A mensagem abaixo pode ser de interesse para membros desta lista.
[]s, JM
---------- Forwarded message ---------
From: Lawrence Paulson
Subject: [isabelle] DeepSeek experiment
Hello all, you may be interested in a little experiment I tried using the
DeepSeek LLM yesterday. I asked whether it could translate a theorem header
from n-dimensional real vectors to abstract Euclidean spaces. My query:
> The following theorem header fixes variable "a" to type "real^'n". Changing
this type to "'a::euclidean_space" requires modifying the header, replacing $
and χ by their analogues for Euclidean spaces.
>
> proposition
> fixes a :: "real^'n"
> assumes "m ≠ n" and ab_ne: "cbox a b ≠ {}" and an: "0 ≤ a$n"
> shows measurable_shear_interval: "(λx. χ i. if i = m then x$m + x$n else x
$i) ` (cbox a b) ∈ lmeasurable"
> and measure_shear_interval: "measure lebesgue ((λx. χ i. if i = m then x
$m + x$n else x$i) ` cbox a b)
> = measure lebesgue (cbox a b)"
On its first attempt it produced something that seemed to be almost right but
clunky and referred to something non-standard. Pushing it a little further, I
got the following, which is far as I can tell is correct. (Isabelle accepts
it.)
> proposition
> fixes a :: "'a::euclidean_space"
> assumes "m ≠ n" and "m ∈ Basis" and "n ∈ Basis"
> and ab_ne: "cbox a b ≠ {}" and an: "0 ≤ a ∙ n"
> shows measurable_shear_interval: "(λx. ∑i∈Basis. (if i = m then
(x ∙ i) + (x ∙ n) else (x ∙ i)) *⇩R i) ` (cbox a b)
∈ lmeasurable"
> and measure_shear_interval: "measure lebesgue ((λx. ∑i∈Basis.
(if i = m then (x ∙ i) + (x ∙ n) else
(x ∙ i)) *⇩R i) ` cbox a b)
> = measure lebesgue (cbox a b)”
Everything was accompanied by quantities of plausible sounding narrative, e.g.
"this version is more abstract and elegant”. A sample:
> This version is more type-safe because:
> • No need to worry about index bounds
> • The basis vectors are first-class entities
> • The notation more clearly reflects the mathematical intent
> Note that in this formulation:
> • Basis is the canonical basis of the Euclidean space
> • ∙ is the inner product (giving the component in a particular basis
direction)
> • *⇩R is the scalar multiplication
> • The sum reconstructs the vector from its components
> Would you like me to explain any part of this formulation in more detail? Or
would you like assistance with proving this version of the theorem?
My impression is that it was grabbing material from a variety of places, in
particular the AFP entry Ordinary_Differential_Equations. But the word
"elegant" doesn't appear there and in fact appears in the entire distribution
+ AFP only 42 times.
I recall doing the precise task above in the past and needing the better part
of an hour to get it right. I didn't try to port the proof (which was long and
complicated) and I would be surprised if it could get that right.
I'm not sure how much more can be accomplished with the free version of
DeepSeek. I have also seen Claude Sonnet 4 swallow entire Isabelle codebases
and summarise them. These tools may already be useful for tasks such as
identifying duplicate material.
Larry
--
LOGICA-L
Lista acadêmica brasileira dos profissionais e estudantes da área de Lógica
<[email protected]>
---
Você está recebendo esta mensagem porque se inscreveu no grupo "LOGICA-L" dos
Grupos do Google.
Para cancelar inscrição nesse grupo e parar de receber e-mails dele, envie um
e-mail para [email protected].
Para ver esta conversa, acesse
https://groups.google.com/a/dimap.ufrn.br/d/msgid/logica-l/CAO6j_Lhv0KBCz0pkfnPuZf%2Ba3H5wLswzYC%3Dwz%2B34fUBUDnJV%3Dg%40mail.gmail.com.
