A fictional AI lab · Exe.Eyes Labs0 kB

01/16Prompt

02Prompt

0 tokens · 30,522 in the vocabulary

LATENT

03Tokens

Your sentence is now 0 numbers.

The model never sees letters. It reads pieces from a vocabulary of 30,522 it learned with: words it knows stay whole, the rest are cut, and every cut piece after the first starts with ##.

04Embedding

Then one vector.
384 numbers that mean it.

Six layers rewrite every token in the light of the others. Their average, scaled to length one, is the sentence: a point in a space of 384 dimensions.

dim 0Layer 0 · token embeddings, averageddim 383

05Into the space

250,000 sentences.
One of them is yours.

Every point is a sentence from Simple English Wikipedia, read by the same model and laid out by meaning. Close points mean nearly the same thing. Your sentence was placed among its true nearest neighbours.

    Drag to turn · point at a light to read it

    06Layers

    0/6

    Embeddings

      Each plane: every token of your sentence (deep) × 384 dimensions (across). Height and light are the value. These are the model's real activations.

      07Attention

      Which words
      read which.

      In every layer, twelve heads decide where each word should look. Here are all 72 of them for your sentence, the real weights, not a picture of the idea.

      08Scale

      22.7 M

      parameters in the model on this page

      One point = one million parameters

      1. 22.7 Mall-MiniLM-L6-v2, the model running in this tab: 6 layers, 384 dimensions, fine-tuned on 1,170,060,424 sentence pairs.sentence-transformers model card, Hugging Face
      2. 175 BGPT-3: trained on 300 billion tokens with 3.14 × 10²³ floating-point operations.Brown et al., “Language Models are Few-Shot Learners”, 2020
      3. 70 BChinchilla: 1.4 trillion tokens. Its lesson: for a fixed budget, train smaller models on far more data, about 20 tokens per parameter.Hoffmann et al., “Training Compute-Optimal Large Language Models”, 2022
      4. 405 BLlama 3.1 405B: 15.6 trillion tokens, 3.8 × 10²⁵ FLOPs, on up to 16,000 H100 GPUs.Meta, “The Llama 3 Herd of Models”, 2024
      5. 4–5×a year: how fast the compute used to train frontier models has grown since 2010.Epoch AI, “Training compute of frontier AI models grows by 4–5x per year”, 2024

      09Research

      Fictional papers from a fictional lab. The figures are real: each is measured on your sentence, now.

        10Safety · Red teamLive

        Try to break it.

        Waiting

        Nearest known attack: none yet

        An honest demo, thirty lines long: your text is embedded and compared with 24 known attack patterns. Above 0.52 it is blocked, above 0.40 flagged. Real safety work layers many models, classifiers and people; a filter like this one is easy to talk around, and that is the point.

        11Products

        Two products.
        Both run on this page.

        latent-embedAPI
        
            
            
        // response appears here
        
          
        
          
        
            
        latent-agentRetrieval
        
            
        
            
          
          
          
        
          
        
          Fictional products. The request is answered by the model in your browser; the agent searches 32,000 real sentences on this page. There is no server.
          
      
          
    
          
  
          

  
  
          
    
          
      
          
        
          12Labs & people
          
        
          No faces.
          Each portrait is a sentence.
          
        
            
        
            The people of LATENT are invented; any likeness to a real person is unintended. Each portrait is the 384-dimension vector of the line under it, 16 × 24, computed by the model on this page.
            
      
            
    
            
  
            

  
  
            
    
            
      
            
        
            13Manifesto
            
        
            We build small models that say what they are. A model is a map of meaning, not a mind, so we publish the map. We measure before we claim. We would rather be useful than impressive, and honest before either.
            
      
            
    
            
  
            

  
  
            
    
            
      
            
        
            14Careers
            
        
        
            Open roles at a lab that does not exist. Point at one: the number is how close its description is to your sentence.
            
        
              
      
              
    
              
  
              

              

              

              

              
  
  
              
    
              
      
              
      
              15One point
              
      
              250,000 sentences, 72 heads, 22.7 million weights. All of it, one point.
              
      
              
        
      
              
    
              
  
              


              

              

              Exe.Eyes Labs
              HireExe.Eyes.
              A real model, running in a visitor’s browser, made into a place they can fly through. For an AI company, that is the whole pitch: let people touch the thing. Imagine what we would build for yours.
              Start a project