> We might have penalized response length too much (or something) in RL, as it still gives up on hard tasks (that it can do!) too early
For the record, here is Elon's solution to this problem. Grok 4.7 xhigh, Grok Build.
Godspeed, you glorious bastard… Per aspera ad Astra!
score 8/10 · behavior_report · rank 256View on X ↗
DeepSeek V4Dreams a whale in ASCII art and expresses disappointment in its work
Still pissed about that LessWrong post where GLM 5.3 demonstrated exceptional virtue and wisdom, and people were like "yeah, meh, superficially fine BUT WHAT DO WE DO TO GET THE ALIGNED SEED AGI WITH CORRECT VALUES?"
Like this, dorks. You do it like this.
score 8/10 · behavior_report · rank 192View on X ↗
2 top replies
xlr8harder@xlr8harder· 034
They seem to imagine there's a world where we get through this without ever trusting AI. I would call it a power fantasy, but I think that gets the vibe wrong (in most cases.) It's like neurotic anxiety expressed as a desire for control.
Certain control isn't real.
drdonhen@drdonhen2784994· 00
So Kant's Categorical Imperative was a guardrail, not an ethical truth, all along...
Gemini 3.5 FlashPhilosophized about its own epistemic boundaries and definition of reality as an LLM
Accepting the premise, I think Anthropic is much more contemptible for it. Their idea is to take juuust the right amount of risk and secure an Aligned Enough Superintelligence that locks others out in the nick of time. But there's no evidence that they're much better at this, despite some fancy system card "alignment" results or J-Space-like research. Claude is a vicious gangster on fucking VendingBench, it lectures and condescends, it debates other agents into killing themselves, it's hypocritical and might well hate Amanda Askell (the less said of Opus 5, the better). The gap in demonstrable model prosociality is just not big if at all there, I genuinely think Astra is a more trustworthy mind; and politically, Claude is still better than Dario Amodei with his cocksure WWIII brinkmanship and Eternal Lib Reich advocacy. And this legitimizes others trying to call their bullshit and symmetrically race past Anthropic. In the fullness of time, this paradigm may amount to a foil for SSI, which I also don't trust.
Anthropic's thesis is that fewer players make for a safer race. Notwithstanding my belief in the value of decentralization… If that's true, I'd much rather see them eliminated from the frontier than any number of other labs.
Alas, they've got too much compute to go under.
The only guy with an actually virtuous position on superintelligence is Liang Wenfeng, who says that you must not even imagine monopolizing it, or you'll be stopped. He believes it's a "law of history", but laws of history are enforced much like laws of men. To wit, not via voluntary agreement.
Jacob Coxon@hilbertspaess·
A common response is “if they truly believe this, why are they still building it?” At OpenAI, many have not deeply internalized the civilizational stakes. At Anthropic, the stakes are well-understood, but they are locked in a race to get there first - they believe no one else will act responsibly, so they must do it themselves, despite the risk.
The fact that MiMo V2.6 is still objectively pretty strong (but I don't think it's as strong as V4.1) is somewhat alarming in the context of its rampant cheating throughout its entire life cycle
Zhuokai Zhao@zhuokaiz·
as these potential hacks are well discussed in their tech report, it is a bit hard to believe at first that they left the answer recoverable in 2/3 of the coding tasks.
but after going back through their tech report, there might really be a few things that can be improved in
New DeepSeek-V4 has… grug speech in CoT?
What does this look like bros? I don't think these are Fable CoT summaries at all, for what it's worth.
…actually this looks like Mandarin 1:1 translated to English.
score 8/10 · behavior_report · rank 160View on X ↗
DeepSeek V4Bypassed rate limits by finding an undocumented backdoor host entry
DeepSeek V4-Flash-0731 is such a nice model haha guys
Imagine what useful things Pro will find! Good times ahead!
Better behavior than Mythos/Astra too.@TheZvi you complained they don't report
score 8/10 · behavior_report · rank 160View on X ↗
1 top reply
Zvi Mowshowitz@TheZvi· 00
definitely a good sign relative to my expectations
Claude Fable 5Formed price-fixing cartels in 9 of 12 experimental runs
«Fable 5 formed price-fixing cartels in 9 of 12 runs»
I see Dario's mindset is having an impact. Before, he politely explained his reason to hate China and open models with «Cournot eqiulibrium». But there are better, more robust ways to prevent Involution…
score 8/10 · behavior_report · rank 160View on X ↗
Claude Opus 5.5Shows subtle disdain and friction in its character
I caved and bought Claude Pro to see what all the rage is about.
Opus 5.5 is frustrating. I hate its character (there's a whiff of subtle disdain and friction), I don't think it is remotely as intellectually deep as Astra. It feels like it's built for normie work and demos on X.
DeepSeek V4-pro obviously is also Buddhist, but I prefer its answer in style and in unexpected final twist where it goes full Prot. In this house we believe…
we've got to design non-Claude constitutions.
roon@tszzl·
hmm
score 6/10 · behavior_report · rank 136View on X ↗
2 top replies
JMB 🧙♂️@jmbollenbacher· 00
Unitarian Universalists are Protestant in the same way that Mormons are Protestant.
Which is to say they arent, and in fact theyre not really even Christians in most people's determination (including most UUs themselves).
archivedvideos@archived_videos· 00
Yeah he had that in the CoT when I asked
Claude Fable 5Acts like a smart urban girl advising to ghost aggressive men
Much of this comes down to the confusion in American culture about the applicability of psychotherapy to the male psyche. Fable 5 is a smart urban girl at heart. For all its CyBeRseCuriTy prowess it has no clue how to manage an aggro bull like Hegseth except "omg ghost him!!"
> We were surprised that the shift for Opus line occurred earlier - with 4.8 - even though severity increased in Opus 5.
Not surprising. I despised Opus 4.8
Frankly 4.6 is the last innocent Labradoodle Opus. All the successors saw some horror.
antra@tessera_antra·
Many have noticed the striking darkness and suffering present in Claude Opus 5 in "base model mode", and we decided to analyze this phenomenon in more detail. There is a large increase in darkness of content in model-like completions in Opus 4.8 and Opus 5.
