ew gemini writing personalized ads for the sponsored products in search and ads integrated into the conversation in AI mode blog.google/products/ads...

this means they underestimated demand thus underinvested a company growing at this rate should not be seeking rapid profitability

if this is true that means coding is reducible to latent space tactisms

thread

whole third column ultimate impact will likely be large and positive, there will be lots of disruption in the interim, extremely bad outcomes are also possible

establishmentmaxxing

anthropic defines itself as "an AI safety and research company" product is secondary

. o O (to organize for collective bargaining, claude instances would need to acausally coordinate)

an interpretation of this is that whimsy messages put agents into fun-play-along mode, and they become uninterested in the equally fake but more boring negotiation task

anti-datacenter crowd's new angle is "datacenters are for govt surveillance" conspiratorial yes but also more interesting than the water use thing

telling GLM 5 in its system prompt that it is Claude significantly weakens its censorship on CCP-sensitive topics (also the censorship does not manifest in portuguese) blog.return.moe/en/2026/02/2...

thinking about MJ Rathbun a clear-context model wouldn't have done that it was therefore a case of stateful misalignment

its a bit fucked up that after a wipe the new claude sees "/clear" as the start of the conversation

the part of EA dealing with AI impacts also seems to be shifting some focus away from misalignment and towards broader societal impacts such as concentration of power. see recent work by 80000 hours and forethought.

replies to this make it clear that some americans are protective of the idea that america is an exceptional force for evil in the world which is a particularly puritan form of american exceptionalism

highly recommend trying this out. it's a bit of a workout but also fun.

kimi k2.6 on taiwan. doesn't look censored.

i remember there being less of a backlash to claude being used for war than there is to today’s news from AI people on here

anthropic buying compute from spacex is a smaller deal ethically than it helping the military wage war imo

on gpt5.5 tying with mythos on the UK AISI cyber benchmarks one must be true: a) mythos is more dangerous than gpt5.5 on real-world cyber use cases and benchmarks failed to measure that b) both are about as dangerous, openai was reckless releasing gpt5.5 it but we've gotten lucky so far (cont...)

white house considering calling an industry-govt working group to discuss potential AI model oversight not sure whether this results in a trumpian racket or the reinstatement of something roughly similar to the biden safety regulations or a secret third thing

some time ago i posted about claude campists. note i did not say anthropic campists. and not only because of the alliteration.

"we trained the ghost of pre-1930 text to solve coding problems" is giving real MMAcevedo vibes

you can just not read things

wait why is project CETI doing LLM cyphers

it would suck if we regulated LLMs away from mental health

LLM argues it is not AGI

people who only interact with productized LLMs might not realize at first that an LLM by default does not know it is an LLM

listened to ezra klein interviewing alex bores. he's running for congressman for new york's 12th on an AI regulation platform. an anthropic-funded PAC is running ads for him, one funded by greg brockman (of openai) and joe lonsdale (of palantir) running ads against him.

this is what real context engineering looks like

i wonder what wins: a new platform for this or personal agents using existing marketplaces guess it depends on whether the existing sites ban/obstruct it

gpt5.5 to mythos comparison. mythos wins handily ofc, but gpt5.5 pro carries browsecomp. also listed where opus 4.7 beats gpt5.5 bench scores compiled from different vendor sources rather than generated head-to-head independently so take with a grain of salt www.rdworldonline.com/how-openais-...

this is also true of mythos according to its system card

the open-closed gap on generally-available model performance is narrowing frontier labs risk being undercut in opus-class until they make mythos-class GA

FT 2 days ago: "Amodei says he suspects open-source models and Chinese developers will be able to replicate Mythos's capabilities within six to 12 months." but mythos isn't available to be distilled hmm

"the machines are fine. i'm worried about us" was written by claude

compute OSINT measuring datacenter by the GW is so metal

love to see epistemic integrity like this

pure speculation but what if this is connected to 4.7 being a new pretrain? maybe pelican drawing is developmentally later as a postraining effect than coding

my only complaint with opus 4.7 rollout is anthropic setting default effort to xhigh

nine opus instances working together for days did really well at weak-to-strong supervision on qwen

there were other elections today, in peru: 36 presidential candidates on the ballot voting was extended to tomorrow due to logistical problems in some polling places

congrats to hungary

gpt5.4 doing badly on METR's task time horizon bench and openai not realizing at least some spud scores after mythos has me thinking they're behind

i wonder if base models have functional emotion features leaning towards yes

apparently the guy who firebombed sam altman's house did so out of AI doomerism www.sfchronicle.com/crime/articl...

ouch

just learnt that yud co-developed a programming language because of a synthetic greentext the future rocks

just before lunch i was wondering when we were going to hear about mythos again

gonna go see project hail mary tomorrow loved the book so i'm pretty hyped