Einstein i llms

Does do you need to reach the next level of artificial intelligence?

We’ve heard a lot about this in the past year – we saw a crisis perceived in the laws of scaling, where motorists and jokes in the industry were concerned that we would end the power to keep the lifting systems.

But we also had a lot of debate about what constitutes the general artificial intelligence – when we can say we have reached that point, and what it means.

A great example of this debate comes in the form of an essay written by Thomas Wolf, apparently partly in response to Dario Amodei’s Grace’s ‘Grace Machines’.

This thought was included in the podcast of Brief Brief, where Nathaniel Whittemore spoke about its premises and consequences for the future.

Einstein and good student

One thing we know from history is that Einstein was not excellent in school.

The theorist and a prominent mathematician had trouble getting into the classical form of education on his day. He did not seem to be someone who had them all together, and Wolf suggests that it is really the norm.

“History is full of geniuses that fight during their studies,” Wolf writes. “Edison was called” added “by his teacher. Barbara McClintock was criticized for the ‘strange thought’ before winning a Nobel Prize. Einstein failed in his first attempt at the ETH Zurich entrance exam. And the list continues.”

It is a mistake, he claims, to think that you can simply escalate good students and get a genius intellect.

I liked Wolf’s analogy saying that Copernicus “went against his training data group” suggesting that the Earth orbits the sun – yes, we know that now, but at the time, it was a revolutionary idea.

“To create an Einstein in a data center, we do not only need a system that knows all the answers, but rather one that can ask questions that no one else has thought or dared to ask,” Wolf writes. “One who writes” What if everyone is wrong about it? “When all textbooks, experts and ordinary knowledge suggest differently. Simply consider the crazy change of the paradigm of particular relativity and intestines that it took to form a first axiom as ‘let’s assume that the speed of light is constant in all reference frameworks’ by opposing the common meaning of (those) days.”

Mentioning Jennifer Doudna and Emmanuelle Charpentier and their work in Crispr, he noted how thinking outside the box can mean the use of a product designed for radically different use cases.

These rare paradigm shifts, he said, often receive Nobel awards.

“True scientific advances will not come from answering well -known questions, but from asking new challenging questions and questioning shared concepts and previous ideas,” Wolf adds, calling the premise of the famous Douglas Adams’ 42 “, where we know the answer, but not the question.

“In my opinion, this is one of the reasons LLM, while they already have all the knowledge of mankind in memory, have not created any new knowledge by linking previously unrelated facts,” Wolf writes, calling knowledge a “intangible fabric of reality”. “They are mainly making” multiple filling “at the moment – filling the interference gaps between what people already know.”

On the other hand, he suggests, once they become capable of overcoming that thought in the box, the sky is the limit.

“Now we are building very obedient, non -revolutionary students,” Wolf writes. “This is perfect for today’s main purpose in the field of creating excellent assistants and highly compatible assistants. But until we find a way to stimulate them to question their knowledge and propose ideas that potentially go against past training data, they will not yet give us scientific revolutions.”

Competing in academics

After the break, Whittemore entered his personal experience, talking about the study of an academic decathlon, where students would study throughout the year.

Whittemore revealed that after competing in the top five rankings nationwide for two years, he tracked the top five students in the later competition in life to see what kind of trajectories their careers received.

“The only thing they all shared was a crazy readiness to work hard,” he said. “But most of them, as I would discover later, following their time through the college and in their careers, were very thinkers inside the box. They came from schools who had good programs who knew what to do to return the champions, and so they put the result … But none of them were the demolition, none of them were.”

In contrast, he spoke about the types of people you read about in the technology media, the main motifers who become household names.

“They had a concern, a curiosity, a group of qualities that led them to want more and be willing to play out of the system rules to get it,” he said.

In the use of LLM, he noted, we may want to challenge this.

“We assume this right line among today’s LLM, which are essentially like the best academic students of the Detelon we have ever imagined, after reading all things, we have studied all things, and who can now remember all things and tell you all things, but who are not creating anything for themselves,” he said, asking if we can think in voice. “Given how much we show scientific achievements and scientific progress, such as the overthrow of the AI, universally agreed, I really think these questions are worth thinking, and it is worth it to be really excavated.”

Multi -agent processes

With this he said, Whittemore also acknowledged that we are in an era of thinking about the multi-agent progress with him. He spoke about how the supercharges agents’ network skills cooperation, and why this will matter in the future.

I personally quote Minsky’s Minsky’s company here, partly because of its bonding myth, but also because it is a fundamental premise that continues to return – that the human brain, as powerful as it is, is not a computer but hundreds of connected computers, and to compete with people, artificial intelligence units will need the same interconnection and interconnection. You can call that ensemble lesson, network knowledge or group recognition. Or you can simply use the general term “cooperation”.

The conclusion is that the real agi can mean systems that can not only get better in answering the questions, but also become able to ask them.