Skyfleet a feed to follow replies, quotes and reposts made to the bluesky team and bsky official accounts //Made at BlueskyFeeds.com
Feed on BlueskyFeeds Stats
- đ Liked by 13 users
- đ Updated over 2 years ago
- âď¸ Provider blueskyfeeds.com
Bsky.Team Likes over time
Like count prediction
The feed Bsky.Team has not gained any likes in the last month.
Feed Preview for Bsky.Team
Sorry for hiding the facts in a detailed blogpost that explained what happened.
1
1
24
i find them sometimes incredibly useful and sometimes maddeningly stupid. i donât like that theyâre managed by those companies and are centralized. i donât know what the word âlikeâ means to you here.
1
0
2
this is different from âsloppyâ llm usage in the sense that the produced object is a mathematical proof. so it doesnât matter if itâs sloppy or not. a proof either checks or it doesnât. if it checks then itâs true, you donât need to trust the llm. youâd still need to verify proof statement ofc
0
0
2
the thing about llms is that they donât get tired and are good at fiddling at a problem when thereâs good feedback from tooling. which proof checkers provide. so i think itâs possible that llms will make proof-based programming more mainstream. humans canât be bothered to do that
1
0
7
well itâs not probably what you would expect.
the part that enables better guarantees and standards about software (to the extent itâs possible at all) is mathematical proofs. but programming with proofs is too difficult for humans to do so this hasnât enjoyed much attention in mainstream coding
2
0
0
a microblogging website isnât one of such systems by definition of what you want it to do
0
0
1
they require guardrails and context to wield them well but they can definitely produce good code.
1
0
0
basically computing is just too flexible. there are approaches that bring more structure into it where possible but each is a genuine innovation or a breakthrough and also takes a while to mature. itâs a developing field. ironically llms can unlock some types of hardening that were tricky before
1
1
4
but even without prerequisites, reading top down doesnât tell you enough about program behavior. itâs too dependent on how state evolves over time. thereâs even mathematical limits on what can be determined by reading the code (even if you had arbitrary amount of time like centuries)
1
1
1
understanding prerequisite systemsâ internal behavior all the way down is not possible because the underlying systems go way too deep. it would be like several lifetimes of learning that get outdated every ten years
1
1
6
software is just too hard for humans to think about. the only way is to break it down into thousands of slices where people specialize in each slice. but the issue with that is that mistakes can compound and nobody sees the whole picture. how would you fix that?
2
1
0
i mean just take your operating system. how well do you know it, to the sand level? youâre building on top of all that. now make it distributed
2
0
4
i feel like watching the android system log in real time for a few seconds aged me by at least several days
austin@aparker.io
you have no clue how many errors your phone or computer is currently experiencing. it is more than 0. Bluesky, currently, is experiencing errors, many of which are completely out of their control. This is the nature of software.
5
3
126
this is like a puzzle to me. i want to understand the train of thought, whatever it is!
2
0
8
but following these standards is actually possible because theyâre always applied locally. with software this doesnât work because itâs an endless tower of abstractions that no human can comprehend fully in principle. and all parts are moving and changing, sometimes in realtime. this wouldnât work.
1
0
1
i mean yeah but one of these is stuff you touch and another is a microblogging site. itâs very different areas with different tradeoffs. if someoneâs life depends on this site not going down, that sounds like an avoidable problem
1
0
7
i mean itâs a balance. thereâs definitely some manual/automated QA being done (in addition to all tests) and some degree of manual control over deploys. but ultimately thereâs just no testing that accurately predicts behavior with million users other than rolling out the change (maybe to a %)
1
0
19
i sort of agree with this but thatâs more of a âlacking observablyâ than âcode qualityâ issue. worth improving maybe
austin@aparker.io
well Iâm biased here but comprehensive tracing would have made it really easy to spot (although I also agree with your point)
1
0
2
my impression from the post that it wouldnât break until it happens at a certain scale which artificial testing doesnât reveal
1
0
1