5 Predictions for Intelligent Apps in 2023

What can we expect to see in the New Year?

How DALLE-2 interprets “AI Predictions for 2023”

The new year is almost upon us, and as 2022 winds to a close, the tech world is inundated with an endless stream of recaps of what happened over the past year, and what we can expect to see in the future. Making predictions in technology can often feel like a Sisyphean challenge given how quickly the world changes (would anyone last year have guessed that an AI chat app would exceed 1M users in five days?), but they are fun to do, so we’ll add our voices to the choir below.

Here are five areas in the intelligent applications world to watch out for in 2023:

1) The Arrival of GPT-4

In terms of the most hotly anticipated product launches for 2023, few rank higher than GPT-4. We expect OpenAI to announce the release of GPT-4 sometime in the first half of 2023, and like many others, we believe it will be a complete step change from GPT-3. Many of us have seen just how powerful ChatGPT is, which is fine-tuned from a model in the GPT-3.5 series, an improvement on GPT-3. However, we believe GPT-4 will continue to take another major leap forward, due to a few key factors:

• Availability of better and more recent training data: One of the critiques of GPT-3, is that the model is frozen in time and doesn’t know about anything prior to 2021. This is because the model was trained on data and information that only goes up to 2021. We believe GPT-4 models will be trained on current and up-to-date data, greatly expanding the possible use cases and queries.

• Optimality: As Albert Romero eloquently points out, due to the highly expensive nature of training AI models, companies have historically needed to make tradeoffs between accuracy and cost. OpenAI will likely achieve another breakthrough with hyperparameter training in GPT-4, resulting in achieving better performance at similar cost levels.

• Reduced biases: One of the common complaints about ChatGPT is it can be prone to bias. OpenAI even stated in its blog that ChatGPT “will sometimes respond to harmful instructions or exhibit biased behavior.” We believe GPT-4 will curb many of these biases due to building upon richer and more robust data sets, though by no means will be perfect.

We’re excited to see the power of GPT-4: whether it will be able to get a perfect score on the SAT; build an intelligent application from scratch with just text inputs; create a short Netflix documentary on the impacts of Climate Change; or book a 10-day backpacking trip in the Swiss Alps at the lowest possible price point. GPT-4 may even start to eat into Google Search, as we are starting to see with ChatGPT…

Ben Tossell @bentossell

ChatGPT is just mass scale user testing in final stages of GPT-4 prep right?

2) Generative Apps Go Mainstream

The second half of 2022 saw an explosion of “Generative AI”, with applications ranging from Jasper to ChatGPT becoming popular in the AI/tech/creative world. However, many of these apps were still limited to a relatively small group of early adopters or technologists willing to dip their toes into something new. We believe 2023 will be the year generative apps truly go “mainstream”, where virtually all corners of the general public will be exposed to gen AI, due to a few key reasons:

• Multi-modality: Most generative apps today are largely centered around textual use cases (text completion, text summarization), but we believe in 2023 we will start seeing true multi-modality: apps that can handle not only text but also video, speech, music, protein generation, actions, and other modalities we can’t even fathom right now. We also predict that there will be an increasing number of Generative Apps that will be fully automated (i.e., AI can autocomplete tasks without any human intervention).

• Traditional Incumbents “baking in” Generative AI: We are already starting to see a whole host of software companies like Coda, Notion, and Canva leveraging OpenAI and Stable Diffusion to help users expand their text, image, and creativity. We believe that later-stage software companies across all verticals will implement features in their platforms to become more competitive and this will create an interesting dynamic for new investments. For example, imagine CVS leveraging FMs to better predict and issue prescription refills, or AT&T replacing legacy call trees with enhanced voice-based commands in customer service.

• Commoditization of basic generative features: We predict that basic features like creating images for static usage (i.e., PowerPoint slide backgrounds), text autocomplete (i.e., suggesting the next word that comes in a sentence), and summarization will become table stakes for all applications. In order for generative apps to be defensible, they will need to create a flywheel that includes (1) data-evaluation datasets (i.e., data as the customer interacts with the model); (2) strong UX/UI; and (3) strong GTM/distribution.

3) Costs of Training & Building Models Get Cheaper

Building, training, and deploying foundation models can be an expensive exercise; for example, training GPT-3 reportedly cost upwards of $12M for a single run. The massive expenses associated with these projects historically meant that only those with large pocketbooks or the ability to raise tens of millions in venture capital could build FMs. However, we think that will change dramatically in 2023.

• Model hosting costs will meaningfully decline: Purpose-built infrastructure, reduced production run costs, and less required training time will contribute to a drop in build costs. One example of this is the recent launch of a new GPT training service from Cerebras System and Cirrascale that offer a rental service for training your GPT models. Other companies such as SambaNova Systems, Graphcore, and Intel’s Habana Labs are in the business of reducing costs for training and building FMs.

• As the cost of training and building models gets cheaper, models will be controlled by a broader range of companies beyond hyperscalers and other large tech companies: Many early-stage companies, like Jasper, are eager to research and build their own large-scale models for specific business applications; however, the infrastructure that currently exists makes it difficult. We are starting to see other companies like Adept building and training their own models.

