The AI Unbundling of Upwork
"$250b of India’s GDP exports are essentially GPT-4 tokens… what happens now?" -- Daniel Gross
Summary
Freelance knowledge work has become a popular way for individuals to work and companies to get work done. The rise of freelance work was material, until the end of the ZIRP era. Now reduced investment, generally, by businesses and the concomitant birth of Generative AI, might undo this. Upwork and Fiverr CEOs say that AI is actually driving higher value projects, but this role might be temporary. As models improve, the need for humans to orchestrate across Gen AI apps will decrease. As people become more used to using Gen AI, they won’t need to turn to freelance experts to help. The Gen AI absorption is nigh
.
Contingent worker is increasingly popular for workers and employers
Freelance work continues to grow in popularity. According to research by Upwork, nearly 40% of the workforce freelances1. This masks a material difference by worker age cohort. 52% of the youngest part of the workforce (Gen Z) freelances, only 26% of the Baby Boomer generation does. Freelance work is popular choice, but there are challenges, and freelance growth is slowing.
Upwork and Fiverr GMV has stalled out
The Gross Merchandise Value of work done on Upwork and Fiverr has stalled out. Two explanations: 1) the macro economic situation 2) the rise of Generative AI.
When rates started rising, businesses put investment plans on hold to see the impact to the economy. This brought reduced investment in headcount, software, and contingent work. The other, concurrent trend is the rise of Generative AI.
Generative AI is eating up the low end up freelance work
The low end of freelance work, the work that ChatGPT and other LLMs can handle, is getting done by Gen AI. An August 2023 study suggested this and now the CEO’s of Fiverr and Upwork are confirming.
The 2023 study looked at the number of jobs and monthly revenue for job categories deemed most likely to be impacted by ChatGPT. The study found those categories saw declines after the launch of ChatGPT.2
While the lower end of the work market is getting taken by Gen AI, it is opening up more work at a higher end of the market. During an earnings update in February of 2024, the CEO of Fiverr, when discussing the impact of Generative AI:
We see a category mix shift from simple services, such as translation and voice-over, to more complex services, such as mobile app development, e-commerce management, or financial consulting.
Freelancers are moving to the orchestration layer
Freelance work is moving to the orchestration and development layers of work. Right now the majority of Gen AI use is done in a zero shot way. Give the model a prompt, get back an answer. Humans think through the actual use of the model output, judge the model output, and copy and paste to connect the model to other services. Workers leveraging AI to get work done are faster, more efficient, and higher quality3.
Upwork is learning into this. Upwork has provided Gen AI tools to its freelancers and it sees adoption of these tools growing. When asked about Generative AI’s impact on the marketplace in a February 2024 earnings call, CEO Hayden Brown talked about the use of AI tools by freelancers:
Sure, Ron. On the AI front, we're seeing really good traction so far in terms of the adoption of the tools that we've been launching for talent. So, Upwork Chat Pro, which was built on OpenAI’s GPT-4, has more than 150,000 signups. We've seen tools like Jasper, Adobe Tools, Amazon CodeWhisperer, and others definitely getting a lot of trial interest from talent. And overall, as with this platform shift to AI, what we've seen in other previous shifts around new tools and technologies, freelancers are always the fastest to move these technologies and ramp up their skills. And so this is what's happening right now. They're eagerly adopting these tools and really using them to drive performance metrics around productivity and quality of work. So, even though it's still early days in some of this, the indicators are there that this is happening. And we're going to continue to pursue our strategy this year and empower our talent with the best possible tools out there.
Part of the reason this works so well is that most model use is done in a zero-shot way: provide a prompt, get a response. What we see is that the same model, when used with more sophisticated techniques, can produce significantly better results.
New AI techniques will eat the orchestration layer
Andrew Ng, well known Data Scientist and Entrepreneur, highlights how new model techniques will enable models to go up the stack from 0 shot to orchestration4.
Today, we mostly use LLMs in zero-shot mode, prompting a model to generate final output token by token without revising its work. This is akin to asking someone to compose an essay from start to finish, typing straight through with no backspacing allowed, and expecting a high-quality result. Despite the difficulty, LLMs do amazingly well at this task!
With an agent workflow, however, we can ask the LLM to iterate over a document many times. For example, it might take a sequence of steps such as:
Plan an outline.
Decide what, if any, web searches are needed to gather more information.
Write a first draft.
Read over the first draft to spot unjustified arguments or extraneous information.
Revise the draft taking into account any weaknesses spotted.
And so on.
The improvements in model capabilities, and output, from more advanced techniques are material. From the same post, a comparison of AI performance on a coding benchmark.
These techniques enable AI to do more AND take over the orchestration work that humans are doing. Importantly, the pace of these improvements are fast.
Generative AI’s speed improvement is faster than anything we’ve seen before
Large language models are doubling in capabilities about every 9 months5. This comes from improvements in scaling, training, architecture, and even prompting. Models are getting better - fast.
Models are developing quickly, new techniques add new capabilities to existing models. These changes will bring about the fully autonomous agents. These agents will materially impact how Freelance AI gets done. An early glimpse of these agents is Devin, the first AI Software Engineer.
Combing model capabilities and external tools - Devin
Building on all these advances, Cognition Labs introduced Devin6, what it’s billing as the world’s first AI developer. Devin can:
Build and deploy apps end to end
Autonomously find and fix bugs in codebases
Train and fine tune its own AI models
Do real work on Upwork
Upwork’s CEO noted AI use cases in more complex projects, “we're really seeing the full range of AI experts being called upon, whether it's for training models, whether it's for data labeling and curation.” Except that work is about to be eaten up, too. Devin can do these tasks. The rise in these model capabilities will lead to the Upwork unbundling — by AI.
The Gen AI Apps Unbundling Upwork
Across developing, sales, marketing, creative, translation, customer service, and more Gen AI apps are gaining capabilities. Sequoia put out an early version of a Gen AI Market Map that is out of date, but still captures key companies.
The process is not: Gen AI app exists —> user switches to the app. Instead there’s a smoother phase shift.
How the unbundling process will unfold
Traditional views of technology adoption draw on Everett Roger’s work Diffusion of Innovations7, the common ‘innovators’ and ‘early adopters’ bell curve. Another perspective is ‘Six Stages for Learning to Use Technology’8 by Anne Russell of Queensland University of Technology. She added views: (1) awareness, (2) learning the process, (3) understanding the application of the process, (4) familiarity and confidence, (5) adaptation to other contexts, and (6) creative applications to new contexts.
With Gen AI, we are in awareness and learning the process, for most organizations. In this instance, consultants and experts play a significant role in helping to comfortably introduce the technology into an org. It could look like a Freelancer making use of the tools or a large consulting firm selling in technology changes. As users understand how to actually use the application, and gain confidence, they will go straight to use, skipping the expert. This is when the true Upwork unbundling will start.
https://www.upwork.com/research/freelance-forward-2023-research-report
https://www.deeplearning.ai/the-batch/issue-241/
https://arxiv.org/pdf/2403.05812.pdf
https://www.cognition-labs.com/introducing-devin
https://en.wikipedia.org/wiki/Diffusion_of_innovations
http://www.russellsynergies.com.au/pdf/RussellSixStages96.pdf