This article was originally published on Forbes on Feb 17, 2023,01:18am ESTForbes on Feb 17, 2023,01:18am EST
Earlier this month, Google’s stock (Alphabet) tumbled 7% when chatbot Bard was unable to complete a search with 100% accuracy. During a demonstration, Bard returned incorrect information about which telescope was the first to take pictures of a planet outside the Earth’s solar system. This was a minor mistake given how far large language models and generative AI has come, rather it was the timing that was a bit flawed as OpenAI’s ChatGPT, the chatbot powering competitor Microsoft Bing, had been dominating headlines since its November 30th launch.
Microsoft, being an opportunist, took it a step further and announced Bing would now be powered by a faster and more accurate version of GPT-3.5 one day after Bard’s failed demonstration: “We’re excited to announce the new Bing is running on a new, next-generation OpenAI large language model that is more powerful than ChatGPT and customized specifically for search. It takes key learnings and advancements from ChatGPT and GPT-3.5 – and it is even faster, more accurate and more capable.”
Both companies have been preparing for this moment for many years. Microsoft invested $1 billion into OpenAI a few years ago with a new $10 billion round announced last month. Meanwhile, Google acquired DeepMind in 2014. Google also previously developed conversational neural language models such as LaMDA, which is used by Google’s Bard for its conversational AI technology.
As much fun as the media has had lately poking fun of Bard, there have been similar, compelling reports of ChatGPT-powered Bing also having accuracy issues.
Point being, both are in the early stages and mistakes are being blown out of proportion. Which brings up more important questions for investors – given that technology can require many iterations, what’s the right timing for generative AI and chatbots to drive real advertising revenue?
Investors can get burned by being too early. For example, autonomous vehicles (AVs) were promised in 2019, and the Metaverse has not driven any real gains despite a large media push in early 2021. How does AI compare in terms of time to market?
Secondly, Alphabet has a lot of turf to defend. It won’t only be Bing, but also browsers like Opera that will incorporate ChatGPT into its sidebar. From there, it’s easy to imagine other competitors may crop up over time, some replacing search engines entirely with conversational AI applications powered by speech recognition, which are otherwise unimaginable today.
We look at these key points below for a 360-degree view on Google’s stock given search is on the precipice of its first major shift in over two decades.
Background on AI-Powered Search
“AI is the most profound technology we are working on today. Our talented researchers, infrastructure and technology make us extremely well positioned, as AI reaches an inflection point.” -Sundar Pichai, Alphabet’s Q4 earnings call.Q4 earnings call.
Despite the mishap with Bard, it would be a human-generated mistake to think Alphabet does not command a place of leadership right now in generative AI. Alphabet was one of the first tech companies to focus and invest on AI and natural language processing (NLP). We pointed out to our premium research members in July of 2022 that ChatGPT is based on transformer architecture that Google initially introduced in 2017 when we saidpremium research members in July of 2022 that ChatGPT is based on transformer architecture that Google initially introduced in 2017 when we said:
“Transformers are becoming one of the most popular neural-network models by applying self-attention to detect how data elements in a series influence and depend on one another.
Sequential text, images and video data are used for self-supervised learning and pattern recognition, which results in more data being used to create better models. Prior to transformer models, labeled datasets had to be used to train neural networks.
Transformer models eliminate this need by finding patterns between elements mathematically, which substantially opens up what datasets can be used and how quickly.
Google first introduced transformer models in 2017 and transformers are used in Google and Bing Search. Transformers also led to BERT models, which stands for Bidirectional Encoder Representations from Transformers, and is commonly used for text sequences. Transformers are also used in GPT-3 (it’s the T in GPT) which improved from 1.5 billion parameters to 175 billion parameters. GPT-3 has the ability to report on queries it has not been specifically trained on.”
Earlier this month, Google’s CEO, Sundar Pichai, gently reminded the AI community of how cutting edge Google’s research is when he stated, “Transformer research project and our field-defining paper in 2017, as well as our important advances in diffusion models, are now the basis of many of the generative AI applications you're starting to see today.”
BERT was designed to help Google better understand search intent, as despite billions of searches every day, about 15% of those searches are for brand new terms. This prompted Google engineers to develop a model that could self-learn.
The result is that searches results are more accurate by taking into consideration the nuances of language.
Sign up for I/O Fund's free newsletter with gains of up to 221% -Click hereSign up for I/O Fund's free newsletter with gains of up to 221% -Click hereClick here
Multitask Unified Models (MUMs) are 1,000X More Powerful than BERT
Multitask Unified Models (MUM) were introduced in 2021 to further address conversational nuances and is 1,000 times more powerful than BERT. MUMs will have a large impact for search users as it decreases the amount of effort put into seeking the desired information. It’s not only addressing the 15% of search based on new terms, rather it’s a powerful iteration that returns search results that more closely resemble how humans interact.
