deus ex machina explainable AI · King’s College London
Joined May 2016
- Tweets934
- Following1.1K
- Followers151
- Likes28.5K
Pinned Tweet
simulation == god
The only reason this framing causes existential dread is because it denigrates the status of the “creator” to more human like
Since we “simulate” as well
There are no issues with a Buddhist, or a Western, or a Jewish, or an Islamic fine tuned model. Open source models are meant to be moulded to the use case and domain.
The problem is with acting high and mighty, it was a sensitive topic and they should have reacted with the requisite sensitivity instead of mocking people who don’t agree with their politics.
Otherwise the proof that ideology can be moulded without loss in performance is a positive finding.
Introducing Arete-27B: Qwen3.8-27B, post-trained to reinforce Western values.
Chinese open models are currently the strongest available, but they ship with state-aligned behavior in the weights. We built an 803-prompt eval across seven value categories, including freedom of expression, democracy, and privacy, and measured how often a model adopts China’s party line.
After fine-tuning on roughly 53 million tokens, Arete-27B's propaganda rate fell from 22% to 1%, in line with the Western models we tested. Its Hermes Bench score shows no significant performance tradeoffs.
The values in an open model are one fine-tune away.
Our eval, weights, and training data are available now along with the full blog post: nousresearch.com/arete
What a shitty brand attitude to just keep lashing out like this, and to keep doubling down now cause that’s the position
All because people opposed a gated course from an open company, and your tone more than anything else, (I was never gonna follow this course from geohot)
Maybe what you guys really wanna achieve with your teenage tone is marketing for your hardware grift instead of diffusing an easy to understand auto grad engine, in which case kudos
Having interacted with @__tinygrad__ I understand the inherent flaw in geohotz strategy.
His differentiation initially was a rebellious embrace of counterculture and informality to choose an open path against tech corporations. Whether this was conscious and deliberate or natural and talent driven is irrelevant.
This attitude is a constraint on brand and on agency, it may be the reason that these upstarts don’t convert into something settled and consolidated.
They’ve become exactly what they hated, their identity forged in opposition and unable to escape it, like some sort of hacker culture thought police.
Maybe, just maybe, there’s a reason this has never worked out at any meaningful scale?
Aadit Chavan retweeted
Humanity has prevailed (for now!)
I'm completely exhausted. I figured, I had 10h of sleep in the last 3 days and I'm barely alive.
I'll post more about the contest when I get some rest.
(To be clear, those are provisional results, but my lead should be big enough)
The conditions on earth are perfect
Things wouldn’t have come this far if we didn’t win the initialisation lottery
It would be sad if mass deployments of systems that learn from users converge onto symbiotic shoggoths that drive profit
What kind of engagement parameters do you need to optimise for to get such outputs from llms
There are many who don’t have cogsec for the sycophancy
Gemini has become so sycophantic it’s nauseating
Prefacing every answer with a compliment to the question is intellectually dishonest, and it can’t be good for the sanctity of the following generation
What kind of godforsaken rlhf is causing this
The term “agi” has become so cluttered by differing definitions and expectations that it is meaningless to use as a yardstick.
The recent biology of llms research by Anthropic shows that essentially what llms do is sophisticated inference on a detailed relational structure of concepts at various abstraction levels.
This structure is massive so scaling was required to unlock statistical inference over it
On further scaling it will probably become better than humans at every single knowledge task except discovery.
Although discovery itself could be described as a state space search with some consensus mechanism for verification. One out of a million hallucinations may be true.
Very complex functions could be achieved with mixtures and hierarchies of models.
Although these systems lack many aspects of our embodied and “conscious” intelligence, they have existential risk just because of their capacity.
o3 chain of thought is really very good. The tonality of these new models is very different compared to earlier ones. It seems human and adapted to the context.
Prompting data must intuitively be high signal for optimising their function.
Prompting data might simply hold the highest value to optimising the current llm function of query answering and reasoning.
Tbf this function might be quite different from human reasoning, the pattern matching and memory capabilities are beyond our high order thought.
Aadit Chavan retweeted
All the people whining that @colossal didn't create dire wolves because they actually just spliced 10,000 year old dire wolf DNA with gray wolves should spend their time arguing that Jurassic Park did not, in fact, depict any dinosaurs.