Ross Bates

Panta Arithmos
CTO @ Worlds

[Ross Bates @rossbates]

Explainability is the killer feature of any AI product. You can't reliably reproduce a result you can't explain.

[Ross Bates @rossbates]

We keep trying to build machines that think like we do, using assumptions about how we think that were wrong to begin with. You are not agent plus model, you are agent as model.

[Ross Bates @rossbates]

@machinesports @MikeSirois @mikelikessports LFTP!

[Ross Bates @rossbates]

@jkeatn @ulkar_aghayeva The OG of scientific humility.

[Ross Bates @rossbates]

Full post...

https://t.co/F5ZrdmR2GX

[Ross Bates @rossbates]

Neural networks discover structure. Symbolic methods traverse it. Both camps are blind to the other.

Why it was always both.

[Ross Bates @rossbates]

@fchollet Calling it a "replacement for NNs" repeats the mistakes the original symbolic AI crowd made by assuming there's one process for everything. NNs discover structure, symbolic/inference processes traverse structure. Both a version of search.

[Ross Bates @rossbates]

@ebarenholtz If I'm understanding correctly, we agree that the usage of language by itself indicates nothing beyond the signal. And, if we are applying this heuristic fairly and equally, then this must apply to humans as well. Words are a messaging protocol.

[Ross Bates @rossbates]

@ebarenholtz If you accept bats have sonar qualia from a sense we don't have, why would machine inputs be different? By demonstrating they can handle images, audio, and text using the same embeddings, LLMs have shown that language is not a materially unique or privileged representation.

[Ross Bates @rossbates]

a penny for your thoughts (rate limits apply)

[Ross Bates @rossbates]

The nightmare scenario is AI companies have discovered that "smart enough you don't leave + dumb enough you keep talking" is where max profit lives.

[Ross Bates @rossbates]

@mitsuhiko substrate (n): the thing underneath, which I'm hoping you understand, because I don't

[Ross Bates @rossbates]

@signulll @_moonaki_ Unfortunately accountable to investors can mean it's only a matter of time until they push you to monetize your data. To be clear, I love the idea and want you to prove me wrong with a creative technical solution.

[Ross Bates @rossbates]

@signulll @_moonaki_ I know you are just trying to lighten the mood with lol, but without identity you have no accountability. Shit goes wrong, you disappear, and that's the end.

[Ross Bates @rossbates]

@daniel_mac8 @synfinner Are you saying that through the process of habituation people are then unable to notice when their partners habits change?

Getting used to behavior makes it easier to detect outliers, seems that would be the evolutionary advantage in the first place.

[Ross Bates @rossbates]

@jjacky Not questioning your understanding. Just pointing out that rules and similarity aren't collapsible. If you are literally just looking for fulltext+vector, use something like Qdrant multi-stage.

[Ross Bates @rossbates]

@Kev96790724 @NotSpacewear @MoeCanDoIt great mindset, which requires patience because you’re forced to reckon with how much better frontier models are.

And it’s hard not to be frustrated. It’s best to be inspired by what we know is possible, and build/maintain data sovereignty instead of trying to reclaim later.

[Ross Bates @rossbates]

@Kev96790724 @NotSpacewear @MoeCanDoIt Agree, it’s all the teams messing around with parameters and system prompts trying to make Cowork usable for nontechnical people.

Those of us who actually understand how to use the model are collateral damage while they try to juice the user numbers for an IPO.