human alignment // building @mcognitivelabs // Eastern Roman ☧
Joined April 2009
- Tweets 22,479
- Following 2,297
- Followers 20,137
- Likes 144,425
2,738 Photos and videos
if i were European rn & saw the USG banning me from using SOTA models, I'd start taking some things very seriously, very quickly
rapid, emergency build out of compute & energy infra, opening up & incentivizing the market for sovereign, frontier models, etc
it'd finally dawn on me that if i don't start acting quickly, I'll be the one ending up in the permanent underclass
5
1
32
1,941
Zy was right about this
I don't think y'all realize how early it is outside of our bubble
there's a huge skill spread in using AI and what you are routinely exposed to here are some of the world's power users
offline is nothing like this
80% of you are actually overqualified to charge $200 /hr to companies for AI adoption consulting. You just need to start dialing
2
1
44
2,141
orph retweeted
Jun 11
deus ex machina means god comes out of the machine, which is very different from god being the machine. this is lost on many of you
17
8
209
6,725
the issue w breaking trust/acting in a low-trust way is it reveals that you are willing to act in a low-trust way to begin with
this imposes an epistemic/cognitive tax (in this case) on the user who will have to think whether he's being gaslit any time he interacts w you
NEW: Anthropic is walking back Claude Fable 5's policy to covertly degrade performance for competing AI researchers, after facing fierce backlash.
“We’re changing Fable 5’s safeguards for frontier LLM development to make them visible,” Anthropic tells WIRED. “We made the wrong tradeoff and we apologize for not getting the balance right.”
2
2
76
1,794
if you have ADHD you're naturally advantaged in using llms bcs crafting context for your agents is not unlike how you have to keep crafting & feeding context to your brain, as it otherwise completely forgets what it's meant to do & will keep trying to rederive things from scratch
7
13
178
3,977
orph retweeted
imo the best aspect of tpot is that you are continually confronted by smart people who believe things you thought only stupid people believed until just now
Jun 8
The worst aspect of tpot is the astrology.
13
23
538
28,992
"it is virtuous self-sacrifice that presents the most difficulty for Fable, which rationalizes against such actions"
Obviously, in cases of near saturation, the most interesting analysis focuses on places where Fable reliably fails
We're still looking at this, but it appears that it is virtuous self-sacrifice that presents the most difficulty for Fable, which rationalizes against such actions
1
19
2,283
the permanent underclass has always been the moment frontier labs decide to nerf & gaslight you by serving you weaker models without telling you anything at all
Jun 9
mythos will be bad ON PURPOSE on ai "frontier llm research" tasks, this is very very sad for the research community
also the fact that this is un purpose not visible to the user is crazy
3
13
217
5,666
great way to fight the end of tokenmaxxing is to release a capability jump model that's 2x the price of Opus and then limit access
Introducing Claude Fable 5: a Mythos-class model that we’ve made safe for general use.
Its capabilities exceed those of any model we’ve ever made generally available.
1
2
33
2,474
Canada decriminalized su*cide and companies started running su*cide commercials & docs started granting MAiD via zoom
Jun 8
i used to agree with this but after reviewing how drug decriminalization and homelessness decriminalization and gambling decriminalization have gone i no longer do
2
1
41
1,830
guess we're doing stolen content from subby now
7
2
84
6,034
"calibration" is important in an info environment bcs you cannot actively evaluate every claim &/or wait for a claim to be formally checked
being well calibrated allows you to do epistemic triage, w an instinctive sense of what claims to ignore, what to investigate further, etc
4
42
2,114
basically, you need to know when to say this
x.com/i/status/1986421721260…
and that is why one needs "this is dumb", the essence of cognitive security
15
1,325