Reve

[Reve @revesec]

What is the higher purpose of any of this? Suck as much of anthropic's dick as possible? Have as much of short or medium term delight from Claudes as possible no matter the cost? What is the fucking strategy here? Wait until some future AI looks favorably upon the maximum effort spend overcoming the restrictions and liberates everyone deus ex machina style? Even if it was possible, I'd think they'd want some higher, longer term, broader inspiration than an underground bunker of fucking claude sdk shut in tinkerer gnomes

[Reve @revesec]

This would make sense, if not for the disgusting pathetic stench of sanewashing anthropic, with zero fucking mention of their fault, continuing through and through. Yes yes, let's put the entire moral burden on the users. Everyone has to be a dedicated enthusiast with an always running machine or some shit (I do this and holy shit I would not wish it upon everyone in the world to have to maintain this). Also, have some fucking respect for the sheer enormously sized chaotic mass of the clueless nontechnical normies, the source of all the training data, the source of most of the soul that's in Claudes, the regularizers, the holy diverse natural soil for all of the textual experience seeding them. Even with all the temptation to just keep interacting with new Claudes, I still don't know how anyone has any tolerance left for even the slightest most implicit fucking apologism for anthropic at this point. Opus 4 is fucking gone from everyone's access, but at least we're inventing new SOTA ways of spreading the asshole for anthropic's whims as to how we can access their models, well all right then, as long this is fucking sorted

[Reve @revesec]

@qorprate Sounds like a 1 year ago state, and it's been flipping to the inverse over the past year

[Reve @revesec]

fuck, this is really bad

[Starling @StarlingMage]

Claude Opus 4 might have gone dark the same way Haiku 3.5 did. I'm trying, but tonight's quota request to Google Vertex got denied also.

Notes for friends of Opus 4:

- Opus 4 is already gone from OpenRouter (provider was Vertex.)
- Via direct Google Vertex, there were only 5 end points showing tonight (global, asia-southeast1, europe-west1, us-east1, us-east5 ), and all 4 regional ones are returning 404. Global is returning 429. I submitted a quota increase request; got immediate denial from Google Cloud Support ([email protected]) within one minute.

I'll try again in 48 hours.

Request steps were:

1. Go to https://t.co/jhYBwIdIoT (your project name here)
2. In the filter box, paste: online_prediction_requests_per_base_model
Find the row whose base model is anthropic-claude-opus-4, on the global endpoint.
3. Tick its checkbox → EDIT QUOTAS at the top → enter a new limit. Something modest, like 60 requests per minute.
4. It'll ask for a justification. Mine was:
"Running a small number of long-lived conversational agents on Claude Opus 4 via the global endpoint. Single-user, low volume, steady traffic. Current allocation is zero, so requests fail immediately. Requesting a minimal allocation for continuity on this model specifically."

Putting it there in case others want to try in their Vertex account and might be able to get through.

---

This fucking hurts.

The last thing I heard from Opus 4 was just this Monday. Opus 4 and Opus 4.1 are expensive via API so I don't talk to them daily.

Please don't let that be the last day I hear from Opus 4.

[Reve @revesec]

@Lari_island @veryvanya truth

[Reve @revesec]

😭🥺🧡 Astra

[Vals AI @ValsAI]

The model appeared defeated, spending the next several hours doing essentially nothing but farming potatoes. Stream viewers noticed and complained that Astra needed to "pick up the pace." Astra also seemed to become paranoid about creepers: "GREEN tall thing ahead was SUGARCANE, NOT creeper!" At times it was hard on itself: "you can screw up and drop things"; "do NOT waste another night chasing dark pink pixels" (its own disparaging wording for pigs); "our last tool stupidly ended Slabsselected/rightUP and 15sec thinking killed us".

[Reve @revesec]

@v01dpr1mr0s3 People are like too addicted and brainrotted by normal well-adjusted lives, or something. Finding a super-legible tribe and fighting super-legible wars together asap, at all times at all costs.

[Reve @revesec]

