of raw logs.

If LLMs aren't an example of trickle down, I don't know what is

Paying for something that has been scraped, trained on hundreds of billions of lines of code (mixed up GPL and other licenses) and being resold back to you is trickle down?

Did they force anyone to focus on reviewing those results instead of other ones? People choose to focus on them for a reason.

I think it's always better to publish preliminary results early (as long as you clearly say so, and they did), and let mathematicians decide which of those are worth focusing on.

They did not do this in a vacuum. But it doesn’t matter much. Benchmarking needs a benchmark.

But distributed over thousands of people, I see it happening. Maybe a bit slower, but still.

> distributed over thousands of people

Can you elaborate what does that look like?

A few people spend all their focus and tokens on one problem, and after some time I expect one of them to get at least closer to a solution. They can either work independently, or even better they team up and create a shared scratchpad for their agents. I don't see limits to scaling this, compared to OpenAI running however many subagents per problem.

And that multiply those few people by number of different problems.