|
Who will win the AI race? What it takes to become a prompt engineer
It’s silly to bet on any one company to win the AI race because there are too many factors that could rapidly change the equations in this field. The factors include partnerships, acquisitions, the emergence of new technologies, and global regulations. That said, here is what the major tech companies are currently doing to stay relevant in the AI race by enhancing their products and platforms while partnering and investing in other AI startups.
Google / DeepMind / Anthropic: Graduating from Cacofonix to Shakespeare
When Microsoft partnered with OpenAI, many began to write off Google, whose mission was “AI first”. Google was ideally positioned to be the winner in the AI race because its transformer model, which can predict the next word, sentence or even para, is the foundation for all large language models, or LLMs. But that was not to be since OpenAI’s generative pre-trained transformer (GPT) and the GPT-powered chatbot ChatGPT garnered more than 100 million users in just the first two months of the launch (31 December 2022) with an interface that enamoured users.
That Google’s Bard was making blunder after blunder only added to Google’s woes, resembling the bumbling village bard Cacofonix in the comic strip ‘Asterix’, and helping ChatGPT’s cause.
But just when we thought that Google would fall behind in the AI race, the company said it would combine its AI research units -- Google Brain and DeepMind. Google also invested over $300 million in Anthropic, an AI company founded by former employees of OpenAI, but the exact figure has not been publicly disclosed. The investment gave Google a 10% stake in Anthropic and allowed the company to scale its AI computing systems using Google Cloud.
Anthropic was founded in 2021 by Dario Amodei and others who were previously involved in the development of OpenAI’s GPT-3 language model. The partnership between Google Cloud and the unicorn Anthropic allows the latter to use Google’s infrastructure to train and deploy its AI models.
Google has also rejuvenated Bard and made it available in 180 countries, including India. Bard uses Language Model for Dialogue Applications (LaMDA), a transformer-based model invented by Google in 2017. It learns by “reading” trillions of words that help it to pick up on the patterns that make up human language. While Bard, Bing Chat, ChatGPT, and even Hugging Chat are learning and evolving, Bing and Bard deal with current data (ChatGPT is not connected to the internet, and the data cut-off remains September 2021) and can cite sources which are personally very comforting. ChatGPT also has numerous plugins and application programming interfaces (APIs) to make it work with current data, and it’s just a matter of time till these differentiators vanish.
Meanwhile, Google announced that Bard would use the upgraded version of the Pathways Language Model called PaLM 2, which is trained on text from more than 100 languages. It is being touted as Google’s answer to OpenAI’s GPT-4. According to Google, this has improved Bard’s ability to understand, generate and translate nuanced text, including idioms, poems and riddles, across various languages.
Besides passing advanced language proficiency exams at the “mastery” level, Google says PaLM 2’s dataset includes scientific papers and web pages containing mathematical expressions. For instance, Workspace features help users write in Gmail and Google Docs and organize in Google Sheets, while Med-PaLM 2 can answer questions and summarize insights from medical texts. Google says it’s adding multimodal capabilities to synthesize information like X-rays and mammograms to improve patient outcomes. (Google plans to soon open Med-PaLM 2 to a small group of Cloud customers for feedback). Sec-PaLM is a specialized version of PaLM 2 trained on security use cases.
PaLM 2 will be made available in four sizes -- Gecko, Otter, Bison and Unicorn. Gecko is lightweight and can work on mobile devices, and is fast enough for great interactive applications on-device, even when offline, according to Google. At Google i/O this month, the company announced new updates to Bard, including image capabilities, coding features and app integration. As an example, Bard will use Google Lens to analyze photos, detect dog breeds, and even draft a few creative captions.
More importantly, Google hopes to make Bard citations more precise. “If Bard brings in a block of code or cites other content, just click the annotation, and Bard will underline those parts of the response and link to the source.” Bard will also be able to tap into all kinds of online services with extensions from outside partners like Adobe Firefly.
When you interact with Bard, Google collects your conversations, your general location based on your IP (internet protocol) address, feedback, and usage information. Bard also uses your past interactions with it and your general location to generate its response, but Google insists that Bard conversations are currently not being used for advertising purposes.
