lunarmony

Mixture of Logits was actually already deployed on 100M scale+ datasets at Meta and at LinkedIn (https://arxiv.org/abs/2306.04039 https://arxiv.org/abs/2407.13218 etc.). The crucial departure from traditional embedding/multi-embedding approaches is in learning a query-/item- dependent gating function, which enables MoL to become a universal high-rank approximator (assuming we care about recall@1) even when the input embeddings are low rank.

Researchers have discussed limitations of vector-based retrieval from a rank perspective in various forms for a few years. It's further been shown that better alternative exists; some low-rank approaches can theoretically approximate arbitrary high-rank distribution while permitting MIPS-level efficient inference (see e.g., Retrieval with Learned Similarities, https://arxiv.org/abs/2407.15462). Such solutions are already being used in production at Meta and at LinkedIn.

And the original article got 77 upvotes somehow

why does this require "a nonexclusive, royalty-free, worldwide license"?

ResNet used in all GPTs and RoPE (used in most modern Transformers since LLaMa) are both from China. People really need to read up on literatures :)

Real strength isn't talking about masculinity. It's taking responsibility for your team's actions, and sticking to your principles even when it hurts your profits.

A company spokesperson claimed that the change is intended to “better protect” Google Search against malicious activity, such as bots and spam, and to improve the overall Google Search experience for users. The spokesperson noted that, without JavaScript, many Google Search features won’t work properly and that the quality of search results tends to be degraded.

“Enabling JavaScript allows us to better protect our services and users from bots and evolving forms of abuse and spam,” the spokesperson told TechCrunch, “and to provide the most relevant and up-to-date information.”

Calculators can do math per CEO of Texas Instruments. (h/t @naval)

More seriously, I don't believe the current progress will expand beyond fields where there's clearly defined rewards (eg narrowly defined task of producing some outputs given well-defined inputs).

fwiw from my experience, the really strong new grads have either done relevant projects back at college or can come up with reasonably good answers on the spot (my team does a lot of pure algorithm/modeling/distributed system work so those are actually practical questions).

I do agree leetcode makes those questions significantly less useful

Goodhart's Law: "When a measure becomes a target, it ceases to be a good measure."

if anything, the user number may be inflated by the bots..

One could argue the same applies to the ongoing israel-hamas war too. The ideal social network should build consensus not encourage polarization.

> cobalt was created for public benefit, to protect people from ads and malware pushed by its alternatives

can't say the same for bytedance, which is designed to exploit users with various ads

multi-tier payment structures (subscription + ads + paid placements) are very common for most news industry players..

What makes you believe TikTok's goal is not to trick users into spending their whole lives on the platform?