Keeping Up with the AI News: A 2024 Quiz

(c) Orion 5 is not a model launched in 2024.

Google launched Gemini 2.0 in December. Meta launched the open source Llama 3, 3.1, 3.2, and 3.3 in 2024. OpenAI released GPT-4o ("o" for "omni") in May. Orion is the rumored internal codename for GPT-5, successor to GPT-4, but it is yet to be launched.

(b) Notion, which has an AI offering (who doesn't?), but it's descriptively named Notion AI.

"Copilot" was everywhere in 2023 and 2024. GitHub released its coding assistant Copilot in 2022. Back then, it was $10 USD / month. In December 2024, GitHub introduced a free tier for Copilot. Microsoft, the parent company of GitHub, also has a Copilot. But instead of a coding assistant, it's an "everything assistant", integrated across multiple Microsoft productivity tools like Word, Excel, PowerPoint, and Outlook.

Salesforce, one of the largest enterprise software firms in the world (and parent company of Slack, the team communication platform), introduced its Einstein Copilot feature in February 2024. Einstein Copilot is deeply embedded into your Salesforce experience, helping you manage your customer data, and engage with your customers.

The "Copilot" naming trend is reflective of the "assistant" functionalities of AI in the last few years. But this is evolving. We are now shifting towards "Agentic AI" - a more autonomous AI helper that can complete tasks involving many steps, including interacting with external resources.

(c) $100 Million to $1 Billion USD

OpenAI's GPT 4 reportedly cost $100 Million to train. Google's Gemini came with a $191 Million price tag. In a podcast interview in April 2024, Anthropic CEO also confirmed the $100 Million estimate. However, Dr. Amodei further stated that model training cost is shooting up:

"Today’s (April 2024) models cost order of $100 million to train, plus or minus factor two or three.

The models that are in training now and that will come out at various times later this year or early next year are closer in cost to $1 billion. So that’s already happening. And then I think in 2025 and 2026, we’ll get more towards $5 or $10 billion." (Source)

To put it in perspective, the 2024 film Wicked had a budget of $150 Million. The Hoover Dam cost $790 Million. And GitHub was acquired by Microsoft for $7.5 Billion.

But if you think the barrier to entry for creating a frontier model is becoming too high, there is also an opposing point of view. Professor. Ng shared an analysis from ARK, who predicted that AI training cost will go downward:

"A foundation model that costs $100M to train this year (2024) might cost $25M to train next year." (Source)

(d) $157 Billion

In October, OpenAI raised $6.6 Billion at a post-money valuation of $157 Billion. How much is $157 Billion? Ukraine's 2024 national GDP (nominal) was $184 Billion. In fact, about 2/3 of the countries in the world have a smaller GDP (again, nominal) than $157 Billion.

Incredibly, the $6.6 Billion round was not the largest of the year. That first place would go to Databricks, who raised $10 Billion in a Series J in December.

(c) $56 Billion

Generative AI investment continued fervently in 2024. The $56 Billion total is a near double of the $29.1 Billion sum in 2023.

(d) $2 Trillion

We have gone from the monetary scale of Millions, to Billions, and finally now to Trillions. It was only several years ago, in 2018, when Apple became the first trillion dollar company. It was an unfathomable milestone that took Apple a little over 40 years to reach.

But what was unfathomable in 2018 is now surpassed by Nvidia's 2024 achievement. The dominant AI chip maker added $2 Trillion in market value in one single year. It's historical.

(c) Claude Ballad

2024 was the year of "Claude Computing". Throughout 2023, OpenAI's GPT models were the only viable choice for production-ready, high quality, general purpose AI. But that changed in June 2024 when Anthropic's new Claude 3.5 Sonnet model was released. It was benchmarked to outperform GPT-4o, and replaced OpenAI's models to become the default choice for many AI-powered services like Perplexity and Cursor.

The faster variant, Claude 3.5 Haiku (with less parameters), also became available in October 2024. The Opus model is the biggest in the Claude family, and therefore typically most performant. However, Claude 3.5 Haiku actually beat Claude 3 Opus (from a previous generation) in many benchmarks. Small(er), fast, and intelligent? Claude 3.5 Haiku you really are the whole package.