What works for Google is that it is the most frequently-used search engine worldwide other than in Russia (where a little over 60% use Yandex) and China (where Baidu is the most used search engine with over 85% market share). In other countries like Japan and Mexico, too, people use Yahoo along with Google (In the first quarter of 2022, nearly 56% of the respondents in Japan said that they had used Yahoo in the past four weeks. In the same year, over 27% of users in Mexico said they used Yahoo). But overall, Bing accounted for a mere 8.23% of the global desktop search market, while market leader Google had a share of around 85.53% as of March 2023, according to market research firm Statista. And in January 2023, Google accounted for 96.46% of the global mobile search engine market worldwide, while competitors like Yandex, Baidu and Yahoo! accounted for less than 1% each on a global scale.
Source: Statista
Further, Google not only leads search engines but the browser and digital ad markets too. Microsoft recorded $18 billion in revenue from digital advertising in CY22 as opposed to Google’s $168.44 billion net revenue in the same period.
So, it appears to be advantage-Google, at least in the consumer space for now -- unless Bard runs afoul of regulators or is perceived to hallucinate (the tendency of AI chatbots to make up information) more than Bing and ChatGPT since most comparisons of Bard Vs ChatGPT Vs Bing tend to be subjective.
Microsoft / OpenAI: Playing on enterprise strength and gaining a foothold in the consumer space
Futurist and author of Rise of the Robots, Martin Ford, tweeted on 8 February: “A Tech Race Begins as Microsoft Adds #AI to Its Search Engine”. There’s a lot of truth to this because the enterprise space is dominated by Microsoft, and its partnership with, and stake in, OpenAI made it a force to reckon with in the AI space while also strengthening its presence in the consumer AI arena. Things will only improve when OpenAI launches ChatGPT for Business.
Microsoft is already touting Bing Chat with Edge as ‘Your copilot for the web’, and it already owns Github (also a repository for open-source AI tools) and the Azure platform that provides these tools to businesses. In the first quarter of this calendar year, Microsoft reported a profit of $18.3 billion on revenue of $52.9 billion as Cloud and AI more than offset its fall in revenue from licensing Windows software to computer makers. This augurs well for the tech company, and Google can ill afford to let its guard down again. Microsoft’s investment of 10 billion in OpenAI has been worth over $200 billion to the company, having seen its shares rise by 16% as of 24 April 2023, according to Data Hub.
Meta: Balancing the metaverse with AI
While Google bet on AI, and Microsoft caught up with its investment in OpenAI, Mark Zuckerberg chose a different path. In October 2021, Zuckerberg made his Meta announcement, insisting that the new identity of his company should be “metaverse-first” and “not Facebook-first”, and many CEOs appeared keen on exploring the metaverse to get the first-mover advantage. Unfortunately, history was not kind to Zuckerberg, and by October 2022, Meta had already sunk in $10-15 billion in the metaverse. It became clear that Meta was burning cash in the metaverse, and investors and shareholders were not amused.
Meta, it appears, is now striking a balance between the metaverse (which it hasn’t abandoned as many think) and AI. It has created an AI Sandbox to “act as our testing playground for early versions of new tools and features, including generative AI-powered ad tools”. Meta Advantage is its portfolio of automation products that use AI and machine learning to help optimize campaign results, personalize ads. It also announced Meta Lattice, a new AI-driven model to improve the performance of its ads on its networks.
Further, while Meta had to shut down two of its AI chatbots called BlenderBot and Galactica, due to glitches, it has got credibility back by releasing a Large Language Model Meta AI (LLaMA) and publicly sharing the code for researchers to test it. Meta’s LLaMA claims it requires “far less computing power and resources to test new approaches, validate others’ work, and explore new use cases”. You may read more about this here (‘Five trends that may change the course of Generative AI models’).
International Business Machines (IBM) / Hugging Face: Regaining ground in AI
While IBM today may be known for its prowess in quantum computing, it has a very credible history in AI too. On 18 June 2020, for instance, the world sat up and noticed how an AI system had engaged in the first-ever live, public debates with humans. At an event held at IBM’s Watson West site in San Francisco, a champion debater and IBM’s AI system, Project Debater, began by preparing arguments for and against the statement: “We should subsidize space exploration”.
IBM later held a second debate between the system and another Israeli expert debater, Dan Zafrir, that featured opposing arguments on the statement: “We should increase the use of telemedicine.” In development since 2012, Project Debater was touted as IBM’s next big milestone for AI.
Interestingly, a year later, at Think 2019 in San Francisco, IBM’s Project Debater lost an argument in a live, public debate with a human champion, Harish Natarajan. But even though Natarajan was declared the winner, 58% of the audience said Project Debater “better enriched their knowledge about the topic at hand, compared to Harish’s 20%. IBM’s Deep Blue supercomputing system beat chess grandmaster Garry Kasparov in 1996-97, and its Watson supercomputing system even beat Jeopardy players in 2011.
