Denis Mongin
@denismongin.bsky.social
Started with lasers and nonlinear optics, now with biomedical informatics, data and statistics in health sciences.
Research in #Metaresearch, #rheumatology and #publichealth at HUG and University of Geneva.
#Rstats ninja, #Openscience
Research in #Metaresearch, #rheumatology and #publichealth at HUG and University of Geneva.
#Rstats ninja, #Openscience
Just for the sake of it, a reminder that iceberg do not look like that, because they would not float like that.
Go try it:
joshdata.me/iceberger.html
Go try it:
joshdata.me/iceberger.html
November 30, 2025 at 6:51 PM
Just for the sake of it, a reminder that iceberg do not look like that, because they would not float like that.
Go try it:
joshdata.me/iceberger.html
Go try it:
joshdata.me/iceberger.html
simple but powerfull !
Example: using 3LLM agreing among 4 reach accuracy at the human gold level while having an answdêr for 90% of the abstracts !
The plus: It is robust to LLM performance!
With "bad" LLMs, we obtained less agreement and thus reject more abstracts, but kept the accuracy
⬇️7/8
Example: using 3LLM agreing among 4 reach accuracy at the human gold level while having an answdêr for 90% of the abstracts !
The plus: It is robust to LLM performance!
With "bad" LLMs, we obtained less agreement and thus reject more abstracts, but kept the accuracy
⬇️7/8
November 18, 2025 at 9:30 AM
simple but powerfull !
Example: using 3LLM agreing among 4 reach accuracy at the human gold level while having an answdêr for 90% of the abstracts !
The plus: It is robust to LLM performance!
With "bad" LLMs, we obtained less agreement and thus reject more abstracts, but kept the accuracy
⬇️7/8
Example: using 3LLM agreing among 4 reach accuracy at the human gold level while having an answdêr for 90% of the abstracts !
The plus: It is robust to LLM performance!
With "bad" LLMs, we obtained less agreement and thus reject more abstracts, but kept the accuracy
⬇️7/8
We then systematically tested combination of N LLM (N between 3 and 7), and and varied the "enough" of "we keep the answer when enough LLM agree".
The higher the "enough", the less abstract with an answer, but the higher the accuracy of the answers.
⬇️6/8
The higher the "enough", the less abstract with an answer, but the higher the accuracy of the answers.
⬇️6/8
November 18, 2025 at 9:30 AM
We then systematically tested combination of N LLM (N between 3 and 7), and and varied the "enough" of "we keep the answer when enough LLM agree".
The higher the "enough", the less abstract with an answer, but the higher the accuracy of the answers.
⬇️6/8
The higher the "enough", the less abstract with an answer, but the higher the accuracy of the answers.
⬇️6/8
Ok, I know bluesky is not AI/LLM friendly,
and I actually agree with most of the reasons why.
But hear me out: we propose here is to use small open-weight LLM (<10B), that you run locally (RTX 3090), to extract information or classify text.
A task the LLM are really good and efficient at.
⬇️2/8
and I actually agree with most of the reasons why.
But hear me out: we propose here is to use small open-weight LLM (<10B), that you run locally (RTX 3090), to extract information or classify text.
A task the LLM are really good and efficient at.
⬇️2/8
November 18, 2025 at 9:30 AM
Ok, I know bluesky is not AI/LLM friendly,
and I actually agree with most of the reasons why.
But hear me out: we propose here is to use small open-weight LLM (<10B), that you run locally (RTX 3090), to extract information or classify text.
A task the LLM are really good and efficient at.
⬇️2/8
and I actually agree with most of the reasons why.
But hear me out: we propose here is to use small open-weight LLM (<10B), that you run locally (RTX 3090), to extract information or classify text.
A task the LLM are really good and efficient at.
⬇️2/8
Ho we actually do
November 10, 2025 at 6:44 AM
Ho we actually do
Crazy.
Plus I would so much tend to distrust anything coming from a platform selling the "data you need in less than 2 hours". Without any info about the actual "people" producingthe data.
How is this even accepted?
Plus I would so much tend to distrust anything coming from a platform selling the "data you need in less than 2 hours". Without any info about the actual "people" producingthe data.
How is this even accepted?
October 24, 2025 at 11:43 AM
Crazy.
Plus I would so much tend to distrust anything coming from a platform selling the "data you need in less than 2 hours". Without any info about the actual "people" producingthe data.
How is this even accepted?
Plus I would so much tend to distrust anything coming from a platform selling the "data you need in less than 2 hours". Without any info about the actual "people" producingthe data.
How is this even accepted?
Mmmm
Rather suggests random people from internet (and i did not find any info about these people in the article ) tend to make the error
Rather suggests random people from internet (and i did not find any info about these people in the article ) tend to make the error
October 24, 2025 at 11:29 AM
Mmmm
Rather suggests random people from internet (and i did not find any info about these people in the article ) tend to make the error
Rather suggests random people from internet (and i did not find any info about these people in the article ) tend to make the error
First author doing lab work, increasing chicken weight by 62% (works also for finance, professional success and better relationships)
Looks like a story for @gmilgram.bsky.social
Looks like a story for @gmilgram.bsky.social
October 24, 2025 at 6:19 AM
First author doing lab work, increasing chicken weight by 62% (works also for finance, professional success and better relationships)
Looks like a story for @gmilgram.bsky.social
Looks like a story for @gmilgram.bsky.social
Quand on sait qu'elle possède à priori pour 600'000 euros de bijoux qu'elle a oublié de déclarer, c'est assez drôle
October 19, 2025 at 5:17 PM
Quand on sait qu'elle possède à priori pour 600'000 euros de bijoux qu'elle a oublié de déclarer, c'est assez drôle
October 9, 2025 at 6:45 PM
J'espère que le plupart des gens trouvent au contraire cela rassurant pour l'état de droit
September 25, 2025 at 12:06 PM
J'espère que le plupart des gens trouvent au contraire cela rassurant pour l'état de droit
Thank you conseil fédéral of #switzerland, it is the right moment to cut scientific funding by 10% in 2026 and 2027.
