Reply to thread

Message

<blockquote data-quote="Gorgon Zee" data-source="post: 9294742" data-attributes="member: 75787">“Randomly” should not be thought of as synonymous with “completely at random”. Most modern data algorithms work “randomly” but their outputs converge to a state that is definitely not completely random. GenAI is no different; the process of convergence from a very random state to something that is useful is indeed “random”, but weighted towards outputs that are consistent with the input data the model was trained on.Essentially, the majority of machine learning and AI algorithms want to produce outputs that could hide in the input and not be recognized as different. Even a basic linear regression model does the same. If you feed it data on heights and weights of people as training data, then if you give it a weight, it tries to produce an output that would look plausible given the inputs. When you ask a GenAI model to produce text for an input, it does exactly the same thing — pick the most plausible sequence of words. The two differences are:<ul>
<li data-xf-list-type="ul">The linear regression model has 2 parameters that were trained. The GenAI one has 70,000,000,000</li>
<li data-xf-list-type="ul">Linear regression can be solved exactly, and so is deterministic in output. This is not the case for most other modern data models.</li>
</ul></blockquote>

[QUOTE="Gorgon Zee, post: 9294742, member: 75787"] “Randomly” should not be thought of as synonymous with “completely at random”. Most modern data algorithms work “randomly” but their outputs converge to a state that is definitely not completely random. GenAI is no different; the process of convergence from a very random state to something that is useful is indeed “random”, but weighted towards outputs that are consistent with the input data the model was trained on. Essentially, the majority of machine learning and AI algorithms want to produce outputs that could hide in the input and not be recognized as different. Even a basic linear regression model does the same. If you feed it data on heights and weights of people as training data, then if you give it a weight, it tries to produce an output that would look plausible given the inputs. When you ask a GenAI model to produce text for an input, it does exactly the same thing — pick the most plausible sequence of words. The two differences are: [LIST] [*]The linear regression model has 2 parameters that were trained. The GenAI one has 70,000,000,000 [*]Linear regression can be solved exactly, and so is deterministic in output. This is not the case for most other modern data models. [/LIST] [/QUOTE]

Verification