If you’re still wondering what happened to Watson, IBM has reincarnated it as WatsonX. It is IBM’s answer to AI-powered enterprise solutions from Google, Microsoft, and Amazon Web Services (AWS). WatsonX comprises a studio, data store, and governance toolkit to enable AI workflows that “are built with responsibility, transparency and explainability”. In a press statement, IBM explained that its AI development studio would offer access to IBM-curated and trained foundation models and open-source models while its data store would enable businesses to gather and clean training and tuning data.
More importantly, just as Microsoft partnered (and invested in) with OpenAI and Google combined its Google Brain and DeepMind units, IBM announced a collaboration with Hugging Face, following which the watsonx.ai studio will build upon Hugging Face’s open-source libraries and offer thousands of Hugging Face open models and datasets.
And when quantum computers eventually overcome the hurdles of error correction and fault tolerance, among other challenges, they may accelerate the progress of AI in ways we can’t imagine today. Companies like IBM--along with Google, Microsoft and Intel--will then get a new lease on life.
Amazon: Providing a platform for AI models
Amazon.com Inc. has joined Microsoft and Google in the generative AI race with its AI tool called Bedrock. Amazon SageMaker already allows developers to build, train, and deploy AI models and allows customers to add AI capabilities like image recognition, forecasting, and intelligent search to applications with a simple application programming interface (API) call. Amazon Bedrock allows LLMs or foundational models (as they are also referred to) from AI21 Labs, Anthropic, Stability AI, and Amazon to be accessible via an API. Further, if the code completion tool GitHub’s Copilot offers complete code snippets based on context, Amazon has announced the preview of Amazon CodeWhisperer--its AI coding companion.
Elon Musk
Elon Musk, who co-founded OpenAI along with Sam Altman in 2015, left its board in 2018 but is reportedly setting up a team to develop an AI-powered chatbot called ‘BasedAI’. Musk is wary of Microsoft’s investment in OpenAI and has accused it of “pushing for profits” instead of developing AI for the human good. Elon Musk has founded a new artificial intelligence company called X.AI Corp, according to reports from The Wall Street Journal and The Financial Times. The company is incorporated in Nevada and lists Musk as the sole director, with Jared Birchall, the director of Musk’s family office, as the secretary. The state filing was made last month. Musk is yet to unveil the details. It’s worth watching this space.
Nvidia / OpenAI: Fuelling and benefitting from the AI gold rush
AI has made rapid progress over the last decade five years, primarily due to three factors: better algorithms, more good quality data, and the phenomenal rise in computing power. Nvidia has benefitted from the third factor, powering AI models with its graphics processing units (GPUs) that are used in gaming. OpenAI, for instance, used H100’s predecessor — NVIDIA A100 GPUs — to train and run ChatGPT and will be using the GPUs on its Azure (Microsoft’s) supercomputer to power its continuing AI research. Meta is a key technology partner of NVIDIA and developed its Hopper-based AI supercomputer Grand Teton system with GPUs. Stability AI, a text-to-image generative AI startup, uses H100 to accelerate its video, 3D and multimodal models.
According to ‘The Motley Fool’, Nvidia’s stock has gained more than 150% since mid-October, “fueled by excitement about the prospects of AI”. However, the US-based financial and investing advice company does caution that Nvidia’s competitors will not sit quietly. “The other risk, albeit remote, is that one of Nvidia’s competitors develops a superior semiconductor for AI applications. The most recent on-again, off-again rumour is that Microsoft has joined forces with rival Advanced Micro Devices, better known as AMD, to create an Nvidia-killer AI chip, though that remains unconfirmed”. Besides, Google uses tensor processing units (TPUs), and Intel uses central processing units (CPUs). For now, though, Nvidia appears to be the winner. But the order can change rapidly in this field.
