awitt

I'm not saying LLMs are useless. But the value they have provided so far does not justify covering the country in datacenters and the scale of investment overall (not even close!).

The only justification for that would be "superintelligence," but we don't know if this is even the right way of achieve that.

(Also I suspect the only reason why they are as cheap as they are is because of all the insane amount of money they've been given. They're going to have to increase their prices.)

Firstly, it's not really good enough to say "our employees use it" and therefore it's providing us significant value as a business. It's also not good enough to say "our programmers now write 10x the number of lines of code and therefore that's providing us value" (lines of code have never been a good indicator of output). Significant value comes from new innovations.

Secondly, the scale of investment in AI isn't so that people can use it to generate a powerpoint or a one off python script. The scale of investment is to achieve "superintelligence" (whatever that means). That's the only reason why you would cover a huge percent of the country in datacenters.

The proof that significant value has been provided would be value being passed on to the consumer. For example if AI replaces lawyers you would expect a drop in the cost of legal fees (despite the harm that it also causes to people losing their jobs). Nothing like that has happened yet.

The thing that drives me crazy is that it isn't even clear if AI is providing economic value yet (am I missing something there?). Right now trillions of dollars are being spent on a speculative technology that isn't benefitting anyone right now.

The messaging from AI companies is "we're going to cure cancer" and "you're going to live to be 150 years old" (I don't believe these claims!). The messaging should be "everything will be cheaper" (but this hasn't come true yet!).

Related to this, I've been wondering how much we should actually be depending on RLS for. There are known timing attacks against it [1].

[1] https://www.cs.tau.ac.il/~mad/publications/sigmod2023-rls.pd...

Yes, that's true. And not even ClickHouse considers their MergeTree engine to be an LSM-tree (see https://github.com/ClickHouse/ClickHouse/blob/bfc0260f359a0c...).

That doesn't make it wrong, or a bad architecture. It still takes ideas from LSM-trees, and it has similarities. But it can't be called an LSM-tree.

I actually think you're describing a WAL.

The original purpose of an LSM-tree is to take I/O off the critical path of a write (there are other reasons to use them though, for example reducing space amplification).

I would argue that by definition an LSM-tree buffers committed writes in memory, and that means you need a WAL for recovery.

If you are going to immediately flush the memtable then IO is on the critical path. And if you have fine grained updates you'll end up with lots of small files, which seems like a bad thing. It could be reasonable if you only receive batch updates.

LSM-trees do need a WAL. The entire idea of LSM-trees is that writes are buffered in memory and written out all at once. But a particular write doesn't wait for the memtable to be flushed. For that reason you still need a WAL (there is committed state in memory).

See https://research.facebook.com/publications/optimizing-space-... and https://www.cs.umb.edu/~poneil/lsmtree.pdf.