| … | … | @@ -138,7 +138,69 @@ same time, transformers, such as BERT, and now, more generally DL-NN |
| 138 | 138 | LLM's have shown that statistica techniques, taken from corpus |
| 139 | 139 | linguistics, are far more adept at encoding the structure of language. |
| 140 | 140 | |
| 141 | | but also |
| 141 | Thus, part of the problem is to obtain structural factoids. That is, |
| 142 | they need to be perceived, sensed, obtained from the external |
| 143 | environment. This requires the perception of not just individual data, |
| 144 | e.g. a single birth record, but also the perception of the "shape" of |
| 145 | the structure: birth records have a date and a name; but the date may |
| 146 | only appear at the top of the page, with a list of names below it. Some |
| 147 | classes of birth records provvide additional info, others do not. An |
| 148 | extreme form is the obituary: all of us know what to expect; but the |
| 149 | actual informational detail can be spotty or it can provide a wealth of |
| 150 | info. Roughly speaking, no two obituaries have the same "shape", even |
| 151 | though they all have at least a few "jigsaw tab connectors" which can be |
| 152 | joined into the existing weave of genealogical cloth. |
| 153 | |
| 154 | That is, not only must the structural factoids be perceived, using some |
| 155 | sort of sensory system, but also the structure itself must be perceived, |
| 156 | presumably with a sensory system built into a post-processing stage. |
| 157 | That is, the data flow must consist of blobs of data, together with a |
| 158 | description of the encoding of that blob of data. The encoding is itself |
| 159 | the result of an act of sensory perception applied to the data stream, |
| 160 | but at a meta-level: i.e. after preliminary analysis has been done. |
| 161 | Thus, this is a recursive process. |
| 162 | |
| 163 | Figuring out what to perceive is also a learning process: examining and |
| 164 | comparing input steams allows common patterns to be perceived in those |
| 165 | streams. The common patterns are "learned". The common patterns can then |
| 166 | be used to classify the input stream into buckets. Thus, classification |
| 167 | can also be thought of as a sensory organ. It's not instant-on: it takes |
| 168 | a while to come up to speed. Once running, it can provide descriptions |
| 169 | of data. Which perhaps sounds fine, but raises the problem of the |
| 170 | homunculus: who, exactly is perceiving the structural forms that the |
| 171 | classification susbsystem is generating? |
| 172 | |
| 173 | Clearly, there are a number of recursive issues. Transformers side-step |
| 174 | these issues into two ways. The low-level issue of perceiving structure |
| 175 | is swept aside by only asking for the probablity P(X_n | Y_1, ..., Y_k) |
| 176 | of some occurance of X_n conditioned in a set of (earlier) Y's. This |
| 177 | pushes all structural information into the weight matrix, in some opaque |
| 178 | and non-interpretable way. Yes, leaps and bounds in interpretability are |
| 179 | being made: the tens of thousands of employees of the |
| 180 | multi-hundred-billion-dollar AI companies are not sitting on their |
| 181 | hands. This project is not a single-handed reversal of those efforts. |
| 182 | |
| 183 | The second way that transformers avoid the problem of the homunculus |
| 184 | is that they don't: the actual human user of ChatGPT or Claude or Grok |
| 185 | has at least some vague idea of what they wish to accomplish by using |
| 186 | these systems. It remains an open question in neuroscience as to exactly |
| 187 | how humans decide what to do next. |
| 188 | |
| 189 | To recap: the goal of the project is to perceive structure, and to |
| 190 | provide a formal, mathematical, algorithmic description of how structure |
| 191 | is perceived. A subgoal is to boot-strap into perception, i.e. avoid the |
| 192 | homunculus. |
| 193 | |
| 194 | Although the word "structure" here is meant to evoke a sensation of good |
| 195 | olde-fashioned symbolic AI, there is no presupposition of avoiding |
| 196 | neural nets. Indeed, the structure of a given fixed DL-NN is described |
| 197 | symbolicly: its written in ROCm or CUDA, its written in Python or |
| 198 | TensorFlow. These are structural, symbolic languages. Insofar as there |
| 199 | is progress at the meta-level, then it is in the use of vibe coding to |
| 200 | design transformers. Which is indeed an interesting approach. But the |
| 201 | goal here could be rephrased as saying "a system that can vibe-code |
| 202 | itself". And the basic research to acheive this. |
| 203 | |
| 142 | 204 | |
| 143 | 205 | |
| 144 | 206 | Status |