Intel / HuggingFace: Betting on powerful CPUs over super-expensive GPUs
With GPUs doing most of the heavy lifting for deep neural net training, Intel has been sharpening its focus on GPUs to handle the AI and Generative AI waves. But it simultaneously insists that since many AI solutions involve a mix of classic machine learning algorithms and small-to-medium complexity deep learning models, there is no need for a GPU when CPU designs such as Xeon will suffice. “We’re making Xeon run even faster with built-in acceleration and optimized software,” says Intel. Its strategy makes sense as we see AI’s proliferation from the cloud to the client to edge (our mobile devices, such as smartphones and those we use in cars, etc.). Intel is launching a software toolkit called AI Software Suite that provides both open-source and commercial tools to help build, deploy, and optimize AI workloads. Intel is already a member of Hugging Face’s Hardware Partner Program, and it is partnering with the latter to build hardware acceleration to train, fine-tune and predict with Transformers.
Intel has a point. GPUs are very expensive. Estimates suggest that OpenAI will need more than 30,000 of Nvidia’s A100 GPUs to commercialize ChatGPT, and Nvidia’s latest H100 GPUs are selling for more than $40,000 on eBay. It’s hardly any surprise, then, that Microsoft is developing a chip that it has internally code-named Athena and made it available to a small group of Microsoft and OpenAI employees, according to The Information.
Open-source LLMs: Joker in the pack
In my previous newsletter titled ‘Five trends that may change the course of Generative AI models’, I covered this as the first trend. The reason: Open-source LLMs have not only come of age but are providing developers with a lighter and much more flexible option. Developers, for instance, are flocking to LLaMA--Meta’s open-source LLM. Meta says it has trained LLaMA 65B and LLaMA 33B on 1.4 trillion tokens. Its smallest model, LLaMA 7B, is trained on one trillion tokens. According to Meta, smaller models trained on more tokens —pieces of words — are easier to re-train and fine-tune for specific potential product use cases. Pythia (from EluetherAI, which itself is likened to an open-source version of OpenAI) comprises 16 LLMs that have been trained on public data and range in size from 70M to 12B parameters. And in a leaked document accessed by Semianalysis, a Google employee claimed, “Open-source models are faster, more customizable, more private, and pound-for-pound more capable. They are doing things with $100 and 13B params (parameters) that we struggle with at $10M (million) and 540B (billion). And they are doing so in weeks, not months.” You can read more about this here. Open-source LLMs are clearly the joker in the pack.
Intel has a point. GPUs are very expensive. Estimates suggest that OpenAI will need more than 30,000 of Nvidia’s A100 GPUs to commercialize ChatGPT, and Nvidia’s latest H100 GPUs are selling for more than $40,000 on eBay. It’s hardly any surprise, then, that Microsoft is developing a chip that it has internally code-named Athena and made it available to a small group of Microsoft and OpenAI employees, according to The Information.
|
|
POINT TO PONDER: If Generative AI wins the race, will all humans be the losers?
‘May you live in interesting times’. Whether this quote is Chinese or not, it does present us with a dilemma when dealing with Generative AI. The latter, which comprises AI-powered tools like ChatGPT, Google’s Bard, Bing Chat, Hugging Chat, DALL·E 2, MidJourney, etc., is making waves because they can not only generate or write blogs, articles, captions, resumes, cover letters, product descriptions, reviews, marketing messages, short films, images, videos, software code, and provide templates for marketing campaigns, but also synthesize business and analyst reports, and aid in speeding up new drug discovery, in creating entirely new materials and generating synthetic data too.
But does this mean Generative AI is becoming sentient by achieving human-like intelligence or even surpassing it? I have talked about this earlier too, when discussing “Will machines think like humans with GPT-4?”. This is just the starting point, and we all will have to revisit our opinions as Generative AI becomes stronger and starts writing its own tools.
I had written about this in the “Rise of autonomous AI agents: AutoGPT, BabyAGI, AgentGPT, ChaosGPT, GodMode, and even TruthGPT”. AutoGPT, for instance, is an experimental open-source attempt to make GPT-4 fully autonomous. Developed by Toran Bruce Richards, it surfs the internet, generates sub-tasks, and initiates fresh agents to perform the tasks. With the help of GPT-4, Auto-GPT can generate code autonomously and subsequently “debug, develop, and self-improve” the same through recursive mechanisms. The project is hosted on GitHub.
While big tech companies would want us to believe that we are close to AGI (You may read this 154-page document titled ‘Sparks of Artificial General Intelligence: Early experiments with GPT-4’'this paper here:), I believe that even as such developments cause a stir because these AI models excel humans in specific tasks, the current developments do not necessarily indicate that a true AGI system is anywhere close to the ones that are shown in movies such as Skynet, Terminator, Transformers, iRobot, HER, etc.
