Victor Geislinger
@victorsvector.com
4.9K followers 200 following 2.5K posts
Combination of machine learning engineer / data scientist and teacher (depending on what hat I'm wearing that day) Focusing on ML/AI @Google (obligatory 'opinions are my own') github.com/MrGeislinger
Posts Media Videos Starter Packs
Pinned
Hi! I'm Victor!
Here's a quick intro :D

These days I focus on ML/AI at Google but have been a teacher, data scientist MLE, etc.

I try to be genuine & positive so when I leave this world it's in a better state than I found it
You can likely find my other online presences on GitHub.com/MrGeislinger
MrGeislinger - Overview
Combination of machine learning engineer / data scientist and teacher (depending on what hat I'm wearing that day) - MrGeislinger
GitHub.com
Best I think people really can do is give compassion & empathy ♥️

Been more or less in all of those problems but my experiences are still unique. And frankly, I don't know how well anything I did in the past would apply to the current state of things
Oh I love it because I literally thought you meant "want" but that actually makes a lot more sense

(new vocab word of the day!)
It's also really funny because I was like 'wow I can imagine this being really great with slides!' and then I saw the pdf at the end lol

I also don't want to be that guy, but I noticed the smallest typo "(as they are often wont to do)"
(but maybe you're going for that human aesthetic 😁)
... and I'm back!!

Wow that was really good— I've been wanting to look at data mesh more closely (I literally have one of the books you mentioned). I really like the idea of federation and feedback loops
So I'm hearing a challenge to read it faster than 30 minutes??

on your marks... get set...
I love that— thanks!!
Definitely moving up my list! (It's available at my library now but I have the Twilight series++ to focus on first lol)
The failures definitely seem more consistent with a ball of relationships over a fundamental 'understanding' of cute concepts that it extracts from

It also more closely matches how neutral nets work compared to something being more emergent with a core concept of how things work
Quick note about the Winograd schema challenge since I say I'm personally not convinced

Some benchmarks show ~95% passing this for some models. But many still question LLMs ability to fully solve. This is in part because we can't tell if it's 'just good at pattern matching' since they still fail
The hard part is defining if enough relationships approximate or are equivalent to a 'world model'. Some folks think LLMs/AI can get there in practice if we make them large enough

Personally, I think 'world model' is one of those concepts that sound concrete but starts crumbling when applying it
It's a good question— for me, I see it as a type of generalization that emerges from using 'known' concepts (not necessarily physical)

Answering what does 'it' refer to in 'the trophy didn't fit in the suitcase since it was too big' requires understanding relative size
some more pictures of the situation
Black hose on roof amongst gravel with end in the puddle of water. Hose leading off the roof to a ladder to the ground. A window sits behind ladder Office with a hole in the roof where water is being caught by a larger plastic bin.
sooo the roof started leaking *bad* in my office

But no worries, we have a siphon going so maybe this will relieve some of the puddle forming above me?
(thanks physics for the assist)
I'm w/ my kids to get a small toy after our family vaccination trip, and after witnessing a complete freakout until ~5s after getting their shots & them being perfectly fine, seeing how it really is just anticipation of the needle

Going to pretend it's metaphor for something but don't know what yet
I've been on this earth for over 3 decades & I still don't understand why the option for hand dryers to blow into your face exists
White hand dryer with black instructional drawing to first dry your hands and second a face looking at the heat by moving metal valve with hand
What a great observation! 😆😂😂
<eos>
That's awesome! Congrats on the monthiversary!!

You know you're at the right place when coworkers are begging for my stat toks lol
Maybe you were looking for 'resistive heating' but it's still kinda weird calling it that for something like saying my computer is 100% efficient in hear generation
Ooh that looks like it'd be fun to read (especially for today)

Would love to know what you think but for now putting it on my list!
📚📑
Not going to deny that future developments could produce more capable models that we could say exhibits theory of mind or a world model

But the skeptical scientist in me says it's not yet convinced current models are there (or even close) based on the evidence where models have difficulty
Good reminder to check your local library to see if they have have free access to @oreilly.bsky.social's digital catalog

(you won't be surprised to hear that every Bay Area library I've checked has the O'Reilly catalog available for free)
The Illustrated Guide to AI Agents

New book announcement!

Thrilled that together with @maartengr.bsky.social , we're writing a new book titled “An Illustrated Guide to AI Agents” and published by @oreilly.bsky.social.
Always loved O'Reilly having early releases in their digital library! Really helps when tech is evolving sooo rapidly!

Super curious to see the illustrations for all the different concepts :D
Totally with the **may** here. Haven't been convinced LLMs have a 'world model' like we think of it based on where we see it 'fails'

I find it interesting from these papers seeing connections across different modalities. But I also would think due to language focus it'd pass Winograd first, if ever
Personally, I think the word "clearly" is doing a lot of heavy lifting here XD

LLMs definitely have pattern matching that we would find consistent with a world model but they also fail in some ways that we suspect a 'world model' would have little problem with (like the Winograd schema)
things just got interesting

I still hold up LLMs having trouble with the Winograd schema challenge (en.wikipedia.org/wiki/Winogra...) as the counterexample of them building a 'world model'. But these papers do seem to suggest the opposite is not out of the question
Winograd schema challenge - Wikipedia
en.wikipedia.org