For 99% of people I don't see the usecase (except for privacy but that ship sailed a decade ago for the aforementioned 99%). If the argument is inference offline - the modern computing experience is basically all done through the browser anyway so I don't buy it.
GPUs for video games where you need low latency makes sense. Nvidia GeForce Now works but not for any serious gaming. But when it comes to LLMs at least, the 100ms latency between you and the Gemini API or whichever provider you use is negligible compared to the inference time.
I'm sure giants like Microsoft would like to add more AI capabilities, and I'm also sure they would like to avoid running them on their own servers.
Another thing is that I wouldn’t expect LLMs to be free forever. One day, CEOs will decide that everyone has become accustomed to them - and that will be the first day of a subscription-based model and the last day of AI companies reporting financial losses.
It’s not. The last “algorithm” release was a random grab bag of code which existed in some of the Twitter repo that might have been tangentially related to recommendations/feed.
Anon, when I was looking through this source dump I saw a huge range of timeouts used in various services, do you know if there's any writeup or explanation as to how the engineering team settled on those values?
True, but on the other hand, it's not exactly surprising if they're redacting details related to spam detection, algorithmic manipulation, foreign interference, and other such adversarial phenomena.
(I'm not necessarily saying that's what's going on. But I do seem to recall that when reddit open-sourced, they deliberately chose to redact info related to vote manipulation/spam detection/etc.)
Do you also believe it plausible that aside from redacted parts, this module has not been changed in any way on production servers for over 2 years (last commit is spring 2023)?
i.e., it's not production code and the guy who said it is and claimed to work at Twitter was lying, as previously pointed out. If only some people would read for comprehension before commenting ...
That's obviously not what "production code" means here. You just said that "Do you also believe it plausible that aside from redacted parts, this module has not been changed in any way on production servers for over 2 years (last commit is spring 2023)?" was a good point, so you know what is meant. And further up in the thread there's "I've always wondered - how can I as a non X engineer be sure that the code on GH is actually deployed on their servers?". I just made the point about ignoring the context and then you do it again. That is indistinguishable from trolling. Feh. No more responses from me.
Seems like a semantic argument over the definition of the term "production code". It's OK for you to use that term differently from anonym29 does. Cheers mate.
Even if this is the actual production code at this very second, it won't match prod for long if they continue this pattern of only dropping an update every two years or so.
Honest question. Would you even dare to say you work at Twitter and then spill the beans on some very public lie or misdirection? It’s trivial to match your writing style between your HN comments and your work emails to identify you. Musk is famously a very petty, bitter, and vindictive person with an easy to bruise ego.
I don’t have any knowledge of the reality inside Twitter but I also have no reason to believe the company would be transparent given the many past controversies, or that any one employee would be able to look at this code which has obvious redactions and say “everything else is definitely 100% prod” and not exactly what GP suggested.
because the person you replied to said they worked (past tense) at twitter, unlike you who says they [currently] work (present tense) at twitter
why would they tell someone not working at twitter anymore to stop working at twitter? and how does that amount to "biased, hypocritical, one-way persecution"
I agree in general it isn't. But in this case Musk claimed that was the point of open-sourcing the algorithm. Transparency on what they are or are not suppressing.
When Tesla "open sourced" their patents, they required companies taking them up on it to, not reciprocally, not copy their "designs". So you get access to their patents in exchange for vague restrictions broader than the patent or copyright system.
Close to the end of the paper they casually drop that they solved all the millennium problems including the Riemann Hypothesis and have Lean proofs for all of them.
I don't know enough physics or mathematics to be able to tell whether any of this holds or if it's o3 generated engagement bait.
On further inspection the author seems to have a bored ape as their X profile pic so I'm going with engagement bait. Pretty high quality bait though.
I haven't played around with hypervisors much but the whole point of k8s is not just isolation but all the primitives the control plane gives you which you don't need to implement. Things like StatefulSet, ReplicaSet, Volumes, HorizontalPodAutoscaler, Service, DNS, ConfigMaps, Secrets, Accounts, Roles, Permissions etc.
Also the container runtime which is containerd by default I believe can be switched out for micro vms like Firecracker (never done this though - not sure how painful it is).
This is really interesting and a thorough write up. Thanks to the author for sharing their work.
Whenever I read about super low-level optimisation, my immediate feeling is that of gratitude to the author for spending so much time shaving off nanoseconds which the entire SE community gets to enjoy.
I wonder how much time humanity has collectively saved simply as a result of how all these optimisations stack up.
Assuming this is going to save someone 10ns per query in the end and I spent 150h on coding and writing this up, we only(?) need around 10^14 queries to make it worth it!
But yes, as new technologies stack up, we achieve exponential growth in throughput in the end, which usually enables new science :)
Honestly sets as trees isn't original. While I was learning about ZFC I came across some lectures[0] by Richard Borcherds which was the seed of insipiration for this project.
Author here. Wasn't expecting to see this on the front page!
I'm really very far from a mathematician and this was a write up of a fun side project. I think the title would be unforgivably misleading in a formal context (if this was a paper claiming any new insights) but really it was a fun side project I wanted to right about. Maybe you read this and learned a little bit about set theory if you had no idea what it was (much like myself).
In general I resent popular science (especially in theoretical physics) which tries to reduce deep and interesting topics to poorly thought out analogies - but again my positioning here is not to educate per se. Or Michio Kaku style orating which assumes string theory a priori and later you have conversations with people who think string theory is established and tested because they watched a 40 minute video of him on YT.
Having said all this I need to get better and giving titles to the things I write - my other post about trying to build AGI in Rust got similar criticism.
I'm under the impression that, at least theoretically, Von Neumann's principles of self-replication, game theory, or optimization in the context of designing neural network structures.
You could think about organizing a neural network with layers or nodes that are indexed by Von Neumann ordinals, where the structure of the network follows the natural progression of ordinals. For example:
Each layer or node in the neural network could correspond to a finite ordinal (such as 0, 1, 2, etc.) or transfinite ordinal (like ωω, ω+1ω+1, etc.). The way the network expands and evolves could follow the ordering and progression inherent in the Von Neumann ordinal system.
This could lead to an architecture where early layers (low ordinals) represent simpler, more basic computations (e.g., feature extraction or basic transformations). Later layers (higher ordinals) could correspond to more complex, abstract processing or deeper, more abstract representations.
But I'm afraid there is no hardware substrate upon which to build such a thing.
GPUs for video games where you need low latency makes sense. Nvidia GeForce Now works but not for any serious gaming. But when it comes to LLMs at least, the 100ms latency between you and the Gemini API or whichever provider you use is negligible compared to the inference time.
What am I missing?