Lightnews — Scholar-powered news

Joseph Lemaitre @josephlemaitre.bsky.social · Aug 19

last time I tried it the system library autocomplete was super slow... not sure if I had a poor config, but I'd say half a second to get the proposal. I was really neat though

1

Joseph Lemaitre @josephlemaitre.bsky.social · Aug 12

just learnt that vs code recently added native git blame decorations with setting:
"git.blame.editorDecoration.enabled": true
good bye complex extensions, goodbye gitlens 👋

Joseph Lemaitre @josephlemaitre.bsky.social · Jul 24

Great read.
If priced out of a bet, the best things is to sell the opportunity to bet to a higher bankroll party (if possible).
There are many cases in life where uncertainty can be traded at discount for certainty (insurance policy is a bet on a home getting flooded, bank loans, ...)

1

Joseph Lemaitre @josephlemaitre.bsky.social · Jul 17

Oh make sense

Joseph Lemaitre @josephlemaitre.bsky.social · Jul 17

I don’t know R but these things are not plain text files?

1

Joseph Lemaitre @josephlemaitre.bsky.social · Jul 15

I've been sneaking on that course (by Sam Abbott, Thomas Robacker, Nick Reich) for the past few days on github (did not see the deployed version) and it is really good learning material for learning epi forecasting: nfidd.github.io/sismid/

1 2

Joseph Lemaitre @josephlemaitre.bsky.social · Jul 14

Does Lucy get a nametag ?

1

Joseph Lemaitre @josephlemaitre.bsky.social · Jul 11

James does really great science ! This is a great job

James Hay @jameshay.bsky.social · Jul 11

We're hiring a modelling postdoc at PSI Oxford for two exciting projects: 1) modelling the early immune responses to Nipah vaccination, and 2) joining the PRESTO team working on immunobridging in vaccine evaluation studies.

tinyurl.com/5abbxrjh

Get in touch for more info! Deadline 4th August.

Job Details

tinyurl.com

2

Joseph Lemaitre @josephlemaitre.bsky.social · Jul 11

This project is supported by the Swiss National Science Foundation @snsf.ch, on a project grant specifically written for this reproducibility project! It’s a great win 🎉

1

Reposted by Joseph Lemaitre

Dorothy Bishop @deevybee.bsky.social · Jul 10

Wow. This is an amazing project with a huge amount of work behind it. Even if you aren't into drosophila, it's worth a look

Mark A. Hanson @hansonmark.bsky.social · Jul 10

The "reproducibility crisis" in science constantly makes headlines. Repro efforts are often limited. What if you could assess reproducibility of an entire field?

That's what @brunolemaitre.bsky.social et al. have done. Fly immunity is highly replicable & offers lessons for #metascience

A 🧵 1/n

14 43

Joseph Lemaitre @josephlemaitre.bsky.social · Jul 10

.. but found Aider (usually Sonnet Architect + DeepSeek implementor) and Claude Code extremely productive.
Task is also important, it is incredible and safe for plotting and front-end... found it slows me down on many pure modeling task. All anecdotal obviously.

2

Joseph Lemaitre @josephlemaitre.bsky.social · Jul 10

yes, very nice someone studied it so well. I am very curious on their hinted follow-up on agent. I always fount LLM clunky, and cursor frustrating (there is some kind of bad context compression or something) ...

1 1

Reposted by Joseph Lemaitre

Mark A. Hanson @hansonmark.bsky.social · Jul 10

But in fact, fly immunity is highly reproducible! 🎊

~80% of claims could be verified. Moreover, some challenged claims just reflect the field advancing its knowledge/tools.

The lesson: if the tools are good and the research largely exempt from direct translation pressures, science works.

3.5/n

A chart depicting the track of claims into verified, unchallenged, challenged, partially verified, or mixed.

Figure 1: Flow of the study.
Claims were first classified in five categories based on literature. 45 unchallenged claims were selected to be experimentally tested and
classified either as challenged or as verified. We further classify some unchallenged as consistent or inconsistent depending on their
consistency with current knowledge.

