LaMDA was in the news after a Google engineer He claimed he was conscious Because his answers allegedly indicate that he understands what it is.
The engineer also suggested that LaMDA communicate that he has fears, like many humans do.
What is a lambda, and why do some have the impression that it can achieve consciousness?
LaMDA is a language model. In natural language processing, the language model analyzes the use of language.
Basically, it is a mathematical function (or statistical tool) that describes a possible outcome related to predicting the next words in a sequence.
It can also predict the occurrence of the next word, and even the next sequence of paragraphs.
GPT-3 from OpenAI The language generator is an example of a language model.
With GPT-3, you can enter the subject and instructions to write in a specific author’s style, and it will generate a short story or article, for example.
LaMDA differs from other language paradigms because it is trained on dialogue, not text.
Whereas GPT-3 focuses on creating language text, LaMDA focuses on creating dialogue.
Why is it a big deal
What makes LaMDA a remarkable feat is that it can generate conversation in a free-form manner that is not constrained by the parameters of task-based responses.
A conversational language model must understand things like multimedia user intent, reinforcement learning, and recommendations so that the conversation can navigate unrelated topics.
Built on transformer technology
Similar to other language models (such as MUM and GPT-3), LaMDA is built on top of Network of neurotransmitters Architecture for understanding language.
The Google He writes about transformers:
“This architecture produces a model that can be trained to read many words (a sentence or paragraph, for example), pay attention to how those words relate to each other and then predict which words it thinks will come next.”
Why does LaMDA seem to understand the conversation
BERT is a model trained to understand what the ambiguous phrases mean.
LaMDA is a model trained to understand dialogue context.
This quality of understanding the context allows the LaMDA to keep up with the flow of the conversation and give the feeling that it is listening and responding accurately to what is being said.
It is trained to understand whether the response makes sense relative to the context, or if the response is specific to that context.
Google explains it like this:
“…Unlike most other language paradigms, LaMDA is trained on dialogue. Throughout her training, she picks up many of the nuances that distinguish open conversation from other forms of language. One of those nuances is sensitivity. Basically: does the response to the context of a conversation given Boolean?
Pathological responses also tend to be specific, by relating clearly to the context of the conversation.”
LaMDA is algorithm based
Google published its announcement of LaMDA in May 2021.
The official paper was published later, in February 2022 (LaMDA: Language Paradigms for Dialogue Applications pdf).
The paper documents how LaMDA was trained to learn how to produce dialogue using three measures:
- the earth
The quality measure itself is reached through three metrics:
The research paper states:
“We collect annotated data describing how logical, specific, and interesting the response is for a multisession context. We then use these annotations to adjust the discriminator to rearrange candidates’ responses.”
Google researchers used cohorts of diverse backgrounds to help classify responses when they were unsafe.
That labeled data was used to train LaMDA:
“We then use these labels to fine-tune the discriminator to detect and remove unsafe responses.”
Grounding was a training process for teaching LaMDA to look for factual credibility, meaning that answers could be verified by “known sources”.
This is important because, according to the research paper, neurolanguage models produce statements that appear valid, but are actually incorrect and lack factual support from known sources of information.
Human crowd workers used tools such as a search engine (an information retrieval system) to check answers so that AI could also learn how to do so.
The researchers write:
“We found that increasing model output with the ability to use external tools, such as an information retrieval system, is a promising approach to achieving this goal.
Therefore, we collect data from an environment where collective workers can use external tools to research factual claims, and train the model to mimic their behavior.”
LaMDA is trained using human examples and evaluators
Section 3 of the paper describes how LaMDA was trained using a multibillion-dollar set of documents, dialogues, and utterances, consisting of 1.56 trillion words.
Section 4.2 documents how humans categorize LaMDA responses. Evaluations are the feedback that informs LaMDA when it is doing well and when it is not.
Human raters use an informational retrieval system (a search engine) to check answers, rating them as useful, correct, and factual.
LaMDA training uses a search engine
Section 6.2 describes how LaMDA receives a question and then generates an answer. After generating the response, it then performs a search query to check for accuracy and review the answer if it is incorrect.
The above research paper outlines the process of receiving a question, creating a draft of an answer, researching the answer, and then updating it with valid facts:
- User: What do you think of Rosalie Gascoigne’s sculptures?
- LAMDA-BASE: She’s amazing, and I love how she’s changed her work over her life. I like the later work more than the previous ones. Her influence is also interesting – did you know she was one of the artists who inspired Miro? “
The problem with the answer is that it is factually incorrect. So LaMDA performs a search query and selects facts from the best results.
Then he responds with a factually updated response:
“Oh my God, her life path is so inspiring. Did you know that she used to practice Japanese flower arranging before turning to sculpting?”
note the “Oh, wowPart of the answer is; this is a form of speech that learned how humans talk.
It appears that a human is speaking, but it is only a simulation of a speech pattern.
Language models mimic human responses
More advanced language models will continue to improve at simulating feeling.
Talented operators can drive chatbot technology to make a conversation that is a text that would be sent by a living individual.
This creates a confusing situation where something feels human and the model can ‘lie’ and say things that mimic the feeling.
can lie. He could reasonably say, I feel sad and happy. or feel pain.
But it’s imitation and imitation.”
LaMDA is designed to do one thing: provide conversational responses that are logical and specific to the dialogue context. That can give her the appearance of being sensitive, but as Jeff says, she’s basically a liar.
Therefore, although the responses a LaMDA gives sound like a conversation with a sentient being, the LaMDA is just doing what it is trained to do: giving responses to answers that are reasonable to the context of the dialogue and highly specific to that context.
Section 9.6 of the research paper, “Impersonation and anthropomorphism,” explicitly states that LaMDA impersonates a human.
This level of impersonation may lead some people to anthropomorphize LaMDA.
“Finally, it is important to realize that learning LaMDA is based on simulating human performance in conversation, similar to many other dialogue systems…a path toward high quality, engaging conversation with artificial systems that may eventually be indistinguishable in some respects from conversation.” With very likely now a human.
Human beings may interact with systems without knowing that they are artificial, or personify the system by attributing some form of personality to it.”
A question about feeling
Google aims to build an AI model that can understand text and languages, identify images, and create conversations, stories, or images.
Google is working toward this AI model, called Pathways AI Architecture, which it describes in “Word::
“Existing AI systems are often trained from scratch for each new problem…instead of extending existing models to learn new tasks, we train each new model from nothing to do one thing and one thing only…
The result is that we end up developing thousands of models for thousands of individual tasks.
Instead, we would like to train a single model that can not only handle many discrete tasks, but also build on and combine its existing skills to learn new tasks faster and more effectively.
In this way, what the model learns by training on one task — for example, learning how aerial photos can predict the elevation of a landscape — can help it learn another task — for example, predicting how floodwaters will flow through that terrain.” .
Pathways AI aims to learn concepts and tasks that it has not been trained in before, just like a human, regardless of the method (vision, sound, text, dialogue, etc.).
Language models, neural networks, and language model generators usually specialize in one thing, such as translating text, generating text, or identifying what is in images.
A system like BERT can identify the meaning in an ambiguous sentence.
Similarly, GPT-3 only does one thing, which is to generate text. He can create a story in the style of Stephen King or Ernest Hemingway, and he can create a story as a mixture of both authoring styles.
Some models can do two things, such as process both text and images simultaneously (LIMoE). There are also multimedia models such as MUM that can provide answers from different types of information across languages.
But none of them are quite on the Pathways level.
LaMDA impersonates human dialogue
The engineer who claimed to have a conscious lambda It was stated in a tweet that he could not support these claims, and that his statements about character and emotion were based on religious beliefs.
In other words: these claims are not supported by any evidence.
The evidence we have is clearly stated in the research paper, which explicitly states that the skill of impersonation is so high that people may objectify it.
The researchers also write that bad actors can use this system to impersonate a real human being and trick someone into thinking they are talking to a specific individual.
“…adversaries could attempt to discredit another person, take advantage of their prestige, or spread false information by using this technology to impersonate the conversational style of specific individuals.”
As the paper explains: LaMDA has been trained to impersonate human dialogue, and that’s pretty much it.
- Google LaMDA: How the Language Model for Dialog Applications Works
- Can AI perform SEO? OpenAI’s GPT-3 experiment
- Build your own SEO answer box with GPT-3 Codex & Streamlit
Image via Shutterstock / SvetaZi