In a text compressor, the model is updated after each prediction. Of course
you could just turn off updates entirely. For a large model it would be
awhile before the prediction accuracy starts to degrade.

At the physical level, reading and writing a bit both cost kT ln 2 energy
(a few zeptojoules at room temperature) because reading memory means making
a copy.

-- Matt Mahoney, [email protected]

On Thu, May 15, 2025, 11:54 AM James Bowery <[email protected]> wrote:

>
>
> On Mon, May 12, 2025 at 11:33 AM Matt Mahoney <[email protected]>
> wrote:
>
>>
>> Training and prediction costs about the same.
>>
>
> In the limit of a large number of predictions per model-revision that is
> false for the obvious reason that predictions are conditional
> decompressions and the conditions (ie: "prompts") overlap hence the
> predictions overlap -- so tabling/memoization/caching not only works, but
> is an inevitable cost savings.
> *Artificial General Intelligence List <https://agi.topicbox.com/latest>*
> / AGI / see discussions <https://agi.topicbox.com/groups/agi> +
> participants <https://agi.topicbox.com/groups/agi/members> +
> delivery options <https://agi.topicbox.com/groups/agi/subscription>
> Permalink
> <https://agi.topicbox.com/groups/agi/Tdc5c19d0f38aacd6-M9ed2badb276cdd06da046713>
>

------------------------------------------
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Tdc5c19d0f38aacd6-Maa98a03c28681e6a2038d8c0
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Reply via email to