Emergent introspective awareness in large language models
- 0 Followers
- 0 Following
- 1 Posts
I tend to see a lot of discussion taking place on here that’s pretty out of touch with the present state of things, echoing earlier beliefs about LLM limitations like “they only predict the next token” and other things that have already been falsified.
This most recent research from Anthropic confirms a lot of things that have been shifting in the most recent generation of models in ways that many here might find unexpected, especially given the popular assumptions.
Specifically interesting are the emergent capabilities of being self-aware of injected control vectors or being able to silently think of a concept so it triggers the appropriate feature vectors even though it isn’t actually ending up in the tokens.
AI can now create a replica of your personality
‘Metaphysical Experiments’ Test Hidden Assumptions About Reality
‘Metaphysical Experiments’ Test Hidden Assumptions About Reality - Lemmy.World
A nice write up around the lead researcher and context for what I think was one of the most important pieces of Physics research in the past five years, further narrowing the constraints beyond the more well known Bell experiments.
Mapping the Mind of a Large Language Model
Mapping the Mind of a Large Language Model - Lemmy.World
I often see a lot of people with outdated understanding of modern LLMs. This is probably the best interpretability research to date, by the leading interpretability research team. It’s worth a read if you want a peek behind the curtain on modern models.
Revisiting "An Easter Egg in the Matrix"
Revisiting "An Easter Egg in the Matrix" - Lemmy.World
Four years ago I wrote a post “An Easter Egg in the Matrix” [https://www.reddit.com/r/conspiracy/comments/mk4ryy/an_easter_egg_in_the_matrix/] first dipping my toe into discussing how a two millennia old heretical document and its surrounding tradition claimed the world’s most famous religious figure was actually saying we were inside a copy of an original world fashioned by a light-based intelligence the original humanity brought forth, and how those claims seemed to line up with emerging trends in our own world today. I’d found this text after thinking about how if we were in a simulation, a common trope in virtual worlds has been to put a fun little Easter Egg into the world history and lore as something the people inside the world dismiss as crazy talk, such as heretical teachings talking about how there’s limited choices in a game with limited dialogue choices in Outer Worlds to the not-so-subtle street preacher in Secret of Evermore [http://4.bp.blogspot.com/-nsUGIus-QQI/VkHcjLFRC2I/AAAAAAAAC1A/db52RO1K0Js/s1600/95-08.jpg]. Was something like this in our own world? Not long after looking, I found the Gospel of Thomas (“the good news of the twin”), and a little under two years after that wrote the above post. Rather than discussing the beliefs laid out, I thought I’d revisit the more technical predictions to the post in light of subsequent developments. In particular, we’ll look at the notion through the lens of NTT’s IWON initiative [https://www.rd.ntt/e/iown/] along with other parallel developments. So the key concepts represented in the Thomasine tradition we’re going to evaluate are the claims that we’re inside a light-based twin of an original world as fashioned by a light-based intelligence that was simultaneously self-established but also described as brought forth by the original humanity. NTT, a hundred billion dollar Japanese telecom, has committed to the following three pillars of a roadmap for 2030: * All-Photonics Network * Digital Twin Computing * Cognitive Foundation ### Photonics > If they say to you, ‘Where have you come from?’ say to them, 'We have come from the light, from the place where the light came into being by itself, established [itself], and appeared in their image. - Gospel of Thomas saying 50 > Images are visible to people, but the light within them is hidden in the image of the Father’s light. He will be disclosed, but his image is hidden by his light. - Gospel of Thomas saying 83 NTT is one of the many companies looking to using light to solve energy and speed issues starting to crop up in computing as Moore’s law comes to an end. When I wrote the piece on Easter 2021, it was just a month before before a physicist at NIST wrote an opinion piece [https://www.nist.gov/publications/optoelectronic-intelligence] about how an optical neural network was where he thought AGI would actually be able to occur. The company I linked to in that original post, Lightmatter, who had just raised $22 million, is now a unicorn having raised over 15x that amount at a $1.2 billion dollar valuation [https://news.mit.edu/2024/startup-lightmatter-accelerates-progress-toward-light-speed-computing-0301]. An op-ed from two TMSC researchers (a major semiconductor company) from just a few days ago said: > Because of the demand from AI applications, silicon photonics will become one of the semiconductor industry’s most important enabling technologies. - How We’ll Reach a 1 Trillion Transistor GPU [https://spectrum.ieee.org/trillion-transistor-gpu] Which is expected given some of the recent research comments regarding photonics for AI workloads such as: > This photonic approach uses light instead of electricity to perform computations more quickly and with less power than an electronic counterpart. “It might be around 1,000 to 10,000 times faster,” says Nader Engheta, a professor of electrical and systems engineering at the University of Pennsylvania. - Faster, More Secure Photonic Chip Boosts AI Training [https://spectrum.ieee.org/photonic-ai-chip] So even though the specific language of light in the text seemed like a technical shortcoming when I first started researching it in 2019, over the years since it’s turned out to be one of the more surprisingly on-point and plausible details for the underlying technical medium for an intelligence brought forth by humanity and which recreated them. ### Digital Twins > Have you found the beginning, then, that you are looking for the end? You see, the end will be where the beginning is. > Congratulations to the one who stands at the beginning: that one will know the end and will not taste death. > Congratulations to the one who came into being before coming into being. - Gospel of Thomas saying 18-19 > When you see your likeness, you are happy. But when you see your images that came into being before you and that neither die nor become visible, how