2 18 49

Reposted by Joseph Lemaitre

Mark A. Hanson @hansonmark.bsky.social · Jul 10

When everything is considered together (w/ caveats), the most prestigious institutes are the most likely to produce irreproducible work. Journal tier was a more minor effect, and in fact, "high-impact" but not "trophy" journals publish the most reproducible science.

But, many caveats (cont.)

4.3/n

A chart showing various factors and their odds ratios and relative contributions to irreproducibility in a multivariate mixed model.

Figure 10. Multivariable analysis of predictors of claim irreproducibility.
The forest plot reports the odds ratio and the 94% highest density interval of each covariate effect on
a claim becoming challenged.

1 10 25

Reposted by Joseph Lemaitre

Richard Sever @richardsever.bsky.social · Jul 2

more than that but right order of magnitude. I always say $10^1 for preprint; $10^3 for journal article; $10^5 for research. most costs are staff not tech because of responsible screening, customer service, etc.

2 3 13

Joseph Lemaitre @josephlemaitre.bsky.social · Jul 1

Seems like it cannot do these little neat interactive javascript diagrams but claude has gotten better at ascii

Joseph Lemaitre @josephlemaitre.bsky.social · Jul 1

I remember you talking about this :D This is great (I get a Error though, only when I am logged with Claude).
Hope you get back the you get a list of the questions

1

Joseph Lemaitre @josephlemaitre.bsky.social · Jun 30

I'm not in epiengage but I'll try with something more unformed ! Squares with LLMs ability in green field projects vs already existing stuff.

1 1

Joseph Lemaitre @josephlemaitre.bsky.social · Jun 30

Oh no, I now have to test twice the hypothesis 😦

2

Joseph Lemaitre @josephlemaitre.bsky.social · Jun 30

interesting, somehow never managed to use LLM for this, they always compromise my ideas/go too far.... I find them useful for nearly everything else (finding flaws in ideas, developing the concept).

Perhaps my notes are too scattered. Or my prompt is bad.

1

Joseph Lemaitre @josephlemaitre.bsky.social · Jun 30

(Obvious disclaimer that LLM use comes with many shortcomings, and world model is something that is hard to define. Just hope for more nuance.)

arxiv.org/pdf/2403.154... is also a good read (yay for independent researchers).

arxiv.org

Joseph Lemaitre @josephlemaitre.bsky.social · Jun 30

Obviously, this last bit also highlights that the world-model is relatively brittle. But still, it exists.

That's for non-coders. For coders, there is a before and after using Claude Code. A lot of the academic literature on LLMs brushes over code (like the big PNAS piece recently forgot who).

1

Joseph Lemaitre @josephlemaitre.bsky.social · Jun 30

Chess is even my pet example to demonstrate how world-models are built by Transformers + Data.
The proportion of legal moves is very low, and a LLM always plays legal moves and quite good ones.

Gary's claim is caused by the chat environment that destroys the context windows for chess-playing.

1

Joseph Lemaitre @josephlemaitre.bsky.social · Jun 30

This opinion (LLMs have no world model) is repeated at every AI track in IDD conferences, and while it's true for a narrow definition of world model, I am sure it does not square with audience experience.
Contrary to what Gary says, LLMs play good (1'800 ELO) chess: dynomight.net/more-chess/.

🎃nisillos @onisillos.bsky.social · Jun 30

"It is not just that LLM’s fail to induce proper world models of chess. It’s that they never induce proper worlds of anything. Everything that they do is through mimicry, rather than abstracted cognition across proper world models."

garymarcus.substack.com/p/generative...

Generative AI’s crippling and widespread failure to induce robust models of the world

LLM failures to reason, as documented in Apple’s Illusion of Thinking paper, are really only part of a much deeper problem

garymarcus.substack.com

1

Joseph Lemaitre @josephlemaitre.bsky.social · Jun 5

Slightly related to what you describe at the IDD afternoon and your comment about complexity

1 1