Future of Artificial Intelligence

[sodatrain]

I've settled in on using Gemini. It seems to do what I ask it to do well enough. My needs are pretty low key - mostly a search engine replacement. I used it to calculate the volume of some shapes of different size retaining walls the other day. That was neat. The video stuff... Wow. What will the value from that end up being? Mostly entertainment? It seems very expensive to produce and a relatively low value output. Maybe I'm just being ignorant in the novelty of it all.

Google seems to have a huge advantage given the ubiquity of its tools and services. Integrating it into Gmail and Docs makes it so much more accessible and functional to a huge audience with a very low bar.

Who are y'all listening to for thought leadership on AI?

.I've followed Sam Harris for a long time.
.Eliezer Yudkowsky? I'm not smart enough to refute what he says but shit it's scary
. Mo Gawdat of Google Deep mind (formerly of?)

Who are you looking to for expertise?

I don't find myself listening to Sam Altman, Elon, etc. I don't feel like I can trust them to not spin shit.

* Tangential rant...The Diary of a CEO podcast has an excellent host and some great guests. I think it's really good despite my preference for shorter form. But JFC those click-baity titles. Wtf.

[daylen]

One of the pioneers of AI, Geoffrey Hinton: https://www.youtube.com/watch?v=iHCeAotHZa4

For the technical cutting-edge: Andrej Karpathy, Demis Hassabis, and George Hotz.

I like more spiritual stuff like https://www.youtube.com/watch?v=4WToZFALyE4

Though, to really get an idea of what is going on you need to follow people across different ideological pockets. Everyone is biased, including Sam Harris. Being able to hold many possible positions simultaneously is essential. Gary Marcus is like the most anti-AI person in the world right now so he is fun to follow.

David Shapiro's YouTube channel for a more star trekky, optimistic angle.

[daylen]

This text to text, text to audio, and text to video stuff is what gets the populus engaged, but the real benefits I see are in the background having to do with discovering new compute architectures for particularly hard problems in manufacturing, medicine, and nanotechnology. Nothing really changes the calculus of energy sustainability more than dense, cheap, and long-lasting batteries.

[fiby41]

getting people in the doors and locked into their partner cloud provider ecosystem

Incidently GCP has nearly waved off its egress fees starting last month.

[daylen]

Some more well-rounded sources than the ones I posted earlier:

AI Explained - https://www.youtube.com/@aiexplained-official
Dr Waku - https://www.youtube.com/@DrWaku
Machine Learning Street Talk - https://www.youtube.com/@MachineLearningStreetTalk

[sodatrain]

Thx for the recommendations. AI Explained seems really good - just burned thru a few of his recent videos.

[daylen]

Seems like neuroscience is continuously dissolving the boundary between realism and virtualism these days. The body is constantly projecting its own model of the world. Perhaps we have been virtual creatures since our cells could communicate(*). Some feedback is usually considered "more grounded" than other feedback. Like for instance, the pushback of the earth on us falling towards it may seem more "real" than a stock falling in price because we cannot "experience" that (or can we?). Even though they both link concrete realities to stuff happening in our minds (e.g. symbolism and pre-symbolic imagination).

(*) It is becoming more mainstream to consider electrical/cognitive processing being done by all kinds of cells.

[jacob]

AIs in the form of Claude-3 has just gone beyond the average human IQ. I don't know where this was first reported, so I'm just going to give a random link. https://www.linkedin.com/posts/jhammant ... 14944-K3Wf

Before getting too excited, it means that the best AI is currently as intelligent as the average human. This makes sense because it's been trained on a vast pool of humans, which are rapidly converging on average since everybody is online. Still, ... if this number keep increasing to 105 or 110, it might be time to revisit whether the best ideas are better found outside the human species.

[daylen]

I've heard reports on X of PhD researchers claiming that Claude-3 understands their thesis better than any other human. Probably because most other humans didn't care enough. Patience regardless of task is one of the virtues of these machines, unless humans deliberately make them lazy like OpenAI did to ChatGPT.

[Slevin]

The trend is real and positive, as you are saying, but the small numbers and measurements like this are fake and will be for a while. In that some particular cherry-picked benchmarking test scored a high score in this situation. Use them as a trendline and not a real measurement of capabilities. For all we know, the LLMs have been individually fed the answers to the test in the training data, and the numbers are actually artificially low. From what I've read this week (recommend Zvi for a nice weekly roundup on AI and major releases), Claude 3 does seem more coherent and less lobotomized than the rest of the field, which is exciting. It still can't give me contextual and correct answers to specific problems. Vibes, yeah, it's good at vibes.

I'm particularly fond of this tweet from this week: https://twitter.com/patio11/status/1764812608102617457

So I'm wondering what if, hear me out here, everything we've ever said about consciousness and its relationship to the self is fanfic about the human experience, explains nothing in the real world, and will not correctly predict what the tech tree looks like.

.
(Patrick Mckenzie writes his own highly informative and entertaining blog "Bits about Money" where he explains infrastructure and system scale problems in banking and banking / money history).

[daylen]

Sounds like mind-body dualism which seems misaligned with a more modern biological view of cells being both mind-like and body-like. Consciousness seems to be emergent from multi-cellular communications acting as a virtualization over reality. Evidence of this being that consciousness can be turned off with global or local anesthetics applied to cells that continue to function otherwise.

This is essentially the binding problem: do stories or abstractions actually bind to concrete reality? We operate under this assumption whenever we communicate like here on the forum. I can say dog and you can probably imagine a dog. I could say skin a dog and you could probably go skin a dog. Isn't this basically talking about consciousness and its relationship to you skinning a dog?

From an evolutionary perspective does it make much sense for humans and various other animals to invest in energy intensive virtualization capabilities if they didn't in some way bind to real survival outcomes?

[No dogs were skinned in the making of this post]

[sodatrain]

Oh wow. Just watched this and sorta wish I didn't.

It's a brief on the major announcements the last day or two.