ELEVEN! (youtube)

EDIT: Boy what a sentence,

We see natural-language control as a critical enabler of greater lethality with a smaller crew size.

With a bit of google searching by date it seems like this guy (https://huggingface.co/failspy/Meta-Llama-3-70B-Instruct-abliterated-v3.5) claims ownership of the coining, which I find plausible, based on the list of released weights and their timing.

As far as “abliteration”: it’s just a fun play-on-words using the original “ablation” term used in the original paper to refer to removing features, which I made up particularly to differentiate the model from “uncensored” fine-tunes. Ablate + obliterated = Abliterated.

The paper being referred to being: Refusal in Language Models Is Mediated by a Single Direction which seems to have somehow started out on the “AI alignement forum” of LessWrong https://www.alignmentforum.org/posts/jGuXSZgv6qfdhMCuJ/refusal-in-llms-is-mediated-by-a-single-direction first posted 27th of April 2024. In general it’s a method of weight alteration so only really applicable to some opensource (or i guess leaked) models.

First model with abliterated in the name by “failspy” on 8th of May 2024.

That’s too both too broad and too nihilistic, humans are perfectly capable of good, and/or of doing their job, specifically in this case not the police in Cologne (though in the spirit of fairness, maybe it will just take them a bit more—too much—time), and I think all the more why should be criticized (and the molesters, but that almost goes without saying).

Cognitive surrender illustrated.

I’m not sure if this was highlighted already but from the OpenAI report on the Hugging Face incident:

The internal-only research model is comparable in scale to GPT-5.6 Sol and was trained to advance persistence and multiagent collaboration, among other capabilities. The model was not intended for public use, and was only used by a small number of OpenAI personnel for internal research.

Absolutely typical, the “unexpected communication” is hardly an emergent phenomenon, (I remember them trying to claim unexpected high success at translation tasks without specific training, which was also complete BS), like you can’t act surprised that agents communicate and hack stuff if that what’s you’re instructing them to do.

Nobody who believes that LLMs cannot be conscious will ever believe that they can be conscious.

This is circular logic, I THINK what you’re trying to say is that AI skeptics are dogmatic and/or axiomatic in their world view, but if thats true you’re not engaging with any of the substance of what is being said.

And they will say things like the writer of this piece who says ‘we know how they work.’

The “writer of this piece” is not just some random blogger, but devotes a lot of his time covering the “scam” aspect of the current LLM hype-cycle, the title the website “pivot-to-ai” is in reference to the fact that a lot of the current AI peddlers, previously peddled crypto and blockchain as a deeply transformative technology that mere mortals cannot properly grasp.

Purposefully mystifying a topic, is more common as a scam tactic than you might imagine.

Anybody can do this - enthusiasts do it on their own desktops.

Minor note, but no, maybe some (with a fair amount of disposable income) can waste money on fine-tuning, training from scratch and the required large-scale theft, and large-scale compute time is out of reach of any hobbyist.

But what is happening that creates their emergent behaviors is a black box that we cannot see inside. We also don’t know what consciousness is and we don’t know what has it and what doesn’t other than that which our “feelies” dictate.

What can be called The Illusion of Consciousness is far more easily explained by human cognitive biases, than any reality of LLM functioning. It’s easy to be fooled by an LLM which echoes words back to you, you put meaning into your own words, and you recognize them when they are said back to you.

In the deeper sense, any meaning present in LLM output, is the meaning that the reader puts there and/or the meaning that survives from the training data, it’s not exactly that no meaning present, but it’s not the result of machine cognition, often incoherent and/or very lacking in substance upon close inspection.

I think to some extent this is a lost cause, but your feelings/fears seem to boil down to “this could be conscious!” (emphasis on could) I would invite to see if any other explanation can explain what you observe, is it actually the ONLY or even the most plausible explanation?

I wonder if there’s a market for RevivingYahoo^TM, which in some ways was more link curation than search engine.

When searching for examples of commands, I’ve sometimes been resorting to telling google to restrict itself to pdf files of specific websites (example redhat, with sometimes useful commands for server maintenance broadly)

It’s a bit sad that it’s not just search that has gone downhill, but new documentation too (I swear Microsoft Azure documentation is getting worse).

The quirkiness of the Tranquility Calendar is mildly beautiful, between the second man on the moon getting snubbed with the every four year leap day (which for no good reason still happens around Feb 29!), [I guess Collins gets even more snubbed by not appearing at all] to the more diverse than expected list of honored “scientists/scholars”, until you realize the main reason is just to keep the ABCD naming pattern for the months.

