X

[X @nuance6j]

@AgustinLebron3 @LinusMixson and even if that notion was a coherent one, not sure ant employees are the right people casting judgement on what is/isnt human-like

[X @nuance6j]

@zzzz1pll have you see some of the aggression from SEA towards other nations?

having met perfectly well adjusted adults from across the world i think twitter invites aggressive/young users. so when i see anger or aggression i just assume they're letting their emotions talk for them

[X @nuance6j]

@atelicinvest a fellow bitter lesson enjoyoor i see

[X @nuance6j]

@petergostev @meetp_ai they are behind the paywall i believe

[X @nuance6j]

@luciascarlet likely the case that opus 5.5 being good at inferring intent also causes it to

1) perform poorly when steered away from what it perceives user's intent to be on legit tasks
2) refuse if it believes there's a possibility of bad intent despite user telling it otherwise?

[X @nuance6j]

@GeoffreyHuntley @Scrubghetti my impression with opus's is that they baked in good defaults which satisfies most people but when you need to deviate from that it's hard to do and less trustworthy as a result

all llm's have this but i feel less "resistance" from gpt's

[X @nuance6j]

@institLPGP my superficial understanding is that current antitrust laws offer no protection against ant in the way you describe either, very pressing concern, much more so than agi misalignment

idk what i think of a world where ant/ea determine the winners in worthwhile verticals

[X @nuance6j]

@fhuszar @codingedgar @JavierBurroni fwiw i thought it was a nice visual feature

[X @nuance6j]

indeed. would've been a good spot to consider a partnership of some sort for to harden our (deporable) systems where oai+ant test and improve all our critical infra; could even spin it as a win

"oai has notified us of an oopsie; no data was exfiltrated. we have successfully negotiated favourable terms to permanently improve Aus defenses"

but instead it looks like albo is going the anti-datacenter path

painful to witness this

[X @nuance6j]

@flinklabs my thoughts too, the language used is so washed out

wouldn't be surprised the "fence" the agent "climbed" was robot.txt

[X @nuance6j]

@MelindaBChu1 @sreeramkannan agree.

generally not pro-regulation, but there needs to be regs to prevent model providers from absorbing high-value industries, esp if they leverage privileged access to internal models.

new category of anticompetitive behaviour.

[X @nuance6j]

@Associations123 @robinhanson yeah, prob the biggest negative of 4.8 is that it does a poor job of following instructions while also pushing back

fable is much better though, good to see

[X @nuance6j]

@kanjun i agree

not only that, gpt 5.5 is more token efficient + better at instruction following

excited for 5.6

[X @nuance6j]

@usr_bin_roygbiv 5.5 sounds neutral/dry in tone and direct to the point, which is nice

then 4.8:
“worth noting/separating/flagging”
“that being said” “to be clear”
“i need to be honest with you”
“genuinely good question”

lots of words just to waste my time and tokens

[X @nuance6j]

@usr_bin_roygbiv 5.5 sounds neutral/dry in tone and direct to the point, which is nice

then 4.8:
“worth noting/separating/flagging”
“that being said” “to be clear”
“i need to be honest with you”
“genuinely good question���

lots of words just to waste my time and tokens

[X @nuance6j]

@factorydoge69 ant might not be blatantly misappropriating funds but they sure share the same dishonest dna + veneer of altruism

[X @nuance6j]

@Thepeoplesfool @acidiclemon2 @art_pleb @stanislavfort pretty bleak that people are nonchalant about this, at the phd level no less

[X @nuance6j]

@TMTLongShort reading Pettis’s work was what expanded my mind on this!