• An emerging category of ML Infrastructure optimization: We define these two categories as ML Efficiency and ML Hosting. ML Efficiency is a concept that allows companies to train any size model on any number of GPUs and achieve more accurate results and ML Hosting is a concept that allows companies to run or deploy ML models, massively parallel compute jobs, web apps, and more without needing your own infrastructure. We are excited to learn more about emerging companies in these sectors and think this is a hot area to watch out for.

  

  One analyst’s forecast of how training model costs are declining

4) Increased Regulatory Scrutiny & Established Regulations in AI Copyright

As we know from past experiences ranging from aircraft flight to the internet, with new technologies comes increased regulatory scrutiny and structures. We don’t expect things to be any different in the world of intelligent apps. In 2023, we are keeping an eye on a few specific areas that will come under the regulatory microscope:

• Copyright infringement for Generative AI will be front and center: Generative AI models are trained on a large corpus of copyright-protected data, which begs the question - is that legal? The Verge published a detailed article on the key legal questions to consider as it relates to AI models. We believe there will be more active debate about copyright issues and a whole host of early-stage companies emerging trying to promote the responsible use of AI (i.e., Credo).

• Case law will be established about the relationship between training data and the models that result: Earlier this year, there was a class-action law­suit filed in a US fed­eral court chal­leng­ing the legal­ity of GitHub Copi­lot and the related OpenAI Codex. The suit against GitHub, Microsoft, and OpenAI claims violation of open-source licenses and has yet to be settled. This raises the important question of how you can produce output on copyrighted materials. We predict there will be misconduct and misuse of Generative AI that will result in the government stepping in to establish case law on what is acceptable.

• New “AI compliance standards” emerge: Similar to how there are regulations and requirements around consumer data privacy laws (i.e., GDPR, PII, HIPPA, etc.) we believe there will be new requirements for AI compliance that emerge that may affect how generative AI can be used.

5) New Job Types are Created while Traditional Roles are Redefined

While artificial intelligence and machine learning have existed at some level for several decades, top AI and ML talent (often with Ph.D. credentials) have typically been constrained to large tech companies (like Google and Meta) or academic institutions. Now with foundation models democratizing access to AI across many sectors and company sizes, we believe both i) new roles will be created that previously never existed, and ii) current roles may become more obsolete or transition to other areas:

• New roles created: Similar to how the explosion of big data in the past decade led to the creation of several new positions (data engineer, data analyst, etc. ), foundation models and intelligent applications will herald the arrival of newly created positions like Prompt Engineer and Chief AI Officer. In fact, Scale AI recently made headlines the other week when they announced the hiring of Riley Goodside as their first (and likely one of the world’s first) “Staff Prompt Engineer”. We believe we’ll only continue to see companies both large and small embrace and create these new roles in 2023.

• Existing roles expanded or transitioned: The rise of applications leveraging foundation models will make it easier for companies to implement AI/ML without requiring the deep, hard-to-obtain expertise previously needed. For example, software engineers are already leveraging GitHub CoPilot to write up to 40% of their code, freeing them up for other higher-level tasks. We expect that in 2023 we’ll continue seeing traditional roles (like engineers and analysts) redefined in the era of foundation models.

Alexandr Wang @alexandr_wang

Today, @goodside joined @scale_AI as a Staff Prompt Engineer. I am going to assert that Riley is the first Staff Prompt Engineer hired *anywhere*. These large models are a new kind of computer, and I’m excited for @scale_AI to lead the way in programming them.

Bonus: Consolidation in the Modern Data Stack

We predict 2023 will be a year of M&A with consolidation in the Modern Data Stack. In 2021, we saw an abundance of data companies backed by VC investors, many at eye-popping valuations. This resulted in a disjointed and sprawling Modern Data Stack. We predict 2023 will be a year of M&A as many of these point solutions have struggled to gain market traction and many CIOs will start cutting back spend. We believe Enterprise customers will want more consolidated offerings that are easy to deploy and maintain, which may end up coming from their native cloud providers vs. from hundreds of different vendors. We also predict VC investors will be less excited about the Modern Data Stack and more focused on the power of FMs and Generative Apps in comparison.

---

Funding News

Below we highlight select private funding announcements across the Intelligent Applications sector. These deals include private Intelligent Application companies who have raised in the last two weeks, are HQ’d in the U.S. or Canada, and have raised a Seed - Series E round. We saw 12 new funding announcements over the last two weeks.

A special shoutout to our newest Madrona portfolio company, Runway. Madrona invested in Runway’s $50M Series C round alongside new lead investor, Felicis Ventures, and existing investors Amplify, Coatue, Compound, and Lux Capital. We share some more of our insights on why we believe Runway will be the leader of the next-generation content creation suite here and are incredibly excited to partner with Cris and the entire Runway team.

New Deal Announcements - 12/05/2022 - 12/15/2022:

We hope you enjoyed this edition of Aspiring for Intelligence, and we will see you again in two weeks! This is a quickly evolving category, and we welcome any and all feedback around the viewpoints and theses expressed in this newsletter (as well as what you would like us to cover in future writeups). And it goes without saying but if you are building the next great intelligent application and want to chat, drop us a line!