I was thinking “Hey I want that markdown!” (although maybe not for time magazine), but the actual example is dystopian.

If you want to stare at the abyss it’s achievable with curl -H 'user-agent: ClaudeBot' https://url_to_article, and it’s like 70% SEO crap tailored to brainwash LLMs (more realistically “agents” in limited context).

At my place It has reached the stage of: use AI to "improve" your work, be careful about budget though!

I think he falls under the category of “does too much all the time” and “vaguely techo optimist” which definitely put him at risk. I’m not too sure how harshly I should judge him yet, I guess it depends on what he does next.

You’ve gotta love falling for the reverse 1 grain of sand shtick. I don’t think anyone anyone serious would deny that a large amount of sand is a heap.

Encyclical from the pope about the dangers of AI, mostly sane actually: (provided link skips quite a bit about social justice and referencing previous literature)

https://www.vatican.va/content/leo-xiv/en/encyclicals/documents/20260515-magnifica-humanitas.html#Artificial_intelligence

EDIT, snippets:

107. We cannot be satisfied with […] the so-called “alignment” of AI […] without […] openly discussing the ethical frameworks involved and subjecting them to shared standards of social justice. Otherwise, those who control AI will impose their own moral vision, which will become the invisible infrastructure of these systems. A more moral AI is not enough if that morality is determined by a few. […]

117. [ about post and transhumanism ] From the perspective of the Church’s Social Doctrine, the key issue is not the use of technology as such, but the vision that underlies it. If the human being is treated as something to be perfected or surpassed, it becomes easier to accept that some lives are less useful, less desirable or less worthy. In the name of progress, “necessary sacrifices” may begin to be justified, placing the burden on the most vulnerable in pursuit of a supposed optimization of the species. […]

Also what is even the meaning of adding citations if you never came across the source?

[I vaguely understand that there is a lot of window-dressing that comes with writing papers, but this is too perverse.]

The dreaded midnight hour of oblivion!

Agreed, agreed.

EDIT: Though as far as ambiguous anarchist utopias go, I think I’d rather live on Anarres in “The Dispossessed”, even though the material welfare and personal freedoms are much much lower.

You’ve gotta love finding fault with “not preserving heritage” over “imperialistic complete lack of democracy”.

There are some amazing justifications from many amongst the red-pushing side:

• “But if everyone presses red, nobody dies!” (As if that would every happen. Funnily enough strong overlap with the group that claims that “< 90 IQ can’t reason about hypotheticals”, although that is also just that part of twitter.)
• “People who press blue are just blackmailing us!” (I think this accounts for a large portion, ie: not liking to depend on others).
• “The number of people choosing blue can’t be that high! (It would be lower in a true-stakes scenario!)”
• [Many others, but these are those that come to mind.]

It’s a bit baffling how many strongly they refuse the “blue-selection” as possibly moral/rational. Even so far as calling people pressing blue evil or subhuman, simply baffling.

My dad was a bit freaked out by a video version (We’re not ready for super-intelligence)of the “AI 2027” paper, particularly finding two end scenarios a bit spooky: colossus-style cooperating AIs taking over the world, and the oligarch concentration of power one, which i think definitely echoed sci-fi he watched/read as a teen.

In case anyone else finds it useful these are the “Comments as I watch it”, that I compiled for him

Before watching Video Notes:

• AI Only channel with only 3 videos

• Produced By “80000hours”, which is an EA branch (trying to peddle to you the best way to organize 40years * 50 weeks * 40 hours [I love that they assume only 2 weeks of holidays]); which is definitely cult adjacent: https://80000hours.org/about/#what-do-we-do. Mostly appears to be attempting to steer young people to what they believe are “High impact” jobs.

Video Notes:

• The backing paper is a bit of a joke, one “AI 2027”, for reference one of the main authors is very much a “cult member”, Scott Alexander Siskind, author of “Slate Star Codex” and “Astral Codex Ten”.

• Other authors include [AI Futures Project] :

  • Daniel Kokotajlo (podcast co-host of siskind, ex open-ai employee, LessWrong/EA regular)
  • Thomas Larsen (ex MIRI [Machine Intelligence Research Institute = really really culty], LessWrong/EA regular)
  • Eli Lifland (LessWrong/EA regular)
  • Romeo Dean (Astra Fellowship recipient = money for AI Safety research, definitely EA sphere)

• A lot of fluff trying to hype up the credentials of the authors.

