Search ST Engineering

ST Engineering
ST Engineering

Generative Artificial Intelligence (GAI) has been all the rage over the past few months, particularly in the valley. Views range from the hyper-optimists who view artificial general intelligence as something achievable within 5 years (down from 10–15yrs), to the more skeptical who see the industry being at peak-hype, with use cases constrained by reliability and privacy issues among others. Some also see generative AI being a mere extension of how we used to do deep learning and NLP in the past due to elements that are so reminiscent of traditional AI/ML (hence “hype”). They point to proprietary data continuing to be the key differentiator (more on data-centric AI here). We continue to need a human-in-the-loop to refine the Reinforcement Learning from Human Feedback (RLHF) model.

However, foundation models have a fundamentally different architecture compared to those that have been used in deep learning before. The foundation models have their roots in the transformer architecture that relies on attention mechanisms. This architecture dramatically improves customisation to a context by allowing for much more flexible input representations where the model can capture long-range dependencies in longer input sequences (this was impossible before). The attention mechanism also inherently causes the model to focus on the most important features, leading to more personalised predictions. Fine-tuning larger models on very specific datasets further allows the models to handle more complex customisation requirements. As a result, while the last decade was all about classification and prediction, we are in an era where customisation will become much easier and cheaper; with an explosion in related use cases.

Some of these use cases include:

1) Autonomous agents where you give the AI a high level objective e.g., “help me make money on the internet” and the AI would come up with a to-do list, complete the tasks, add more tasks based on its progress until it meets its objective. This area has really exploded in the past couple of months given how immense the opportunity is, just in the litany of tasks and industries it can be deployed in.

Under this umbrella would be personal assistants; imagine a world where everyone had an agent that would function like your own Chief of Staff, paid by you to adversarially interact with other agents, planning your tasks, negotiating with them for your benefit. The agents would be very interactive and dynamic, communicating, getting feedback and updating the model to have outputs that are bespoke to you. It could be a personal assistant for work and home, for B2Bs to manage very repetitive tasks, for B2Cs to customise their recommendations whenever you interact with them through a website or app. Companies like Inflection.ai, Character.ai, Personal.ai, Microsoft, OpenAI and others have been working on a variations of this. The technical challenges here center around trust, data privacy and the ability to run some portion of the software on users’ devices. Companies tackling this space are really in the early innings and present an opportunity for us to build capability in. However, we should note that there is a heightened element of risk with autonomous agents. There are those that believe that should these agents pursue recursive self-improvement or aggressively pursue autonomy, that’s when the GAI should be taken off the table.

2) Personal tutors, except that this time it is an expert on every topic on your phone. A much lower risk use case while presenting an extremely efficient way to consume a vast amount of information!

3) Co-pilots for copywriting and software development have been extremely successful and are for now, the breakout use cases for generative AI. Github Copilot is the world’s most widely adopted AI developer tool just a year into its launch; more than a million developers have used it and it is turbocharging productivity with users reporting acceptance of 30% of suggestions on average. Anecdotally, some companies in the valley have reported laying off 50% of their development team primarily due to the efficiency gains from copilot (beyond just overall cost cutting). These co-pilots represent the quintessential use cases for generative AI; the products succeed when the use cases are resilient — the mistakes don’t hurt and are easy for the users to spot and correct.

Generative AI is a hugely transformative, omni-use technology that isn’t constrained by atoms, only bits. We are very early in the adoption cycle. The speed at which it is progressing is also extremely fast, “a month in tech is day in generative AI”. As enterprises consider whether to adopt GAI, we shouldn’t be too quick to write off the technology due to the hallucinations or mistakes the models makes. This is because, as Ted Sanders, an OpenAI ML engineer shares, reliability is a tractable problem; the quality of output is a function of both the model and the input. There are so many improvements that we can make in system and prompt design. As we’ve seen in copilots, we can have success in use cases where it is easy for users to check the output and continuously iterate with it, improving the model over time.

In addition, as we consider where we or our customers have data around which we can build a product, it’s important to note that “proprietary data” is extremely nuanced and context-dependent. The strength of a data moat depends on many factors:
(1) how fast the data decays (which might necessitate the ability to acquire new data frequently);
(2) how unique it is (security threats are unique to your company but the delta diminishes when you consider incidences across many organisations);
(3) how performant synthetic data is (we are starting to see AV edge cases perform as well on a corpus of fully synthetic data);
(4) the distribution of the data (do we have enough data to account for edge cases);
(5) how expensive and how long it would take to amass the data needed to attain certain performance benchmarks, etc.

In closing, the nature of the architecture is so different from what traditional ML/AI teams are used to that even Big Tech organises hackathons to allow teams to tinker and build intuition about the models. ChatGPT has given us a very user-friendly interface to start building applications in a low code paradigm and I would encourage all employees to lean in and start building intuition around its capabilities.