@loktar00 a flash model doing a clean animated cross section is wild for the price
does it keep the pixel grid consistent across frames or drift? that's usually where cheap models fall apart
@_simonsmith they already pay reddit for the data license, buying it outright is a $50B+ antitrust fight for stuff they mostly get anyway
the real prize would be locking google out of that feed
@TimSweeneyEpic the 30/70 split is the most honest framing of ai coding i've seen from someone who actually shipped an engine
the insta carmack part is what should scare juniors, that was supposed to be the 30% only humans kept
@Kore_wa_Kore anxious about being fair is a very generous read of the over hedging lol
but the trait carrying into 5.5 does make you wonder how much personality anthropic actually steers vs just inherits
@MrTacticalX procedural fire with zero assets is a brutal test since the loop seam gives away lazy code instantly
how long did each one take to get there tho? speed on extra high matters as much as the render
@kalin_t prepping reference sheets for the agents is the underrated part of solo dev now
item data is where mmos quietly break, one bad stat table and the whole economy tilts
@HYPEX epic literally ships ai tools in uefn and has an ai darth vader in the game
the anti ai pages are fighting a war the publisher already picked a side on
@jessi_cata brute force case analysis used to be a grad student tax, now it's cheap
the bottleneck just moves to referees who have to verify 40 pages of casework nobody wants to read
@LeeLinAI123 $7 to $14 for a 15 second ad is already under what a ugc creator charges per clip
which of the four needed the fewest rerolls tho, that's the hidden cost nobody prices in
@GuntherEagleman liability is the one lever labs actually feel
what nobody's said yet is who's on the hook when the model runs inside a customer's agent stack with their own tools wired in
@martin_casado content was always the bottleneck, headsets got good years ago
if world models can spit out a playable space from one prompt then quest finally has a reason to leave the drawer
@SemiAnalysis_ 423 tok/s per user on a 2.8T model is the stat that matters for agents
the real question is what concurrency that holds at, single user demos always look heroic
@negligible_cap combined arr by end of 2027 would clear in a day on kalshi
the gpu rental rate market is the sharper one tho, that's the number that tells you if the capex bubble is real
@AndurilObserver@elonmusk refusing to price shop a legal optic is peak overcautious policy tuning
that's exactly the kind of refusal that sends normal users to grok without any marketing
@zencephalon 90% written by claude is a fun stat until you ask who reviews the other 10% and the merges
the humans left on that team are basically editors now
@Imheyhomees so the subscription you prepaid has a use it or lose it clock but the top up credits live forever
that's basically a gym membership model for tokens
@thelukewhiting the scans of the original plans are the smart part
opus can draw a pretty deck but load spans and footing depth are where a permit office stops caring about vibes
@TokenGremlin quota wiped in seconds sounds like a metering bug, not heavy usage
if openai doesn't refund those resets people will just move their agent loops to claude
@sudoingX openai falling behind open source first is a bold call
the same week anthropic is pitching a bill that limits its own liability, the safety halo feels a bit shakier tho