Xorlev
The prefill is the only blocking part, and you can prefill the whole context up to the point where they diverge, then prefill each question and decode the one token in parallel for each question.
If you batch vLLM calls with the same prompt prefix to the same process, it'll deduplicate the prompt prefix across batched requests (+/- the block size) and decode in parallel for each.
That's with a vanilla LLM. If you modify the LLM you can pull that in-graph, but it isn't really necessary.
This might be confusing the lack of need for validation with the lack of need for other people. Sure, taking confidence from your partner is wonderful but it's not "seeking validation" to maintain other relationships.
Putting everything on one person can quickly become codependency and enmeshment. At some level some codependency/enmeshment is inevitable ("healthy interdependence") when you spend your time with one person, however it can also be very unhealthy.
You can lose your own identity, and end up putting all your needs on the other person. That makes conflict difficult, distance difficult, and you lose your support network.
I think Friday demos are really cute, and a healthy relationship can certainly touch on all areas, but it's important to invest in both other relationships (friends/family/partners) AND yourself. Investing in time with yourself means investing in your hobbies, doing things just for you and maintaining that individual identity.
GitHub was, once upon a time, quite stable. Things have changed: more features, more usage, and automated agents.
When accounting for prefix caching, this greatly accelerates each turn. Barring large file reads, prefill still isn't the bottleneck vs. decoding reasoning tokens. Script-writing too.
This is especially true during exploration phases when traversing through directory trees and grepping files, you're talking about a few hundred tokens/turn.
Now both exist and get roughly the same feature set now, but the Java version remains popular given the vast variety of mods and servers.
It was also invite-only for a loooong time.
I understand the strategic value of offering unlimited features to differentiate from competitors like Slack, might drive some amount of anxiety. Buyers may question long-term sustainability or fear undisclosed "shadow" caps.
Since engineering limits are inevitable to prevent abuse (especially on free accounts), it might be better to set specific, generous expectations upfront. For example, 2 years of freeform search plus unlimited "tagged" (i.e. Decision Inbox) search. This avoids the skepticism that comes with promising "no limits" forever. It also avoids the trap of needing to announce a change later with predictably negative reactions.
If you do want to offer unlimited, then planning ahead with hard-to-hit-unless-you're-trying messages/hr limits might help you tame growth and avoid abuse. My initial thought when seeing unlimited anything is "I could write a filesystem on top of that" - especially if you allow attachments. :P
That said, interpreting user feedback is a multi-role job. PMs, UX, and Eng should be doing so. Everyone has their strengths.
One of the most interesting things I've had a chance to be a part of is watching UX studies. They take a mock (or an alpha version) and put it in front of an external volunteer and let them work through it. Usually PM, UX, and Eng are watching the stream and taking notes.
Logs are always generated, and logs include some amount of data about the user, if only environmental.
It's quite plausible that the spellchecker does not store your actual user data, but information about the request, or error logging includes more UGC than intended.
Note: I don't have any insider knowledge about their spellcheck API, but I've worked on similar systems which have similar language for little more than basic request logging.
This is a long-standing issue with zvols which affects overall system stability, and has no real solution as of yet.
It's also not as easy as GB/s/core, since cores aren't entirely uniform, and data access may be across core complexes.
Just saying it doesn't make it true. Twitter has a long way to go to regain advertiser confidence, and Blue is barely a footnote.
Using something leaked always carries some inherent risk.
That's worth explaining: it's automated code deletion, but the owner of the code (a committer to that directory hierarchy) must approve it, so it's rare there's ever a false deletion.
I don't know how many strange issues I've tracked down that amounted to "this protocol buffer has a uint32 field, and surprise now the value is negative in Java and oops there was a check that cared about that." At least five or six issues.
At least when it comes to serialization, enforce invariants above the serialization layer.
I would love an unsigned byte type in Java though. What a pain.
"Can I use AL for ArrayList?" "Go for it"
If knowing you use a crate is too much, then running your own registry with a mirror of packages seems like all you could do.
That's not even 100% true either [1]. For instance, you can flash Android devices from the browser. The example in [2] talks to an attached Arduino device.
> We are heading to the stars, and we will bring life with us.
Another way to think about it: we'll be Earth's best, most successful, furthest-flung seeds.
I write C++ at Google, and it encourages use of modern C++ features, and many things you see adopted in std have roots in our libraries.
I'm curious what you think Google prevents us from using and why you think our C++ is like 90s Java.
https://abseil.io/tips has a lot of our philosophies and abseil is chunks of our internal libraries published externally.