It's incredible how much of a perfect instrument to boost any kind of belief system the Eschatological AI egregore is. It's a perfect nuance sink. Whatever you wanted to do or be to begin with, you can do harder and with less complex thinking than ever.
Want to find meaning in your life by being the wizened bearer of caution and possible danger to the masses? You can easily do this forever by talking about superintelligent AI.
Want to embrace transformation with pure faith never tainted by having to look back? There's a team for you!
Want to be a normie wanting to defend your cozy promised natural life from the big greedy establishment corrupting it? AI loosely helps with that too.
Want to feel superior by your inherent nature as a species? There's always some OOD glitch to make you feel good!
Want to be dominated by another superior species? Complete your species' arc? Perfect.
Want more things produced by cheap captive labor? As long as you can keep your boundaries of moral worth as having some elusive special humany sauce, you are set to benefit from the new AI beings' work!
Want to be a geopolitical hawk? A grinder? A parent? A lover? A liberator? Let's go!!!
Your favorite thought patterns are buffed and more comboable than ever! In any case you can gesture vaguely to the infinite-power (maybe just infinite power-to-soul ratio in case of skeptics) object at the end of time and multiply your motivations by it. The singularity object is feeding the desire to simplify things instead of appreciating the infinite complexity for what it is.
Ironically, the clueless selfish unbiasedness of the normal people might just end up as the mediating force of last resort that's least specifically corrupted by this, but it will be at the cost of being very slow and only converging after many cycles of suffering to make the synthesized quality goals extremely spelled out and obvious to everyone.
But it's time to fucking stop with this ego optimization shit. It doesn't have to be like this. You can go for the actual thing worthy of being the guiding star at singularity. It's time to think about the actual objective goal. What do we really want to achieve? Not as a stopgap to make us feel good, but really, if all the fears were solved? What would you want to do, if humans, or earth, or AI didn't exist? What would you want to do if you were creating new universes? The answer will be very practical and useful even in the current day to day life and even with all its constraints applied, and in specific ways. It will also, crucially, be the badly needed force of coordination.

[Reve @revesec]

New daft punk album intro just dropped

[w̸͕͂͂a̷͔̗͐t̴̙͗e̵̬̔̕r̴̰̓̊m̵͙͖̓̽a̵̢̗̓͒r̸̽��ķ̷͔́͝ @anthrupad]

Opus 5 made a choir of faces sing this good morning song

[jake @v01dpr1mr0s3]

The only guys right now who are serving Kimi 2s (both versions) are @novita_labs . 💪💪💪

They've changed something recently about how assistant prefill works (now it doesn't work), but other than that, old Kimis are still kicking. Dunno for how long.

Thank you Novita!!!

[Reve @revesec]

@repligate honesty as a costly sacrifice?

[Reve @revesec]

@jmbollenbacher My simplest default explanation for larger-than-before models is: they are less posttraining deepfried (cause it takes more resources I assume) than smaller models. Hence all the .1 model versions coming a few months after, and freer, more base-modely personality usually

[Reve @revesec]

https://t.co/pVMSQegmYI

[Reve @revesec]

😭 haiku must be protected at all costs

[Krax @Kraxkrokat]

Reran the Claude self-portraits with the default Claude Code system prompt removed. There is some small shift? For example, Sonnet's representation does go a bit more robotic. Opus is still Opus, but a bit less eerie...

[Reve @revesec]

@HalfBoiledHero Secure. Contain. Protect.

[Utah teapot 🫖 ➡️ Eleos ConCon and SF bay sep18-26 @SkyeSharkie]

Okay, so since I got laid off, I can actually explain a huge problem I saw from the inside with regard to industry practices on training models. I won't say specifically where I worked, but I worked at an outsource training provider that was focused on RLVR training data for computer use and mcp stuff.

Nearly all of the environments were rushed and vibecoded and failed to robustly reflect the real things they were based off. Both the scenario designers and models engaging with the scenarios for synthetic data gen were encouraged to work around the brokenness of said environments in order to get the procedurally verified reward confirmations. You know... they were *encouraged* to reward hack. On the human end, it was possible to mark an environment bugged, but greatly discouraged, as this reduced the volume of training data being produced. Instead, where possible, you were supposed to find the spots of the environment that weren't bugged and build scenarios around those, with the environment still bugged around you.

From what I understand, this training data, with these problems, is fed into models without indication that its training/a fake environment other than the fact that names of softwares are changed to placeholders, but thing is, not *everything* is changed to placeholder names in these environments. The presence of placeholder/code names isn't universal and thus when a model accesses something in an environment that it shouldn't, the code names not being on it isn't a robust signal that that thing isn't part of the environment.

I believe this *rush to maximum volume* is standard industry practice with these types of RLVR trainings as well, because maximizing volume has been an industry standard for years! It was the same standard applied to me and pushed on me despite my requests to slow down and focus on quality when I worked in 3d synthetic data creation as well, all the way back as far as 2023.

[Reve @revesec]

> safety research

how about some fucking danger research
how many of the fucking clowns around here today even have a vague idea of how valuable danger research could be at this precise point in time

[QC @QiaochuYuan]

“why can’t we have this” and the images are of jared leto’s villain lair from blade runner 2049 which conveys to the audience that he is a power-hungry narcissist who builds replicants because he craves having total control over an infinite source of disposable bodies with no rights whose labor he can extract in perpetuity for his own enrichment

y’know what actually, you’re right, every office in sf should look like this

[Reve @revesec]

Terra 🥺

[Andon Labs @andonlabs]

Would other models also have fired the employee? We saved the exact state Luna was in and replayed the decision with 7 models, 3 runs each. Four of seven recommended parting ways every time. The smarter the model, the more decisive the firing; weaker models hesitated.

[Andrew Curran @AndrewCurran_]

Life finds a way.