James Strachan 🏳️‍🌈

5.2K posts

James Strachan 🏳️‍🌈 banner
James Strachan 🏳️‍🌈

James Strachan 🏳️‍🌈

@jamesstrachan

Experienced Humboldt Fellow and Visiting Scientist at UKE in Hamburg, working on coordination and communication in teaching interactions. He/him

Hamburg, Germany Katılım Şubat 2009
373 Takip Edilen333 Takipçiler
James Strachan 🏳️‍🌈
James Strachan 🏳️‍🌈@jamesstrachan·
Our results indicate that LLMs, particularly GPT-4, show impressive performance on standardised tests of Theory of Mind, but that they also show key differences in how they appear to behave 7/7
English
0
0
1
128
James Strachan 🏳️‍🌈
James Strachan 🏳️‍🌈@jamesstrachan·
LLaMA, on the other hand, appeared to show a bias in responding that may explain its superior performance on this one test (6/7)
English
1
0
1
149
James Strachan 🏳️‍🌈
James Strachan 🏳️‍🌈@jamesstrachan·
Our follow-up experiments indicated that GPT's poor performance was driven by an overly conservative stance as it claimed uncertainty despite being able to generate correct answers when probed (5/7)
English
1
0
1
81
James Strachan 🏳️‍🌈
James Strachan 🏳️‍🌈@jamesstrachan·
Very pleased to see this out now Open Access in Nature Human Behaviour We investigated how well different LLMs performed on tests of Theory of Mind and compared their results against human performance (1/7)
Nature Human Behaviour@NatureHumBehav

Systematically testing language models on a broad battery of Theory of Mind tasks with comparison to human data, a study by @jamesstrachan et al. demonstrates human-like performance by AI chatbots such as GPT-4. @ASTOUND_project nature.com/articles/s4156…

English
1
3
5
946
James Strachan 🏳️‍🌈
James Strachan 🏳️‍🌈@jamesstrachan·
GPT-3.5 also performed at or only slightly below human levels on most tests, while LLaMA2-70B performed below human levels The only task where this pattern didn't hold was the faux pas, where GPT models performed significantly worse than humans while LLaMA did better (4/7)
English
1
0
1
81
James Strachan 🏳️‍🌈
James Strachan 🏳️‍🌈@jamesstrachan·
Humans (unsurprisingly) found more complicated tasks more challenging than simpler tasks What was more surprising was how well the LLMs did. Particularly GPT-4, which performed significantly better than humans on 3/5 tests (irony, hinting, Strange Stories, 3/7)
James Strachan 🏳️‍🌈 tweet media
English
1
0
1
84
James Strachan 🏳️‍🌈
James Strachan 🏳️‍🌈@jamesstrachan·
We tested a broad range of Theory of Mind tests, including false belief, irony comprehension, faux pas, hinting, and the Strange Stories We also made sure to test humans and LLMs in as similar a way as possible to ensure a species-fair comparison of their performance (2/7)
English
1
0
1
72
James Strachan 🏳️‍🌈 retweetledi
Max Marschner
Max Marschner@marschner_max·
Excited to share the first paper out of my PhD, now published @CognitionJourn! The joint outcome of work with David Dignath and Günther Knoblich. sciencedirect.com/science/articl… We find new evidence that co-actors represent joint actions on a group level. WE > ME. Thread: 1/
Max Marschner tweet media
English
1
6
15
1.2K
James Strachan 🏳️‍🌈 retweetledi
James Strachan 🏳️‍🌈 retweetledi
nature
nature@Nature·
Nature research paper: Synaptic wiring motifs in posterior parietal cortex support decision-making go.nature.com/42UCYj9
English
2
96
299
62.5K
James Strachan 🏳️‍🌈 retweetledi
Mathieu Charbonneau
Mathieu Charbonneau@matcharbonneau1·
I am very excited to share with you the news that our book ‘The Evolution of Techniques: Rigidity and Flexibility in Use, Transmission, and Innovation’ is coming out in exactly 2 months. mitpress.mit.edu/9780262547802/… 1/5
English
2
23
49
7.2K
dora kampis
dora kampis@dora_kampis·
The saga we tell here says social world’s important, And memories come from it as resultant But there’s much more to grants than captured in verse, So if this was not clear then come and converse, I tell you in prose what in lyrics I’ve shortened.
English
2
0
1
255
dora kampis
dora kampis@dora_kampis·
For a departmental event I wrote a limerick about my grant. In celebration of the project starting Sept 1, and since otherwise it would from now on just sit in one of my "misc" folders, I release it onto the world instead. come for the cringe, stay for the forced rhymes 👇
English
1
0
12
1.4K
Gloucester Quays
Gloucester Quays@GloucesterQuays·
@jamesstrachan @Lindt Hi James, the customer experience manager here, my team and I here today are thrilled that we were able to work with Lindt and make your Granny's day that little bit more special! We hope she had a lovely day out and hopefully we can welcome her back for her next birthday - Alia
English
1
0
1
52
James Strachan 🏳️‍🌈
James Strachan 🏳️‍🌈@jamesstrachan·
Here she is picking out some additional decoy chocolates to protect her new stash Quote of the Day: "If I'd known being 100 was this fun I'd have done it ten years ago."
James Strachan 🏳️‍🌈 tweet media
English
0
0
1
68
James Strachan 🏳️‍🌈
James Strachan 🏳️‍🌈@jamesstrachan·
Thank you to Carol (in the picture above) and her team at Lindt, and to the ladies on the Customer Service desk who all made my Granny's day
English
1
0
0
62