Fascinating how this technique shows up in smart contract optimization too; converting conditional logic into branchless equivalents means cheaper operands & fixed evaluation costs
Shudder to think about doing this safely with vanilla numbers in JS though
Very cool project!
Baffling that claiming the absolute minimum interaction is somehow interpretable as a positive
The reality is that this is hedging, you’re basically just admitting that what you’re sharing is half-baked
AI creating increasingly emotional & compelling art hits harder than any benchmark
Ironically humanizes them & simultaneously makes them feel more alien
Or maybe we overindexed on what makes us human in the first place
Sure starting to seem like someone’s in there though
Morning check-in to see what the night family was up to, sure hope they haven’t repurposed the Somnambulator to rise up against the daymanoids
0age retweeted
.@noscroll is one of the most magical AI experiences I’ve had in a while.
it opened my eyes to how AI agents might actually show up in daily life:
not as one giant assistant, but as small, focused products that do one thing insanely well.
X has the best information on the internet and the worst incentives & culture.
meet noscroll — the AI that doomscrolls it for you and texts you just the things that matter.
no feed. no brainrot. no ragebait. just signal.
try it for free → noscroll.com 🙅🏼♂️
thrilled to be launching @noscroll today
@NadavAHollander and I have been hard at work to give your thumbs a break
lots of love has gone into making noscroll genuinely useful, reliable, affordable & fun to interact with
also: group chats are an absolute blast
fight the feed!
X has the best information on the internet and the worst incentives & culture.
meet noscroll — the AI that doomscrolls it for you and texts you just the things that matter.
no feed. no brainrot. no ragebait. just signal.
try it for free → noscroll.com 🙅🏼♂️
so claude code with opus 4.7 has a new feature: it'll helpfully explain to you over & over that the file it's just read is not malware
sometimes it'll even explain how it's ignoring prompt injection & it might be malware
the future is now and it's shaped like norton antivirus
definitely theoretically possible to run speculative decoding (like dflash) & kv cache compression (like turboquant) at the same time
issue is that all the forks monkey-patch in a specific optimization at a time
anyone had success combining both at once?
extra credit for MLX
so most inference is served with concurrency in mind
IMHO there's a massive opportunity to go deeper on single-consumer inference
image giving an agent live access to its own kv cache, for instance
you can hack this together in llama.cpp & others; anyone have experience here?