Perfect move, perfect timing.
No planetary problem to be solved,
no energetic problem coming,
science is absolutely not under threat anywhere in the world,
so I think it is ok.
Perfect move, perfect timing.
No planetary problem to be solved,
no energetic problem coming,
science is absolutely not under threat anywhere in the world,
so I think it is ok.
September 18, 2025 at 7:36 AM
Thank you conseil fédéral of #switzerland, it is the right moment to cut scientific funding by 10% in 2026 and 2027.
Perfect move, perfect timing.
No planetary problem to be solved,
no energetic problem coming,
science is absolutely not under threat anywhere in the world,
so I think it is ok.
Perfect move, perfect timing.
No planetary problem to be solved,
no energetic problem coming,
science is absolutely not under threat anywhere in the world,
so I think it is ok.
it works with latest version!
Thanks. I guess it was related to the issue you pointed out
Thanks. I guess it was related to the issue you pointed out
September 1, 2025 at 6:46 PM
it works with latest version!
Thanks. I guess it was related to the issue you pointed out
Thanks. I guess it was related to the issue you pointed out
it is in red in the first paragraph
The gardian has been doing a fair job compared to others like bbc
The gardian has been doing a fair job compared to others like bbc
September 1, 2025 at 4:43 PM
it is in red in the first paragraph
The gardian has been doing a fair job compared to others like bbc
The gardian has been doing a fair job compared to others like bbc
mmmm
if so, I should get the same here, but I do not.
There is something more I think.
if so, I should get the same here, but I do not.
There is something more I think.
September 1, 2025 at 4:37 PM
mmmm
if so, I should get the same here, but I do not.
There is something more I think.
if so, I should get the same here, but I do not.
There is something more I think.
They are the same.
September 1, 2025 at 2:36 PM
They are the same.
Dear {marginaleffects} #Rstats specialists:
I am finally reading the whole marginaleffects.com book from @vincentab.bsky.social . Super didactic, I love it.
I have one question in chapter 5:
Why are these two different ?
I am finally reading the whole marginaleffects.com book from @vincentab.bsky.social . Super didactic, I love it.
I have one question in chapter 5:
Why are these two different ?
September 1, 2025 at 2:06 PM
Dear {marginaleffects} #Rstats specialists:
I am finally reading the whole marginaleffects.com book from @vincentab.bsky.social . Super didactic, I love it.
I have one question in chapter 5:
Why are these two different ?
I am finally reading the whole marginaleffects.com book from @vincentab.bsky.social . Super didactic, I love it.
I have one question in chapter 5:
Why are these two different ?
I have been super impressed by your maps, especially this one:
There is a lot to learn, thank you for sharing your work
There is a lot to learn, thank you for sharing your work
August 28, 2025 at 5:39 PM
I have been super impressed by your maps, especially this one:
There is a lot to learn, thank you for sharing your work
There is a lot to learn, thank you for sharing your work
Je redoute en plus (mais je peux avoir tord) que la sévérité soit en fait proportionnelle au diamètre, et non à l'aire, parce que l'aire est 2 fois plus grande pour 2003 que 2025.
Bref, beaucoup de choses améliorable dans ce graph qu'on voit partout, avec ses variations...⬇️
Bref, beaucoup de choses améliorable dans ce graph qu'on voit partout, avec ses variations...⬇️
August 20, 2025 at 8:13 PM
Je redoute en plus (mais je peux avoir tord) que la sévérité soit en fait proportionnelle au diamètre, et non à l'aire, parce que l'aire est 2 fois plus grande pour 2003 que 2025.
Bref, beaucoup de choses améliorable dans ce graph qu'on voit partout, avec ses variations...⬇️
Bref, beaucoup de choses améliorable dans ce graph qu'on voit partout, avec ses variations...⬇️
3)la sévérité. Non définie, mais l'aire est proportionnelle à la sévérité, avec une aire reference de 1.
1 quoi?
Mystère...⬇️
1 quoi?
Mystère...⬇️
August 20, 2025 at 8:13 PM
3)la sévérité. Non définie, mais l'aire est proportionnelle à la sévérité, avec une aire reference de 1.
1 quoi?
Mystère...⬇️
1 quoi?
Mystère...⬇️
Du coup la aussi entre ces trois cercles, lequel est le plus ancien en regardant les couleurs ? A t on une idée de la différence d'année ?...⬇️
August 20, 2025 at 8:13 PM
Du coup la aussi entre ces trois cercles, lequel est le plus ancien en regardant les couleurs ? A t on une idée de la différence d'année ?...⬇️
2) la palette couleur est mal choisie, et non indiquée ! On nous met deux point extrême, et pas les couleurs intermédiaires, qui varient peu...⬇️
August 20, 2025 at 8:13 PM
2) la palette couleur est mal choisie, et non indiquée ! On nous met deux point extrême, et pas les couleurs intermédiaires, qui varient peu...⬇️