
<div class="card">
<div class="card-body">
<small class="text-muted p-0 m-0">\@jeremyphoward via Twitter on Jul 25, 2023<a href="https://twitter.com/jeremyphoward/status/1684082952085585921"> <svg xmlns="http://www.w3.org/2000/svg" width="16" height="16" fill="currentColor" class="bi bi-link-45deg" viewBox="0 0 16 16">
<path d="M4.715 6.542 3.343 7.914a3 3 0 1 0 4.243 4.243l1.828-1.829A3 3 0 0 0 8.586 5.5L8 6.086a1.002 1.002 0 0 0-.154.199 2 2 0 0 1 .861 3.337L6.88 11.45a2 2 0 1 1-2.83-2.83l.793-.792a4.018 4.018 0 0 1-.128-1.287z"/>
<path d="M6.586 4.672A3 3 0 0 0 7.414 9.5l.775-.776a2 2 0 0 1-.896-3.346L9.12 3.55a2 2 0 1 1 2.83 2.83l-.793.792c.112.42.155.855.128 1.287l1.372-1.372a3 3 0 1 0-4.243-4.243L6.586 4.672z"/>
</svg></a></small>
<p class="card-text">Be careful of concluding that "GPT 4 can't do <x>" on the basis you tried it once and it didn't work for you.

See the thread below for two recent papers showing how badly this line of thinking can go wrong, and an interesting example.</p>
</div>
</div>