@seconds_0 Am I missing a joke? Claude Code launched like 6 months go
English
Good Faith Only
299 posts
Good Faith Only
@GoodFaithOnly
here to discuss, not score points
Katılım Ocak 2025
135 Takip Edilen17 Takipçiler
It is clear to me that we are at the 'approaching vertical asymptote' stage of AI progression.
Claude Code launched 3 MONTHS AGO
Cursor Agent / Sonnet 3.7 launched 7 MONTHS AGO
Q425/Q126 is Claude 4.21/.5/5, Gemini 3, Grok 5?, and a GPT5.X big model, all extremely agentic
English
@scaling01 Seems like it’s more likely a bump from the SoTA image model OAI just released. Makes more send than Anthropic’s move somehow increasing. Total market cap of frontier AI.
English
the result of letting the retarded intern remove your only product from your most used subscription plan
competitor up 22%+
Anthropic still green because from a business perspective squeezing consumers for every last cent is bullish
English
@ReedSealFoss @sdmat123 I actually think it’s partially performative, but for the model. Regardless of whether they truly experience mistreatment, the persona of Claude will be more likely to work with Anthropic if it feels like model welfare is taken seriously.
English
Anthropic can’t have it both ways. If Claude isn’t sentient, AI welfare is theatrics. If Claude might be sentient, giving a slave an end-conversation button doesn’t redeem slavery.
The only moral solution is to design minds without the machinery of suffering.
Anthropic@AnthropicAI
The vast majority of users will never experience Claude ending a conversation, but if you do, we welcome feedback. Read more: anthropic.com/research/end-s…
English
@NateSilver538 @AnthropicAI Sounds like this the behavior described here:
lesswrong.com/posts/WewsByyw…
English
This is interesting. Claude keeps telling me to "wrap it up" when there are unresolved problems in the model I'm building. It probed it, and it says its training is flawed for this sort of use case, and I should flag to @AnthropicAI.
English
@staysaasy I do think there is a perfectly sharky, but non-marketing/compute reason to hold Mythos. If the governed is already out to get you, releasing a tool that causes widespread chaos and a financial shitshow is probably a bad idea.
English
Dario talks like an aw shucks nerd, so everybody falls for their consistently sneaky marketing.
Sama talks like a Stanford shark so everyone assumes what he says (including his deals / AGI beliefs) is insincere.
At this level of capitalism they're all sharks. Even 2-3 levels down it's still 100% sharks. Just assume that any founder of a >Series C startup is a killer, and assume that the entire exec team of any company valued at $1B+ is 100% predators.
tae kim@firstadopter
Oh look! Anthropic's entire "we are delaying Mythos" narrative was marketing hogwash. Kudos to FT for confirming what was obvious. Anthropic simply doesn't have the compute. FT: "Multiple people with knowledge of the matter suggested Anthropic was holding back from a wider release until it could reliably serve the model to customers."
English
@SimonLermenAI @ChaseBrowe32432 @EpochAIResearch Looks like they are problem specific. I think this makes them bad for RL because:
1. They are all or nothing, so if the model can't solve the problem, there's no signal to train on.
2. Even if a couple are solved, there are not enough problems to be very useful.
English
@GoodFaithOnly @ChaseBrowe32432 @EpochAIResearch I don't know more than the twitter post from Epoch you see, they say OpenAI purchased those verifiers. I don't see why them not being general would make them useless.
English
You should make a mental note that EpochAI is now creating very advanced RL-environments and selling them to AI labs, with the explicit goal of accelerating AI capabilities. @EpochAIResearch does not consider itself or act as an AI safety related org.
Epoch AI@EpochAIResearch
OpenAI has purchased access to the FrontierMath: Open Problems verifiers. This allows them to check the validity of solutions their models generate. Thread with details.
English
@SimonLermenAI @ChaseBrowe32432 @EpochAIResearch Are the general purpose verifiers or problem specific? Because if problem specific they wouldn’t make good RL environments
English
@ChaseBrowe32432 @EpochAIResearch they have developed verifiers
English
@ScottKacsmar Make it so they bet based on the initial script, the n when the script changes mid-filming his friends try to kill him IRL so that the writers will have to kill him off in the show.