According to Google, it takes an average of eight queries to answer a complex question. With MUM, this is reduced to one query. You can theoretically ask “should I travel to Hawaii or California this Fall?” and MUM will be able to compare travel rates and weather patterns to answer this question with more depth. Similar to if you ask your friend this question, they might answer “Hawaii is more expensive to travel to but California is prone to wildfires in the Fall, so I would go to Hawaii.” To search for this answer would take many queries, but with MUM, succinct, human-like responses are provided in only one interaction.
Large language models have been helping to improve search results for many years. Therein, Google presents its moat; which is not only a deeply engrained behaviour pattern where search users automatically turn to the multi-decade leader out of pure habit, but that Google search truly presents the highest quality search results today.
Google’s commanding lead on search is not a legacy metric, by any means, rather it symbolizes the lead Google has on data for training large language models.
Bard’s demonstration may have been problematic compared to Chat-GPTs more favorable reviews, however, it’s nothing more than that for now —- which is a mix of bad reviews and good reviews by a limited number of beta users.
TPUs:
This brings us to Google’s TPUs, which are essentially ASICs (application specific integrated circuit) on the efficiency/flexibility spectrum. I first covered the differences between TPUs and GPUs nearly four years ago in 2019 for our premium members when I said:first covered the differences between TPUs and GPUs nearly four years ago in 2019 for our premium members when I said:
“TensorFlow is rising in popularity as a machine learning framework and TPUs primarily run TensorFlow models. This is one of Google’s more successful experiments. They are cheaper and use less power than GPUs and are specifically focused on machine learning.
TPUs train and run machine learning models and power Google Translate, Photos, Search, Assistant and Gmail – i.e., image recognition, language translation, speech recognition and image generation.”
Although there are ongoing debates between TPUs and GPUs, the primary difference is that TPUs are application specific and have been optimized for Google’s AI tools. Meanwhile, Nvidia was the first to break ground in deep learning due to the ease of programming GPUs and the relative speed in which parallel computing can train networks. Nvidia also offers its customers an aggressive product road map.
An example of this is the H100 DGX SuperPods, which we covered for our premium members in July when we said:
“Nvidia and Microsoft recently worked on a Mega transformer model with 530 billion parameters and the future for AI engineers is trillion-parameter transformers and applications. The H100 is already prepping for this. According to Nvidia, the training needs for transformer models will increase 275-fold every two years compared to 8-fold for other models. The H100 GPU with its Transformer Engine supports the FP8 format to speed up training to support trillion-parameter models. This leads to transformer models that go from taking 5 days to train to becoming 6X faster to only taking 19 hours to train.”
As of today, TPUs do not necessarily provide Google an advantage over Microsoft’s partnership with Nvidia. When TPUs were first launched, it was expected that it would provide Google an important lead in launches such as Bard. However, Nvidia has proven to be a more difficult competitor than originally expected, and I imagine Microsoft will not stray from this partnership as the company will instead focus on other areas, such as taking more market share with Bing.
Introduction of Bard powered by LaMDA
LaMDA is a conversational language model that powers Bard. Two years ago, Google launched LaMDA to better mimic open-ended conversations by training the language model on dialogue. The result was a more human-like chatbot that personally knows you well enough to recommend movies or books, is sensitive enough to change an uncomfortable conversation, can discuss its own “death” by being turned off, —- and also has machine vision to where it can look at a picture and discuss the picture intelligibly.
Bard was released this month for beta testers and will be available to the public “in the coming weeks.” As mentioned in the introduction, Bard answers questions with real-time data whereas ChatGPT is trained on data from 2021 or earlier (note: the new Bing version is rumoured to use real-time data, see below).
Bard is also free, and given Google’s search revenue, the company may have incentives to undercut competitors that charge paid plans for conversational AI.
Other ChatGPT alternatives
Anthropic is building a 52-billion-parameter pretrained model called Claude, which is a potential rival to ChatGPT. Google invested $300 million into Anthropic last year, with a similar arrangement as Microsoft and OpenAI, which includes a stake in the R&D of the startup. Anthropic was founded by former employees of OpenAI. Whether Claude can actually exceed Google’s own language systems is yet to be determined, or perhaps Google is simply spreading its bets and wanting access to its competitors’ former talent. Despite being in closed beta, there’s an excellent write-up here about the differences between ChatGPT and Claude.
