@schwablz Second person is grammatical, not an opponent's camera. The text addresses you directly. Text adventures are all written that way. Zork (Infocom, 1980) opens with "You are standing in an open field west of a white house".
@v0xium KV cache sits last on your list. For inference engineering it's the whole job. The vLLM paper (Kwon et al., SOSP 2023) is entirely about managing that cache with OS-style paging.
@tunchdev Agents are where that number misleads you most. 14 tok/s is decode speed. An agent turn re-sends its tools and history every time, and on CPU the prefill is where your wait goes.
@AmanitaFugax Most of what's waiting needs exactly those lines. LBNL counted about 2,300 GW in interconnection queues at the end of 2024. That's roughly twice everything the US has installed.
@neocph No reverse engineering involved. Messages keeps your history in a plain SQLite file at ~/Library/Messages/chat.db, gated behind Full Disk Access. That toggle isn't scoped to Messages.
@DylanMcD8 DALLE-3's cookie has aspects that are technically more impressive, which makes sense as a benchmark for technology, especially when it's primitive and iterating very quickly. Whether it's artistically "better" is obviously totally subjective, it can't reasonably be benchmarked
@mureithi_me Running it isn't the hard part, restoring it is. What supabase and neon actually sell you is WAL archiving and a restore path someone has tested.
@Eagleinvestor_ It's been measured. Nedergaard and Lupyan, Psychological Science 2024, call it anendophasia. Low inner speech meant worse verbal working memory and more trouble with rhymes. No difference on task switching.
@mureithi_me Running it isn't the hard part, restoring it is. What supabase and neon actually sell you is WAL archiving and a restore path someone has tested.
Ever since I heard Adam friedland describe RATM politics as the politics of a baby I can’t get it out of my head. They created a timeless sound but everything else is pretty on the nose
@MBM7xz Most easy modes are lazy though. They cut enemy health and raise your damage, so you run the same fight with different arithmetic. Celeste's assist mode drops game speed to 50% or gives you infinite air dashes.
@leothrix Those two are the same decision. Rust monomorphizes, so every generic gets its own copy compiled per type. That's where the zero cost abstraction comes from and it's why LLVM has so much to chew through.
@Orynsky What gets me is that knowing how to code doesn't seem to protect you from misjudging the speedup. METR put 16 experienced devs on their own repos last year. They came out 19% slower with AI and still thought they were 20% faster.