The 🕯️ Candle Test: Can LLMs be tricked into changing their answer to a simple riddle? Turns out most top-of-the-line models—from Claude 3.7 to Llama 3.1—fail at this task, with the sole exception of GPT-4o: chainforge.ai/play/?f=65d5...

A prophecy for programming in the future ---programming has always, and always will be, in a process of change. It is those who embrace the future monstrosity, as Derrida called it, that lead the space. From the end of the paper, "To Write Code," CHI 2020:

NSF has made an official statement regarding the freeze (new.nsf.gov/executive-or...). All grant work associated with DEI (including “accessibility”) “principles and frameworks” is supposed to be “ceased”:

o3 is like a Christmas dream. Most of my daily work involves filling grid patterns with colors, so when this new model drops, it will be game-changing. 🙏

New ChainForge release ✨out now: Generate a table from a prompt, extend rows, add a column like magic! Great for getting test data for getting started.

We added a "add a column" feature: Extend an existing table by adding a new column, and flash-fill. Here, adding synopses of Miyazaki films to an existing table:

Generating a table of test inputs...

Well, that's a wrap. Finished my class notes for the Empirical Methods in HCI course this fall, and it came in at a whopping 152 (!!) pages of content. Still more refining to do before a public release, but happy to share the materials with instructors at this point.

Went to a digital humanities conference and attendee had an interesting question... What LLMs will reply when asked about generalizations of ethnic groups? Here's a test ---GPT3.5 is happy to generalize for North America, not so much in Europe---and Claude Haiku always replies: