Is Google PaLM A Response To OpenAI’s Recent Ascent?
And why is no-one speaking about PaLM and why can’t Google create the same hype as OpenAI?
I’m currently the Chief Evangelist @ HumanFirst. I explore and write about all things at the intersection of AI and language; ranging from LLMs, Chatbots, Voicebots, Development Frameworks, Data-Centric latent spaces and more.
Introduction
Amongst many other things, one aspect which OpenAI seems to be mastering is online hype. With every new release and feature, Twitter and LinkedIn timelines are filled with related comments, observations, tests etc.
Most recently, it was the release of the DaVinci 3 model and ChatGPT that dominated time-lines. OpenAI seems to successfully navigate the media pitfalls encountered by Meta AI with BlenderBot 3.
OpenAI is also mastering the art of managing a controlled release to enough users to generate social media traction and to the right user; without making the new product generally available.
👉🏼 Follow me on Linked-In For Conversational AI Updates! 🙂
The Threat To Google…
LLMs will change search as we currently know it, and the threat to Google is real…let me explain…
LLMs are Knowledge Intensive, with some connected to the internet and others being a contained time-stamped model, like ChatGPT.
LLMs can facilitate search in natural language with natural dialog-like responses and chain of thought reasoning. Hence context and dialog state are also maintained.
The three distinct advantages of LLMs are:
1️⃣ LLMs are most often curated, self-contained models trained on specific data. Whilst leveraging the web alone and in a direct fashion is universal, it is un-curated and often unstructured in terms of responses.
Is Microsoft introducing a new era of dialog natural language based search via OpenAI and ChatGPT?
2️⃣ LLM systems do not necessitate access to the Internet. For instance, ChatGPT does not have access to the internet for information retrieval.
Especially in the case of OpenAI, the LLM is a very broad domain, knowledge intensive interface for question-answering or fact-checking tasks.
The AI models underpinning the LLM framework searches through a digital archive for relevant information. The more comprehensive the digital archive, the broader and correct the answers.
3️⃣ LLMs allow for users to ask broad general domain questions, with the response in natural language. The response type is configureable by the user and can be in the form of:
- a summarisation
- simplified format
- a response in a different human langauge
- in point or bullet format
- and more…
With multi-turn highly contextual dialogs…as seen below. KI-NLP search results are highly customisable and consumable.
Whilst Google does not handle multi-turn questions, leverages the open-web and formatting of responses are not possible. Hence a LLM model can liberate users from the behest of traditional search engines on various fronts; not to mention inaccuracies and biases gleaned from the web.
Hence a new form of search is manifesting and Google are surely alarmed by this.
👉🏼 Follow me on Linked-In For Conversational AI Updates! 🙂
Why Google’s Response Is Faltering
Firstly, with the launch of LaMDA, claims that the model is sentient did not help their efforts to propagate and popularise their LLM. With LaMDA announced 18 May 2021, there has been little user hype and experience sharing.
Access to LaMDA is only available via a mobile app called the AI Test Kitchen. Hopeful makers can express their interest of experimenting with LaMDA, after which approval sits with Google. (I’m yet to receive approval and I’m currently on the waiting list.)
The focus and aim of LaMDA is to be a breakthrough conversation technology; facilitating chain of thought and complex reasoning.
PaLM was announced 4 April 2022 as a 540 billion parameter model, in the earlier graph you can see PaLM compared to other LLMs in terms of sheer parameters.
However, PaLM does not have matched media and user interest primarily due to the lack of access and real user examples.
Whilst the commercial and end-user intent of other LLM providers are very much visible…take OpenAI for instance, they have adjusted their pricing drastically a few months ago it a bid to facilitate accessibility and use.
Which begs the question, what is Google’s consumer strategy with PaLM?
👉🏼 Follow me on Linked-In For Conversational AI Updates! 🙂
More On PaLM
PaLM is multilingual and as seen below, checks all the LLM boxes in terms of dialogue, semantic search, QnA, translations, dialog management, human langauge to code, etc.
The image below shows the functionality PaLM can branch out into…according to Google.
Below are the results of 29 English based NLP tasks performed by PaLM. Google claims in addition to English NLP tasks, PaLM also performs well with multilingual NLP tasks, including translation. While only 22% of the training corpus were languages other than English.
In this final example below, PaLM is asked via a few-shot learning input to explain a joke:
👉🏼 Follow me on Linked-In For Conversational AI Updates! 🙂
In Conclusion
There has been reports regarding a Code Red meeting at Google and a need to define their AI strategy, and here in lies the Achilles’ heel of LLMs…the go-to market strategy. What are the use-cases and practical implementations of LLMs from a consumer or enterprise perspective.
LLM Generative Models have been implemented as a supporting technology in chatbot development frameworks. Generative Models are also capable of chain of thought prompting which elicits reasoning.
LLM Predictive Models (classification) raises the problem of a data gap, where specific data is required from fine-tuning a LLM for classification.
Recently I wrote about closing the data gap for predictive models (LLM) by utilising a NLU/NLG Design tool like HumanFirst Studio to create LLM training data from highly unstructured data.
I’m currently the Chief Evangelist @ HumanFirst. I explore and write about all things at the intersection of AI and language; ranging from LLMs, Chatbots, Voicebots, Development Frameworks, Data-Centric latent spaces and more.
