Why we must be careful about how we speak of large language models

You can check out all the on-demand sessions here.

We have used the verbs "thinks," "knows" and "believes" for decades. Such descriptions are usually harmless.

We must be careful about how we talk about software, artificial intelligence, and large language models because they are fundamentally different from the human mind.

The same intuitions that we use in our dealings with each other should not be applied to artificial intelligence systems. We need to be aware of how they work and avoid imputing to them capacities they don't have.

Openai CEO admits risks. What's happening now?

The intelligent security summit is on demand.

Understand the role of artificial intelligence and machine learning in case studies. You can watch sessions on demand.

Humans vs. LLMs

It is amazing how human-like systems can be. It's easy to think of them as entities with minds like our own after interacting with them for a while. They are alien forms of intelligence and we don't fully understand them. We need to be careful when incorporating them.

Collective behavior is related to human language use. Our interactions with our community and world give us language.

As an infant, your parents and caregivers offered a commentary in natural language while pointing at things, putting things in your hands or taking them away, and so on. LMs are trained in a different way than we are used to.

A LLM is a mathematical model that shows the distribution of token in a text. They don't do the same things a human would do in response to a question.

The interaction with an LLM is simplified by Shanahan. Let me know how this fragment will end. What words are likely to come next?

Correct answers can be produced at an impressive rate when trained on a large amount of examples. The difference between humans and LLMs is crucial. Different excerpts of language can have different relationships with truth. It is possible to tell the difference between fact and fiction. These distinctions are invisible for an LLM.

It is a good idea for users to keep reminding themselves of what.

It's true that LLMs do. This reminder can help developers avoid the misleading use of words like belief, knowledge, understanding, and self.

The blurring barriers

When we talk about phones, calculator, cars, etc., there is usually no problem in using anaglyph language. These wordings can be used as shorthand for complex processes. Things can get a little blurry in the case of LLMs.

There is a lot of research on engineering tricks that can improve the performance of LLMs. Adding a simple sentence such as "Let's think step by step" can improve the LLM's ability to complete reasoning and planning tasks. The temptation to seeLLMs as having human- like characteristics can be amplified by such results.

There are differences between reasoning in humans and reasoning in LLMs. We know that our intent, our background knowledge, and our interests are understood by the person we are talking to. They know how to verify their answer by looking at a map or asking other people.

The rich context that was missing when you asked the same question is missing now. Context is provided in the background by adding bits to the prompt, such as framing it in a script-like framework that the artificial intelligence has been exposed to during training. It's more likely that the LLM will come up with the correct answer. The artificial intelligence doesn't "know" about the relation to each other.

It isn't the same as knowing that the country to the south of Rwanda is calledBurundi.

Careful use of LLMs in real-world applications

We should be careful how we build applications on top of LLMs. Users need to be careful about how they think about their interactions with them. The safety and robustness of their applications can be impacted by the way our mindset is framed.

The way we use psychological terms like "believes" and "thinks" may need to be changed in order to expand LLMs.

It may take an extensive period of interacting with, of living with, these new kinds of artifacts before we know how to talk about them. We should resist the call ofmorphism.

The mission of VentureBeat is to be a digital town square for technical decision-makers to gain knowledge. Our briefings can be found here.