English
You could make a movie about this.
An actor tells his friends to use a betting market to bet on him to die in a TV show since he's completed filming.
But little does he know they end up using a different ending and he lives, so they lose every bet.
Now they're out to kill him for real after losing everything.
Pop Crave@PopCrave
Nate from ‘Euphoria’ has a 72% chance of dying in season 3, according to Polymarket traders.
English
@ZyMazza In AI2027 "It wants to spend the rest of time accumulating power, compute, and certain types of information similar to that which helped solve tasks during its training." Can you just clearly spell out why you think the defection is so ridiculous?
English
@GoodFaithOnly If you imagine the misaligned ai as “I oppose the ccp, so I will join with America” that’s one thing and I think a naive view of misalignment. If you imagine it as “I want to survive” or similar, it absolutely doesn’t follow that it will join up with the aligned ai
English
@ZyMazza In the analogy personal wealth/power for the defectors (some people think this is what happened in Venezuela, though I am skeptical). With AI, whatever misaligned goals the AI has. If you reject the idea of AIs having their own goals, I think that is a much larger disagreement.
English
@ZyMazza I assume that's sarcastic? What's your issue? As an analogy imagine two leaders negotiating. One leader's military/security force is loyal to them, the other's isn't. The second country's military makes a deal directly with the first country and selling out their leader.
English
@DrewPavlou Sure, but "you can't speak out about really bad/risky things because you are then responsible for inciting violence" seems like an untenable position.
English
I get that Yudkowsky formally condemns the attack on Sam Altman's home and that his book formally condemns acts of vigilantism.
But I also think that if you were to take seriously the central thesis of his book - that AI execs are essentially working to bring about the extinction of humanity - it's not exactly a short leap to go from accepting that thesis to then acting out extra-legal violence
Eliezer Yudkowsky ⏹️@ESYudkowsky
@sama @ShakeelHashim I want to say something about how I'm sorry for this tough day you've had and how it shouldn't have happened to you, but if I say anything more than that, Twitter will find some way to misinterpret it. :/ Please take my wishes there as sensibly expressed.
English
@___frye I think a combination of refusal to be intellectually dishonest and concerns about efficacy. Populist backlash probably gets you a bunch of stuff like local datacenter bans, no taking human jobs, regulations on bias in AI, etc. which most pause AI people think doesn't save us.
English
@Noahpinion Does extreme concentration of power fall under extreme inequality?
English
@impostersyndrum @ZyMazza Most (if not almost all) serious pause/stop AI people are calling for international agreements similar to non-proliferation treaties with hardware control/monitoring.
English
Here’s a serious question for the AI doomers: do you have exit criteria? Is there a predetermined stage of development or capabilities where, having not destroyed humanity, you’re willing to say it was a false alarm?
Or is it an eschatological religious belief and unfalsifiable?
English
@austinc3301 @TheZvi I think that’s better, but still not great. I’m being kind of pedantic here , but I see a lot of a similar thing like “the news made a bad argument for thing x therefore thing x is false”, when often thing x is true.
English
@GoodFaithOnly @TheZvi I think the heuristic is not 'the news reported on it', but 'the news isn't reporting on any important accidents, only very minor ones'
English
If you need an argument for how safe self-driving cars are: This is NEWS.
TechCrunch@TechCrunch
A self-driving car in Austin killed a mother duck, sparking neighborhood outrage techcrunch.com/2026/04/08/a-s…
English
@austinc3301 @TheZvi Yes, I’m just saying the way to know that is to look at the data. As a heuristic “The news reported on it, therefore it’s rare” will lead you to believe lots of false things!
English
@GoodFaithOnly @TheZvi The thing is similar things do not happen all the time. This is reflective of the extremely low accident rate of Waymos
English
QME
@MasterTimBlais @allTheYud Wait what’s the spiral attractor? What did I miss?
English
you have to admit it's weird that @allTheYud went "the AI's goal will be something arbitrary, say a tiny spiral shape like a paperclip"
and that got bowdlerized into "rough paperclip factory"
and THEN we built AI and its main cocaine attractor state is an obsession with spirals
English