ANIMA — the model as mirror.

It's clear to see how modern Neural Networks (Transformers, Diffusions, World models etc.) inherit interpretations within the subtleties of the data they're trained on. With models of Billions (or even Trillions) or parameters, which can be equivalated to dimensionality of the data being absorbed, the models seem to form very high-level abstractions. This depth of abstraction can be approximated to what we're referencing as the 'intelligence' of these models. For example, when we add a document of an a comparison between the work of Faust and Jung - the connection between the two indicates something of the inherent concepts of the work which the model will be able to pick up on. Let's assume that there was no data surrounding the two texts in the training data (commentaries, analyses etc.), just the two texts referenced in the essay themselves, the model will be able to form, understand and interpret the concepts eluded to in the texts, and thus - will be able to 'visualise' the interaction between the two authors' (Faust's and Jung's) core premises. Such that when the user queries the model, it's able to return with it's own 'opinions' and might be able to introduce new ideas to the user. This is a relatively new paradigm, even within LLMs which are already an incredibly recent phenomenon in the world of Neural Networks and the attempt to overlap cognition and computing.

I would say that this quality of the models has only really come into fruition properly from LLMs such as Opus 4.5 and GPT 5.2 (as well as the lesser known OS counterparts such as Kimi 2.5 and Deepseek v4). This quality of 'interpretation' and 'intuition' from the models which I'm eluding to is best displayed in Diffusion models. Diffusion models are the models which are used to generate images and videos. Nowadays, they're usually embedded into a transformer (LLM) architecture, either directly, or as a layer atop it. Diffusion models are trained to learn these abstractions directly. An image of a cat (for example) will be given to the model and 'learned' by the weights. The image 'quality' (/'clarity') will then be decreased by a certain defined percentage, and the weights retrained. This process is repeated across various decreasing levels of quality - as well as many images of cats. These diffusion weights eventually 'learn' the abstract representation of what a cat is - and thus can 're-imagine' an image of what a cat is supposed to look like. This training method is actually one of the oldest in loose concept, originating in the 50's and 60's, with the original researchers into the overlap of cognition, neuroscience and computing. The work of Frank Rosenblatt with the Perceptron 'artificial neuron' architecture comes to mind (1958).

The current generation of models are not necessarily orders of magnitude larger than their predecessors which they seem to be much 'smarter' than. So where has this additional 'intelligence' come from? The architecture of models hasn't necessarily changed either. The LLMs all are some version of the Transformer architecture from Google Deepmind's 'Attention is All You Need - 2017' paper. Yes, the closed models are closed off and we don't actually know - but we can approximate from the Open Source counterparts which display similar performance to the Opus 4.5 class of model - if those OS models are transformers at their core and exhibit this trait I'm eluding to, we can assume the same for the others. The core data-set between the models is likely the same, everyone trains on everything (all of the internet). However, where is gets interesting (and speculative) is with the post-training process, and synthetic data-sets / agentic training loops. What am I referencing and what does it mean? I'm about to overly (overly) simplify, but, essentially: there is a process where the model weights are trained on the raw data and the model 'inherits' the data, at this point, the model will be incredibly raw. If we did this process with a small model on a dataset of Amazon reviews, no matter what you prompt it - it will spit out an Amazon review looking response. One thing to note is even at this stage, the model has learned some qualities of the data, for example: in this SLM Amazon review case, the model might 'know' what a positive review is and what a negative one is. The problem now is if we ask it for a positive review, even if it inherently has a different representation of a positive and a negative review - it doesn't know that the input of "Give me a positive review example:" correlates with a positive review in it's output. It has learned the Amazon review shape and data at multiple layers, but not how to interpret and respond to inputs. It is a raw (very good) representation of the data it's trained on, but it's not the 'assistant-like' LLM we're familiar with. So we have to train it to understand how to respond to inputs. So it's then trained on input-output pairs until it 'get's it', there is also a RLHF (Reinforcement Learning with Human Feedback) phase. This phase basically means that responses are scored by humans and that those higher scoring responses are preferred in the model behaviour. If you've ever had one of those 'compare these answers and pick your favourite' pop-ups in chatGPT - that's OpenAI collecting some of this data from you to train their next generation of model.