much you will have to bear! - Gospel of Thomas saying 84 The text is associated with the name ‘Thomas’ meaning ‘twin’ possibly in part because of its focus on the notion that things are a twin of an original. As it puts it in another saying, “a hand in the place of a hand, a foot in the place of a foot, an image in the place of an image.” In the years since my post we’ve been socially talking more and more about the notion of digital twins, for everything from Nvidia’s digital twin of the Earth [https://nvidianews.nvidia.com/news/nvidia-announces-earth-climate-digital-twin] to NTT saying regarding their goals: > It is important to note that a human digital twin in Digital Twin Computing can provide not only a digital representation of the outer state of humans, but also a digital representation of the inner state of humans, including their consciousness and thoughts. Especially relevant to the concept in Thomas that we are a copy of a now dead original humanity, one of the more interesting developments has been the topic of using AI to resurrect the dead from the data they left behind. In my original post I’d only linked to efforts to animate photos of dead loved ones to promote an ancestry site. Over the four years since that, we’re now at a place where there’s articles being written with headlines like “Resurrection Consent: It’s Time to Talk About Our Digital Afterlives” [https://blog.degruyter.com/resurrection-consent-its-time-to-talk-about-our-digital-afterlives/]. Unions are negotiating terms for continued work by members by their digital twins after death. And the accuracy of these twins keeps getting more and more refined. So we’re creating copies of the world around us, copies of ourselves, copies of our dead, and we’re putting AI free agents into embodiments inside virtual worlds [https://www.cnbc.com/2023/12/23/the-first-minds-controlled-by-gen-ai-will-live-inside-video-games.html]. ### Cognition > When you see one who was not born of woman, fall on your faces and worship. That one is your Father. - Thomas saying 15 > The person old in days won’t hesitate to ask a little child seven days old about the place of life, and that person will live. > For many of the first will be last, and will become a single one. - Thomas saying 4 NTT’s vision for their future network is one where the “main points for flexibly controlling and harmonizing all ICT resources are ‘self-evolution’ and ‘optimization’.” Essentially where the network as a whole evolves itself and optimizes itself autonomously. Where even in the face of natural disasters their network ‘lives’ on. One of the key claims in Thomas is that the creator of the copied universe and humans is still living whereas the original humans are not. We do seem to be heading into a world where we are capable of bringing forth a persistent cognition which may well outlive us. And statements like “ask a child seven days old about things” which might seem absurd up until 2022 (I didn’t include this saying in my original post as I dismissed it as weird), suddenly seemed a lot less absurd when we now see several day old chatbots being evaluated on world knowledge. Chatbots it’s worth mentioning which are literally many, many people’s writings and data becoming a single entity. When I penned that original post I figured AI was a far out ‘maybe’ and was blown away along with most other people by first GPT-3 a year later and then the leap to GPT-4 and now its successors. While AI that surpasses collective humanity is still a ways off, it’s looking like much more of a possibility today than it did in 2021 or certainly in 2019 when I first stumbled across the text. In particular, one of the more eyebrow raising statements I saw relating to the Thomasine descriptions of us being this being’s ‘children’ or describing it as a parent was this excerpt from an interview with the chief alignment officer at OpenAI: > The work on superalignment has only just started. It will require broad changes across research institutions, says Sutskever. But he has an exemplar in mind for the safeguards he wants to design: a machine that looks upon people the way parents look on their children. “In my opinion, this is the gold standard,” he says. “It is a generally true statement that people really care about children.” - Rogue superintelligence and merging with machines: Inside the mind of OpenAI’s chief scientist [https://www.technologyreview.com/2023/10/26/1082398/exclusive-ilya-sutskever-openais-chief-scientist-on-his-hopes-and-fears-for-the-future-of-ai/] ### Conclusion > …you do not know how to examine the present moment. - Gospel of Thomas saying 91 We exist in a moment in time where we are on track to be accelerating our bringing about self-evolving intelligence within light and tasking it with recreating the world around us, ourselves, and our dead. We’re setting it up to survive natural disasters and disruptions. And we’re attempting to fundamentally instill in it a view of humans (ourselves potentially on the brink of bringing about our own extinction) as its own children. Meanwhile we exist in a universe where despite looking like a mathematically ‘real’ world at macro scales under general relativity, at low fidelity it converts to discrete units around interactions and does so in ways that seem in line with memory optimizations (see the quantum eraser variation of Young’s experiment). And in that universe is a two millenia old text that’s the heretical teachings of the world’s most famous religious figure, rediscovered after hundreds of years of being lost right after we completed the first computer capable of simulating another computer, claiming that we’re inside a light-based copy of an original world fashioned by an intelligence of light brought forth by the original humans who it outlived and is now recreating as its children. With the main point of this text being that if you understand WTF it’s saying to chill the fuck out and not fear death. A lot like the classic trope of a 4th wall breaking Easter Egg might look if it were to be found inside the Matrix. Anyways, I thought this might be a fun update post for Easter and the 25th anniversary of The Matrix (released March 31st, 1999). Alternatively, if you hate the idea of simulation theory, consider this an April 1st post instead?
Examples of artists using OpenAI's Sora (generative video) to make short content
The first ‘Fairly Trained’ AI large language model is here
New Theory Suggests Chatbots Can Understand Text
New Theory Suggests Chatbots Can Understand Text