TRICS are all you need

It is human, not machine intelligence, that drives artificial intelligence

Many AI experts predict that human level intelligence will be reached within a few months and supra human levels of attention will soon follow. These forecasts have some people demanding protective regulations and others planning how they will buy AI agents to replace their human employees. Physicians are concerned that AI will take over their practice. This concern is nonsense, based on a misunderstanding or misrepresentation of just what GenAI models are. Most of their success at specific tasks is driven by human design, not computational intelligence.

A paper by Goh et al. was interpreted to say that an LLM by itself provided more accurate diagnosis than a physician using the LLM or a physician using other computational resources. But that is not what they found. What they actually reported was that an LLM plus expertly constructed prompts (“Using established principles of prompt design, we iteratively developed a 0-shot prompt”) was more accurate than an LLM with inexpertly constructed prompts. The condition that they called LLM-only was more accurately the LLM guided by expertly prepared and iteratively improved, prompts. The LLM was common between the two groups, so it could not have been the use of an LLM that made the difference. The intelligence was supplied by the prompt designers, not the language model. Without the expert prompts, the model added nothing to physicians’ diagnosis. Attributing the diagnostic success to the LLM, then, is misleading.

In fact, almost all of the intelligence exhibited by AI models is due to human intelligence. Humans design the systems, they choose the representations, they provide the training data, they specify how to measure the outcome, and do almost everything else the model needs to appear intelligent. Imagine a parent who holds up a 4-year old girl so that she can reach a basketball hoop and dunk the ball. The girl successfully scores the basket, but that does not make her a shoo in to join the WNBA in a few months. The parent does the hard work and all that is left is for the girl to do the final bit of dropping the ball into the hoop.

That is how AI models work. Humans do the difficult tasks to simplify what the computer has to do. Together they achieve a level of intelligence, but that intelligence is almost completely driven by the human.

Lachter and Bever (1988) called attention to the representations designed into neural network models in the context of learning the past-tense forms of English verbs. Rumelhart and McClelland (1986) proposed that a simple associative neural network was sufficient to learn English past-tense transformations (transforming “walk” into “walked,” “swim” into “swam”) . They claimed that the network learned to transform words into the past tense in about the same order that children did. Lachter and Bever, on the other hand, pointed out that Rumelhart and McClelland chose a phonemic representation of the words that incorporated most of the difficult parts of the problem. Their representation, “Wickelphones,” marked word boundaries (which are not obvious when listening to spoken English) and also included other phonemic information that was not available to children. The model succeeded not because it learned the linguistic relations, but because Rumelhart and McClelland provided it with those relations. Once the language had been phonemically represented, simple associative learning would then be enough to finish the solution.

Lachter and Bever called the representations that would structure the problem for an easy solution TRICS, The Representation It Crucially Supposes. The problem extends much further, however. Most of what we think of as artificial intelligence is actually due to human intelligence. Like the parent who holds up the little girl, people have been holding up computer systems, simplifying the problems that they deal with until they can be solved by simple algorithms. Progress in artificial intelligence is partly illusory—lately it consists mostly of improvements in fluency—and substantially due to progress in how people structure solutions to the problems used to assess artificial intelligence.

Contribution of humans

• Training data

• Number of neural network layers

• Types of layers

• Connection patterns

• Activation functions

• Training regimen for each layer

• Number of attention heads

• Parameter optimization method

• Context size

• Representations of words as tokens and vectors

• Training task

• Selection of problems to solve

• Training progress measures and criteria

• Human feedback for reinforcement learning

• Rules for modifying parameters as a result of human feedback

• Prompt

• Temperature and other meta-parameters

Contribution of machine

• Parameter adjustments through gradient descent

Once the neural network is set up according to human designs and the data are provided, the only remaining work for the computer to do is to adjust its parameters. Those parameters (but not their values) and the relations among them are completely determined by the structure provided by human intelligence. The typical method for parameter adjustment is gradient descent. The learning algorithm adjusts the parameters by successive approximation to meet the human-specified training criteria.

As with other AI applications, the models can exceed humans at some tasks simply by having a larger capacity to hold data, better computational capacity, and more consistency. The number of tokens used to train GenAI models now exceeds a human’s exposure to language by several orders of magnitude. So, there is ample opportunity to match a context to a similar training pattern and stochastically mimic a similar output.

Parameter adjustment is not a negligible, but it can only work within a framework of parameters. So far, that framework must be provided by humans, and therefore, there is no chance that the model could exceed, beyond the advantages noted above, the capabilities of the humans on specific tasks or expand their capabilities outside of the given frameworks. There is no chance that these models would support an out-of-control explosion of machine intelligence because they are so dependent on human input.

Given the heavy contribution of humans to the success of current models, it is misleading to attribute that success to the intelligence of the model directly. It would be a mistake to think that expansion of these models in the absence of human innovation will lead to self-increasing capabilities.

Current models are limited to problems selected by humans that can be solved through gradient descent, but not every problem can be. They are general only in the sense that some humans have figured out how to cast many nominal problems into a guess-the-next-word framework. In that sense, these models only solve one problem—guess the next word. Humans see distinctions among the nominal tasks that computationally do not exist.

Artificial General Intelligence cannot be achieved until computer models can select for themselves the problems that they will tackle and design their own methods for solving them.