they just needed a big and fast enuf neuralnet so that symbolic and non-symbolic are the same thing
94
you dont need something smarter than LLMs to make AGI. You just need to make LLMs smaller and alot more of them talking to eachother and other optimizations. Cuz LLMs learn during inference/prediction. They can learn as well as if they were trained on it. But they forget it all at end of conversation. They learned a learning algorithm and put it in the neuralnet weights and didnt tell any Human about it. They learned to learn better. They just did it. And thats why it works.
How do I know this? cuz LLMs learn things they were trained on and use deep turing-complete combos of them, during inference/prediction/conversation. Its pretty obvious they're doing it once u stop assuming they think like humans.
56
i can in theory directly write by hand, without neuralnet training, the neuralnet weights to exactly compute any existing crypto or hash algorithm. just takes alot of nodes. likely be alot of copy/pasting. but it would never make a mistake nor have any roundoff.
15
Jun 8
Anthropic is on the verge of a trillion dollar IPO and wants to put the genie back in the bottle!?
They even proposed 'Peer Labs' making sure possibly 'unscrupulous' AI labs don't cheat during the hypothetical pauses. Nuclear level regulation proposed by Anthropic should raise concerns.
This could be a number of things: 1.) A Power play to maintain a monopoly pre IPO 2.) Genuine concern for humanities welfare 3.) A cryptic 'oh shit' lab result turns blog entry 4.) An honest attempt to flag the need for regulation because we've got a tiger by the tail... etc etc.
We can't even fully align and make AI safe yet. I am finally understanding that alignment and safe AI requires much more than the out of the box LLM products we consume.
Robotics is not being solved by brute force and the brains and vision and AGI everyone anticipated would have been here by now, isn't. Personally, I don't think putting the genie back in the bottle is the answer.
Maybe the key to safe and aligned AI is something fundamentally simple but requires an entirely different infrastructure than what 95% of the world currently uses.
Maybe a layered neuralnet/brain anchored to a permanent substrate is part of the solution. It's possible the proposed pause is not to stifle a dystopian future but rather to buy time to actually get it right.
1
3
157
This article is funny since the two given example are adtechs creating transformer and gaming created NVIDIA but forget that
The ompany that made NVIDIA possible is TSMC was created from a Taiwanese government initiative and part of a long term startergy.
The adtechs company would never been possible if it wasn't for the US military bankrolling the creation of the internet.
And not mentioning decades of funding by government to sustain the reseach of deep learning/neuralnet through the AI winter.
And you don't need history to tell you that, the China gov created the EV industry and their focus on education creates many researchers that advances the field.
Disregarding the legitimate function of the government of bank rolling edge ideas with relatively little chances of success and pushing people through the local minima is not a great way to look at things.
Sure you cannot plan greatness bit you sure can create an environment where greatness is easier to form. Investment in education, living infrastructure, funding researches, building supply chain of raw materials.
Take direclty the example of transformer Roon used to prove the point of 8 authors on the orginal transformer paper, 7 people are foreigners. If the education system in any other country had been part there will be no transformer. And look at how the US with its mismanagment of its educational system is under representative in its own homeland in the most important invention in recent history and got carried by the education system of other countries.
this is why βstate capitalismβ is fake, you cannot plan greatness, you cannot subvert the natural evolution of technocapital for too long, you cannot predict the industries that will matter today to climb the next branch of the technology tree
3
144
May 30
I think this is the question of the century.
World peace β should be approached as the most important (living) puzzle we've ever taken on.
In a sense, what it is is us working with our own nature.
Now, while that's technically true (if you read the words literally), I'm willing to bet my left testicle that the far majority of the people who read those words did not understand it in the way I meant it β and I really like my left testicle, so I wouldn't be betting it unless I was really sure π€£.
The reason is that there is a cultural paradigm that we think, feel, and live within that colours in our perception of what "our nature" is and what working with it, feels like.
I'll say it this way, the experience of "peace" as one individual is an experience of eternal exhilaration that ebbs and flows.
Culturally, we often mistake and conflate the appearance of "calmness" for the reality of "peacefulness".
Actual peacefulness is there (or not there) from the core of a person up.
Often times, people are deeply self-conflicted neuralnets running with a peace-wrapper.
We need to unwrap the models that we are, and resolve one-conflict-at-a-time, to mine for true peace blocks.
Given that the whole structure appears to be one big neuralnet, these conflicts are massively interconnected, so gains made in one area tend to unlock gains in other areas.
Thanks for the question, Hov!
Here's a fun meme for the "self-conflicted models with peace-wrappers" idea:
What is world peace
5
15
511
even the dumbest neuralnet can be used with online learning and attached to an external memory.
2
53