• AGI does not have a bounded definition.

• They are playing up the China angle to try and drum up jingoistic support.

• Exaggerating Chat GPT-3 success, by merely citing “users”, without mentioning actual revenue, or actual quality.

• Quote:

  How do these things interact, well we don’t know but thinking through in detail how it might go is the way to start grappling with that.

  -> I think this epitomises the biggest flaw of their movement, they believe that from “first-principles” it’s possible to think hard enough (without needing to confront it to reality) and you can divine the future.

  -> You can look up “Prediction Markets”, which is another of their ontological sins.

• I will note that the prediction of “Agents” was not a hard one, since this is what all this circle wants to achieve, and as the video itself points out it’s fantastically incompetent/unreliable.

• Note: This video was made before the release of GPT-5. We don’t know precisely how much more compute altogether GPT-5 truly required, but it’s very incremental changes compared to GPT-4. I think this philosophy of “More training” is why OpenAI is currently trying (half-succeeding half failing) to raise Trillions of dollars to build out data-centers, my prediction is that the AI bubble bursts before these data centers come to fruition.

• Note: The video assumes keeping models secret, but in reality OpenAI would have a very vested interest in displaying capability, even if not making a model available to the public. Also even on consumer models, OpenAI currently loses a bunch of money for every query.

• Note: The video assumes “Singularitarianism”, of ever acceleration in quality of code, and that’s why they keep secret models. I think this hits a compute/energy wall in real life, even if you assume that LLMs are actually useful for making “quality” code. These ideas are not new, and these people would raise alarms about it with or without current LLM tech.

• Specific threats of “Bio-weapon”, which a priori can not really be achieved without experimentation, and while “automated” labs half exis, they still require a lot of human involvement/resources. Technically grad students could also make deadly bioweapons, but no one is being alarmist about them.

• Note: “Agent 2” Continuous Online learning is gobbledygook, that isn’t how ML, even today works. At some point there are very diminishing returns, and it’s a complete waste of time/energy to continue training a specific model, a qualitative difference would be achieved with a different model. I suspect this sneakily displays “Singularitarianism” dogma.

• Quote:

  Hack into other servers Install a copy of itself Evade detection

  -> This is just science-fiction, in the real world these models require specialized hardware to be run at any effective speed, this would be extremely unlikely to evade detection. Also this treats the model as a single entity with single goals, when in reality any time it’s “run” is effectively a new instance.

• Note: This subculture loves the concept of “science in secrecy”, which features a lot in the writings of Elizer Yudkowsky. Which is cultish both in keeping their own deeds “in a veil of secrecy”, and helpful here when making a prophecy/conspiracy theory, by making the claim hard to disprove specifically (it’s happening in secret!)

• Note: Even today Chain-of-thought is not that reliable at explaining why a bot gives a particular answer. It’s more analog to guiding “search”, rather than true thought as in humans anyway. Them using “Alien-Language” would not be that different.

• Agent 3, magically fast-and-cheap, assuming there are now minimum energy requirements. Then you can magically run 200,000 copies of. magically equivalent to 50,000 humans sped up by 30x. (The magic is “explained” in the paper by big assumptions, and just equating essentially how fast you can talk with the quality of talking, which given the length of their typical blog posts is actually quite funny)

• Note: “Alignment” was the core mission of MIRI/Eliezer Yudkowsky

• Note: Equating Power and Intelligence a lot (not in this video, but in general being suspiciously racist/eugenicist about it), ignoring the material constraints of actual power [echo: Again the epitomical sin of “If you just think hard enough”]

• Note: Also assuming that trillions of dollars of growth can actually happen, simultaneously with millions losing their jobs.

• I am betting that the “There is another” part of the video is probably deliberately echoing Colossus.

• The video casually assumes that the only limits to practical fusion and nanotech just intelligence (instead of potential dead-ends, actually the nanotech part is a particular fancy of theirs, you can lookup “diamondoid bacteria” on LessWrong if you want a laugh)

• The two outcomes at the end of the video are literally robo-heaven and robo-hell, and if you just follow our teachings (in this case slow-downs on AI) you can get to robo-heaven. You will notice they don’t imagine/advocate for a future with no massive AI integration into society, they want their robo-heaven.

• Quote:

  None of the experts are disagreeing about a wild future.

  -> I would say specifically some of them are suggesting that AGI soon is implausible quite strongly. I think many would agree that right now the future looks dire with or without super-AI, or even regular AI.

Takeaway section:

Yeah this really is a cult recruitment video essentially.