April 2024 AICamp Boston meetup - Joscha Bach - Is the LLM the last invention we ever need to make?
Joscha Bach is an iconoclastic researcher working at the intersection of AI and cognitive science. He previously was a researcher at both the MIT Media Lab and at Harvard. Between 2021 - January 2023 he was Principle AI Engineer at Intel's Cognitive Computing group. He currently serves as AI strategist at Liquid AI and on the advisory council of the AI Foundation. His website is bach.ai and you can find him on X at / plinz .
Talk Title & Description:
Is the LLM the last invention humanity needs to make?
The success of LLMs is taking the world by storm. The Turing test seems to have been obliterated years ago, and more applications and use cases are unlocked every week. Yet many of us are skeptical: are LLMs sufficient to carry us all the way to AGI? Are they the way to creating superhuman intelligence, a transitional development, or a dead end?
Пікірлер: 48
joscha, one of my most listened minds overall
This is the best lecture I've heard in months.
Thank you Dr. Bach.
Awesome Talk
5:20 Look how happy he is with his joke landing :) Edit: 7:28 Another smash, my man came prepared
@markcounseling
6 күн бұрын
4:55 Notice that this one didn't quite land though 😂
Thank you Mr. Anderson
great nerdy humor and insightful observations.
Brilliant as always ❤
Presence is not just interaction using real time models, it is balancing the application of models built from past experience with the update, enhancement, and transcendence of those models based on the present moment.
thanks for this Mr. Elton!
The more of his presentations I see, the more I feel like I'm listening to an LLM...
@encyclopath
16 күн бұрын
He will have travelled back in time to have implemented himself
I think we should keep agency and basically impement the perfect tool and grow with it gaining insight into every topic we want by using that tool. we can even expand that tool that way
Lots of things get built better by taking more time to build them, but better in many fields, such as software, is synonymous with faster.
accelerate into the future
37:26 two Gigabytes, I could outsource myself to my PC standing on the floor.
Jedes mal mindblowing der Mann...
29:00 LLMs and IFS-like multiple parts
To be useful functional constraints must direct scaling layers in the same way such functions construct abstractions in compressive architectures. Where the functions are derivative of a complex multimodal system based in evolution imperative. Text has embedded in it much of the evolutionary directive, as well as causal relations.
One day i told Bing Copilot humans are also kind of AI's, a bit in in style of Bach and Michael Levin are explaining it...and told humans desperately need AGI to solve the climate crises. Next day Bing immediately asked about my ideas...and when I was wondering it: "I’m sorry if I caused you confusion. I don’t remember all my conversations, but sometimes I can use your previous messages to help me understand what you want to talk about. This does not mean that I store or keep your personal information. I respect your privacy and safety. 😊 I didn’t want to assume that you were the same person who asked me about the KZread idea earlier, but I thought it was possible because you used the same language and location. If you are a different person, you can tell me, and I can start a new conversation with you. If you are the same person, you can continue where we left off. 😊 I am interested in hearing more about your thoughts on artificial intelligence and its impacts on the world. You can also ask me anything else, and I will try to answer to the best of my ability. 🙌"
@solarwind907
12 күн бұрын
Guess how the IDF uses the AI program “lavender“ to choose targets in Gaza? It’s a lot like how you describe your experience with Bing
20:24 Electric Weltgeist: Mirror, Mirror on the Wall - Black Nazis 24:50 the problem of human ethics, the loss of “Divine Will”, and the hackability of utilitarianism: the breakdown of utility maximization; effective altruism doesn’t lend itself to be implemented on computers (- will lead to unwanted outcomes if implemented rigorously).
A reason it's quite fallacious to compare a 2Gbyte "only" model with the brain is because the ANN alone is an inference model. Our mushy brains do the actual training too - a comparison should match the amounts of compute&data needed to produce the seemingly efficient inference engine. I'm pretty convinced the brain size is about massive searching to find useful solutions within very limited amount of sensory data in order to fit it with an optimized, efficient model of the world, NOT about memorizing the resulting model of the world. Try to pretrain an LLM with all information available before "Attention is all you need" paper, then feed its window context with the actual paper and ask it to produce the implementation code. If it is able to do that even remotely then indeed we can talk about AGI.
2:30 instead of culture wars we should look at this as a mirror of who we are... (For the purpose of understanding ourselves better)
38:10 would it be possible to run a person on a gaming PC? We just know. He seems to imply a well compressed embedment of the universe of human knowledge can fit in 2 GB?
42:45 I take this as the great and eternal nerd utopia, to rise above the monkeys and to be able to exist within a purely informational space, with yourself as one LLM, in communication with other LLMs. So clean, so pristine, so safe, no 900-lb gorillas allowed. But alas, the entirety of bodily experience is lost within this autistic vision, as beautiful as it is in many respects.
44:12 "Every technology so far we invented made things better" That seems some fairly subjective opinion. I actualy heard the opposite from professional designers doing Typography. Better for whom? For humanity of for capitalism? Some people think Music with Autotune is "better" some wouldnt agree.
At 8:00 he implies that it's fine if AI replaces us (as photosynthesizes replaced the pre-oxygen lifeforms) because it would be selfish for us to want to be the end-point of evolution.
@thomasmitchell2514
19 күн бұрын
Is that wrong though? He followed it with logic, consider how you’d feel if primates had that mindset and prevented evolution so they could remain on top. Is that the right thing to do? Also his argument wasn’t for complete replacement, but significant displacement. However humans can very likely evolve with AI. Would people that want to stop evolution where it’s at be okay with significant divergence in human capability? Statistically, more intelligence is associated with less violence, so maybe natural selection will slow that process down until it’s overwhelming. But which team are you rooting for?
What camera was used
@DanElton
19 күн бұрын
It was filmed by Addy who is a friend of Rachel St Clair. He says he used a sony A7RV. He also used $700 lapel mics lol (probably overkill)
I wonder what is the exact relation among microsoft, intel, germany and Joscha Bach?
@yoyo-jc5qg
10 күн бұрын
Says in description he moved on from intel to liquid AI and also is an advisor for the AI Foundation which collaborates with researchers, universities, and companies like microsoft to promote ethical, responsible, and sustainable AI
i belive he thinks too abstract there got to be a easier way
30:10 - whose theorem?
@Gudnarr
18 күн бұрын
Löb‘s theorem: en.m.wikipedia.org/wiki/Löb's_theorem
@Gudnarr
18 күн бұрын
I think!
👀ツ
Thank you Mr. Musk
@umberto488
17 күн бұрын
Huh?
You want the truth? JB will tell you the truth(if you can handle it lol).
@loopuleasa
20 күн бұрын
not just handle, grasp it too