(a) Devin is a fully automated AI software engineer

The Devin announcement and product demo in March 2024 was one of the most breathtaking moments of 2024.

(d) $500 / month

With great power comes with great... price tag? However, Devin's pricing model is a bit different than other AI subscriptions, which is typically per user. The $500 / month for Devin comes with unlimited seats, but it does have a cap on "compute units".

(d) >25%

In the company’s Q3 earnings call, Google CEO Sundar Pichai said more than 25% of the code at Google is now generated by AI.

Combining with Devin's ability to act as a capable engineer, and many other advancements in coding AI, human coders, myself included, might feel very uneasy.

But perhaps we can be optimistic. When calculator came onto the scene, it took away tasks we once had to do manually (arithmetic), tasks considered a core skill at the time. However, calculator enabled us to do a lot more arithmetics with higher quality. This allowed us to abstract up and perform more complex tasks that involved a lot more arithmetics.

So while >25% of the code sounds like a lot, there's also plenty of repetitive work in software engineering that developers would be happy to offload to AI. Instead of writing type definitions, generating API documentation, creating boilerplate code, and setting up basic unit tests, developers can focus on system and architecture design.

(c) The Nobel Prize in Physics

Professor Hinton received the Turing Award, the highest distinction in the field of computer science and often referred to as the "Nobel Prize of Computing", in 2018. In October 2024, he was awarded the Nobel Prize in Physics.

(b) Predicting the structure of proteins

The 2024 Nobel Prize for Chemistry, like the Physics award, included pioneering AI researchers. Drs. Hassabis and Jumper received the prize, along with Professor Baker, for their contribution to the long-standing protein folding problem: how does a chain of amino acids obtain a specific 3-dimensional structure? DeepMind's AI model for protein folding is called AlphaFold.

(d) Amazon Kendra

With over 200 products, knowing which does what is a major headache for developers using Amazon Web Services. Even just considering machine learning and AI products, there are more than a dozen.

If you are just getting started with building AI applications on AWS, I'd recommend going with Bedrock first. Bedrock gives you access to many frontier models, including those from Meta (Llama) and Anthropic. The usage is also quite similar to API-based pattern when you directly integrate with OpenAI or Anthropic. If you have more specific use case, you can perform model fine-tuning with your own data.

Amazon SageMaker has been popular with data scientists since 2017 (compared to Bedrock which was introduced in September 2023). SageMaker has a feature called JumpStart, which acts as a "model hub" that allows users to access pre-trained models, including foundational AI models. Besides using the models as is, you can also fine-tune them.

Amazon Rekognition provides computer vision capabilities, like object detection and video processing. You can improve model performance by fine-tuning it with your own labeled dataset.

Amazon Kendra is part of AWS AI offering, but it's an intelligent search service, not an AI model provider. You can, however, combine Kendra with AI models to construct a RAG system (Retrieval-Augmented Generation).

(b) Amazon Nova

Even though Amazon has invested $8 Billion in Anthropic, and Anthropic's Claude foundational models are available through Amazon Bedrock, it shouldn't come as a surprise that Amazon has been developing its own AI models as well.

In December 2024, Amazon announced it's line of AI models called Nova. Amazon Nova includes both text and image generation.

Amazon Aurora is a popular relational database service.

Amazon does not have a Sora product. In December 2024, OpenAI released Sora, a mesmerizing video generation model.

Amazon also does not have a Dora product, at least as a part of the Amazon Web Services. There are, however, many Dora products available on Amazon.com.

(b) Willow

We close the AI quiz with a last question on quantum computing. Because there's no rule that there can only be one hype train at a time. And while the AI hype train is going full steam in 2024, the quantum computer hype train got a powerful kick at the end of the year. The breakthrough from Google's Willow quantum chip is that its error reduces as it's scaled up. This is a hopeful step towards being able to assemble many quantum chips with reliable performance.