I have explained this in detail so you may read more about this here:
Interesting debate:
Yuval Noah Harari (Author of Sapiens) Vs Yann Le Cun (Chief AI Scientist at Meta) on AI.
TIP OF THE WEEK
How to become a prompt engineer
Generative AI may take away your job, but it can also help you get a highly-paid job like prompt engineering if you’re willing to reskill yourself.
A prompt is crucial for interacting with large language models (LLMs). Well-crafted prompts are precise and targeted, extracting specific details accurately. Prompts can be in the form of questions, statements, or commands, depending on the desired information and response type.
Prompt engineering, for instance, can be used to analyze customer sentiment and behaviours, enabling businesses to make informed decisions and drive innovation. It can also help researchers to extract valuable information from large language models by breaking down tasks, specifying information requirements, and using specific vocabulary. Prompt engineering is also applicable in fields like healthcare, education, and finance, where it can help in extracting relevant information, analyzing trends, and making informed decisions.
|
|
A prompt is a set of input text or instructions used to guide AI models like ChatGPT, Bing Chat, Bard, Hugging Chat, Midjournery, DALLE-2, etc., to generate a desired output. It is a specific text that prompts an AI model to deliver the outcome you desire. Prompt engineering is the process of creating and refining these prompts to generate the desired result.
|
|
|
|
Unlike a prompt writer (like most of us who enter prompts in the context box of AI-powered chatbots like ChatGPT, Bing Chat, Bard, Hugging Chat, etc.), a prompt engineer would need a combination of technical skills, industry knowledge, and experience.
|
|
|
|
|
Prompt engineers program in prose and send the plain text commands to the AI model, which does the actual work.
|
|
|
|
A more technical area of prompt engineering is fine-tuning the input data used to train AI models. It involves carefully selecting and structuring the input data to maximize its usefulness for training the model.
|
|
|
|
A prompt engineer would need some technical experience in Big Data technologies like Hadoop, Apache Spark, and others.
|
|
|
|
Understanding Big Data technology will also be helpful when working on an AI model since you will need to work with a large amount of data.
|
|
|
|
Experts suggest that knowledge of programming languages like Java, C++, and Python will help you understand the working of AI models and modify the text to get the output you want.
|
While you should take an online course from the likes of Coursera, Udacity, Udemy, etc., to understand more on this subject, these are just some quick tips to help you speed-track your decision. The good news is that you do not require a computer science degree or advanced coding skills to take these courses.
Picture courtesy of Livemint
To kickstart a career as a prompt engineer, Illinois Tech suggests the following steps
|
|
Obtain a degree in a relevant field such as computer science, linguistics, or cognitive psychology.
|
|
|
|
Gain practical experience through internships or entry-level positions in AI or related fields
|
|
|
|
|
Develop a strong understanding of AI systems and natural language processing techniques; consider adding capacity in programming languages (e.g., Python) and AI frameworks (e.g., TensorFlow, PyTorch)
|
|
|
|
Build a portfolio showcasing your work with AI systems, particularly those involving natural language processing and prompt engineering
|
|
|
|
Network with professionals in the AI community to stay informed about industry trends and job opportunities
|
Anthropic, which recently advertised a position for a ‘Prompt engineer and Librarian’ asked for the following traits, which provides us with a good clue of what prompt engineering will demand.
|
|
Have at least a high-level familiarity with the architecture and operation of large language models.
|
|
|
|
Are an excellent communicator and love teaching technical concepts and creating high-quality documentation that helps out others.
|
|
|
|
|
Are excited to talk to motivated customers and help solve their problems.
|
|
|
|
Have a creative hacker spirit and love solving puzzles. Have at least basic programming skills and would be comfortable writing small Python programs.
|
|
|
|
Have an organizational mindset and enjoy building teams from the ground up.
|
|
|
|
You think holistically and can proactively identify the needs of an organization.
|
|
|
|
Make ambiguous problems clear and identify core principles that can translate across scenarios.
|
|
|
|
Have a passion for making powerful technology safe and societally beneficial.
|
|
|
|
You anticipate unforeseen risks, model out scenarios, and provide actionable guidance to internal stakeholders.
|
|
|
|
Think creatively about the risks and benefits of new technologies and think beyond past checklists and playbooks.
|
|
|
|
You stay up-to-date and informed by taking an active interest in emerging research and industry trends.
|
The expected salary range for this position is USD $280k - $375k. Now that you know the skillset you need, ask yourself: Are you up to the task?
|
