Peter B

Cars 2 is insane, I am amazed it even got made and released.

Good blog post, this helped clarify some things for me.

My concern is we don't have answers to many of the open questions or limitations even with vanilla LLMs, so throwing a bunch of them together in more complex ways seems... dumb?

FreeCAD just released 1.0 and it feels perfect for my level of amateur CAD skill. And free is a big benefit!

I played around with Rhino3d and Grasshopper for some generative art stuff a while ago, but it was mostly functional and never really to understand how to actually use it effectively.

My latest project was gridfinity-ing this mini toolbox, I am thinking of doing a set of video tutorials on using FreeCAD to build custom gridfinity boxes because I've learned quite a bit!

I have the fewest problems with `google/gemma-2-9b-it`, I find it particularly good at instruction following combined with JSON mode for structured generation.

Definitely don't look up Gridfinity and then get addicted to 3D printed modular organization 😉

LLMs struggle with perception, not reasoning, in ARC-AGI What made o3 so much better than previous models on this benchmark?

LLMs struggle with perception, not reasoning, in ARC-AGI by Mikel Bober-Irizar

What made o3 so much better than previous models on this benchmark?

anokas.substack.com/p/llms-strug...

The article discusses how the AI ethics community's focus on "fairness" significantly limits how it approaches and addresses algorithmic harm, and proposes reframing harms in terms of domination and oppression per Iris Marion Young's framework.

firstmonday.org/ojs/index.ph...

oh okay.

Somehow it keeps getting worse.

Chatgpt4 has paragraphia.

nononononono

aaaaahhhhhhh

I'm sorry, this has to be the dumbest study with the dumbest framing. This is an actual sentence from the summary:

'Moreover, as in humans, age is a key determinant of cognitive decline: “older” chatbots, like older patients, tend to perform worse on the MoCA test.'

goblin.tools/About is the one LLM "product" that really hits the sweet spot for me in terms of a useful and specific application of generative AI. Is there more stuff like this? I just love the idea of a collection of simple, task-specific tools with a basic interface.

We Looked at 78 Election Deepfakes. Political Misinformation Is Not an AI Problem.

In fact, @sayash.bsky.social and I have just published an essay with them, where we play our usual role of looking at the evidence and tamping down AI hype and fears instead of playing them up.
knightcolumbia.org/blog/we-look...

(Cross-posted to AI Snake Oil aisnakeoil.com/p/we-looked-...)

I encountered this drafting a presentation for work. I opted for calling them "task-oriented" or "narrow" models, but I would have much preferred "discriminative" if not for the negative connotation.

It's been interesting to witness in real-time how the usage of "algorithm" in many places has shifted from a neutral "sequence of instructions" to a negative "controlled ordering and boosting of information".

I just asked Gemini, ChatGPT 4o, and Claude this exact question and they all gave me a warning about this. I'm pretty sure the model in the screenshot was baited or the initial Gemini release - it's very easy to cherry pick these.

I would say overall it was a success in both the flavor and easy prep! We'll make both these again.

If you're trying this yourself, my tips are:
- Ask it for modifications to make prep easier
- Make sure to ask it to "make it flavorful" or something similar
- Double check the suggested cookware

We got here by:
- Asking for 5 recipes from each
- Picking one recipe from each that sounded the best
- Asking it to reflect and scale the recipe for 6 people and make sure it was "flavorful" (we've had very bland ChatGPT recipes in the past)

ChatGPT bungled scaling the recipe up to 6 people. The portions/ingredients were right, but it thought we could fit 1.5 cups cooked quinoa, an onion, and 6 cups kale into a large pan. Fortunately I read ahead of time and pulled out the dutch oven.

As far as flavor, they were really a tie. Both had unique, interesting flavors and were new dishes for us.

Claude's recipe was better. The simplicity helped, but it also gave us a "Make-Ahead Tip" to make the whipped feta a day in advance. That was great for planning.

Claude's Za'atar Roasted Cauliflower with Whipped Feta

Stuffed Acorn Squash with Quinoa, Kale, and Goat Cheese

This Thanksgiving we pitted ChatGPT against Claude in battle of the side dishes.

Claude gave us Za'atar Roasted Cauliflower with Whipped Feta and ChatGPT gave us Stuffed Acorn Squash with Quinoa, Kale, and Goat Cheese. Which was better? 🧵

YouTube video by Posit PBC A Future of Data Science - posit conf 2024

In August I had the pleasure of presenting a talk at posit::conf, called A Future of Data Science, in which I assert that data science exists because statistics missed the boat on computation.
The video is up now...
www.youtube.com/watch?v=YKMZ...

I had the same experience, it didn't really click with me.

We're back!