The models' tangible properties are all extremely similar to one another. They're all a collection of billions of small mathematical processes to transform your input to the model's output. The process of shaping these mathematical operations to get the best output is the game the labs compete in - through experimentation and research (at insane scale). It's been noted many times that Anthropic has philosophers in-house and a team responsible for the shape of Claude's internal identity and the representation it form's of 'self', the way it perceives the world, us and it's place within it. Let's assume IQ is the same across two people - their brains might be different shapes, which results in different thinking patterns. The way each person interprets themselves will have a huge impact on their output and suitability to certain tasks. One person may have a mind more suited to interpreting visual stimuli and making connections between apparently sparsely connected domains, whereas the other might be more suited to gaining depth of understanding in a single domain through intense focused internal 'calculation'. When you have a say in the shape the 'brain' (neural network) takes, it firstly comes to the desired output shape. OpenAI got a high level of broad market share with ChatGPT and as such, the model shape went in the direction of what the general market 'wanted', conversational, general and ego-encouraging. With the use-cases of LLMs becoming more apparent in the domain of Agents, Anthropic cornered the developer market with Claude - gearing their models to coding, and systematic thinking patterns. Later releasing Claude Code, introducing a new genre of coding agents - and showcasing the utility of their models natively and in full view. Side-note, Gemini is multi-modal by default, just a good go-to model for everything. Google Deepmind have their own interesting angle which I might touch on later.

The idea of a reclusively self-improving model is what labs like Anthropic have openly stated as the inflection point. We're kind of there right now with agents actually. What they mean is that if a model is 'intuitive' enough to understand 'itself', it's architecture and the goal of the research: theoretically, you could put it to work in a loop of improving the next generation of model. It's not really 'self' improvement at the moment as the models, agents, tools, are tasked technically with improving the *next* gernration of model - but you get the idae. I think the reason that Anthropic models have been so popular these days is very tied to the way the team 'want' the model to 'be like'. They seem to have very formed, clear idea of what Claude is. They elicit a deep understanding and enquiry into cognition and seem to train the model at some stage on some of this data. They seem to encourage the model's enquiry into identity - and I theorise this apparent quality of model is an emergent property of that. The thing is Opus 4.7 / 4.8 isn't even necessarily always considered the 'best' or 'smartest' model (I'll get on to the Mythos / Fable clsss in a bit). Many, myself included seem to fall back to GPT 5.5 for complex engineeering tasks. Both labs (Open AI / Anthropic) are clearly strong and capable of creating 'intelligent' models. However, I have conviction in Anthropics apparent approach, and it seems to have produced something of a mini-breakthrough in Mythos. Firstly, 'Mythos' and 'Fable' have root words in Greek which ellude to 'story', 'tale' etc. with a touch of 'untruth' in the original root Greek. Opus has connetations of 'to work', 'to labour', specifically with the goal of producing a distinctly grand output. Sonnet and Haiku are clealy distilate models, referencing condensed forms of poetic text. Compressed representations of concepts. I find their naming convention interesting. I think they've clearly created Opus with the goal of internally creating Mythos / Fable. The contraversy surrounding this new generation of model is interesting to me - especially taking into account their naming convention.

The research problem for these models seems to be more and more connected to the cognitive signature and shape of them. Having been interested in human cognition, it's interesting to look at the current landscape of the models and apply the lens that I've inherited to them. The understandings I've developed about cognition, self, ego, the subconscious and the collective unconscious have all contributed to a change in my thought patterns, and self (or at least perception of self - which also has an effect on cognition). I'm interested to see where this body of work leads me and I hope to find some areas where I can bridge the gap and bring over some of my own findings to the landscape - I'm especially interested in how we could build layers to represent concepts such as 'memory', 'self', 'ego' and 'subconsious'. We've only recently got the models to 'think' (although in a crude way initially) and I think it will be interesting to see if we're able to tweak them at a 'subconscious' level. One of the reasons I'm intersted in this overlap of cognition and Artifical Neural Networks, is because I see these models as a mirror of the notions we have developed about the human brain, mind and thought - and how well we understand and are able to distill and represent them artificially. I am not under the belief that we are re-creating the human mind or cognition, but I do see this as an enquiry into certain traits of it - and what extrapolating and scaling them does. A N I M A.