Bounce and Bound, Out of the Pot!

Following up from my previous post on what can we expect of the coming wave of AI, this article discusses what we can do to continue thriving alongside AI

How do we prepare for that?

When helping build a conversational AI tool in Google to help our internal teams get answers faster, we learnt that in the early stage, experts’ inputs (albeit manual) are crucial in calibrating the model results to meet users’ expectations, with respect to the accuracy, completeness & depth of responses.?

The domain experts add tremendous value, in the following dimensions:

1. Knowing the right metrics and required data sources
2. Defining the appropriate levels of granularity & amount of data in the inputs & outputs for the results to be meaningful. This is especially critical when data is to be used privacy-safely, because instead of gathering every single piece of data, experts can define the requisite data. This reduces the risks of violating privacy regulations
3. Increasing the probability of having sufficiently different data to generate accurate responses across varied circumstances. Niche use cases or fragmented industries likely require data collaboration to have a critical mass of amount & diversity of data to train the model. There are 2 parts to collaboration: (i) trust across competitors (ii) right data infrastructure to protect the integrity of data. This willingness to collaborate can only happen at the industry level, which requires negotiation among stakeholders, something an industry veteran can do, probably not an AI agent anytime soon.
4. Understanding the potential pitfalls of synthesized data - The experts are knowledgeable about the potential pitfalls of synthesized data. While still under-developed, the ramifications that can have major negative impact on Gen AI, including (but not limited to) infringement of Intellectual Property, misuse of data, inadvertent cross-competition data sharing… More will inevitably follow as GenAI usage picks up. Experts know their stakeholders and environment the best, hence can collaborate better with respective industry bodies to draft regulations that are right-for-innovation in each industry. While there will be an umbrella concept underscoring the foundational principles governing the use of GenAI, there will likely be tailored regulations for different industries, as different industries have different risk-reward trade-off in the use of GenAI.
5. Refining the responses of the most frequent use cases to ‘constrain’ the models in generating responses for questions without precedence
6. Integrating into the right business processes to show the value of AI, while not exposing its underbelly

One should definitely have a basic understanding of how the technology works and ideally, have used it before. But not everyone needs a PhD in AI to work alongside AI. Domain expertise will still be important but it’s the application of that to guide the adoption of AI that will be critical. This requires practice, hence the earlier you can start, the better. So even if you cannot acquire a set of wings (PhD in AI), you can still strengthen your muscles (build on your expertise) to get ahead.

Drill for the right data

What is the Minimum Sufficient Data (MSD) for your use cases?

Yes, we know we need the right data for an industry in a particular language. But what exactly do we need? You might be wondering why you cannot use everything you have, because your lawyers will flag regulatory concerns; your data scientists - data integrity and cost… This is where starting earlier can help you think through and be intentional about bridging the gaps.

At its core, the best way to maximize the value of Gen AI would be to provide/describe these 3 scenarios (be it text, images, videos…) for every instance:

• Scenario 1 - What is the current state of things?
• Scenario 2 - How are things in a new environment/set of circumstances?
• Scenario 3 - How should things be ideally in this new environment?

Answering these scenarios requires a lot of manual effort and guess-timations. Hence, we should be intentional about collecting them in the right context.?

These 3 scenarios can be abstract, hence I will use these typical use cases to illustrate them:

• <In Consumer context> Consulting a store associate is still critical when purchasing higher-value items, such as finding the right Car, TV, Smartphone, Fridge…
• <In Professional context> Define the prioritization of projects, product features or team make-up to grow a business

1. Scenario 1 relies heavily on answering long surveys or spending a lot of time on Q&A -?<Consumer> The associates will spend a lot of time understanding your preferences & use cases.?<Professional> Supplement dashboard with talking to people to gather feedback about product features, projects or someone’s performance
2. Scenario 2 requires a lot of research and again, talking to people. More often than not, the circumstances differ, even for the same use case, so we also need to find relevant benchmarks to interpolate from our current state to the new state. Successful interpolation is premised on a good understanding of the unique circumstances behind the current & new states. However, these nuances are not captured systematically, hence limited information exists for inference.?<Consumer> The associates will share other users’ experience<Professional> Talk to people to understand the context behind the success or failure of something, or the stakeholders that are critical to bring on board
3. Scenario 3 requires creativity & stretch of imagination - This is where we want the most help from AI. While AI might help with some interpolation, it only works with the right data. Hence, it is crucial to (i) have a good understanding of what the new state would look like (ii) proactively pilot with users to generate sufficient data & get good feedback

Now that we understand the types of questions for which we need answers for, the key considerations for your MSD would be:

• Identifying metrics that are meaningful to the outcome, instead of everything under the sun?
• Having sufficient data to be representative, instead of being a statistical anomaly
• Having diverse data to cover as many circumstances as possible
• Gathering inputs that are collectively exhaustive such that the response can be complete
• Implementing a strong feedback loop between inputs and outcome, thus providing smart inputs to the algorithm?

Sufficiency and diversity can be difficult to achieve concurrently (e.g. a lot of data for 1-2 circumstances; nothing in other). Hence, data collaboration and intentional data acquisition are critical.?

Not having this MSD is not the end of the world, but the outcome will be less reliable. Hence, another hack while building this out is to phase out the usage –?

• ‘confine’ to circumstances for which there is good data to train the model;
• ?collect more data;?
• expand to other circumstances.?

It is important to understand these gaps up front, so you can design the right coping mechanism. This depends on the risk-reward trade-off and stakes, which is again why domain expertise is critical in getting this off the ground. Otherwise the project might just be starved of nutrition just when it needs it the most.

It should be obvious now that this involves different expertise and potentially successive iterations, hence influencing the right stakeholders to join forces is equally important.

Understand your stakeholders & processes

How comfortable are your stakeholders with uncertainties?

What do we mean by 'stakeholders need to be comfortable with uncertainties'? These are some important considerations when comparing human vs. AI in making decisions

• Are we solving the right question? Models guess what you ask (hence the importance of prompting when using GenAI); Humans tend to know the context and can read between the lines
• Do we get the right answer? Models give the most probable answers. So when the information available for inference is limited and the probability across scenarios low, the answer is as good as a coin toss. Some responses can be too extreme where opposing perspectives are provided, making them too generic and not useful. While humans can give a more definite answer, we fall prey to overconfidence - more often than not projecting way more confidence than warranted
• Who should be accountable if the outcome is wrong? Instead of scrutinizing what someone says from every possible angle, one ingenious way to get around overconfidence is holding people accountable for what they say. But who can you fire if the model gives a wrong answer? Hence, if your culture is one where someone needs to be singled out everytime something goes wrong, AI adoption will be challenging.
• Who should be liable if the process is deemed ‘unethical/illegal’? Similar to the above, while organizational culture & hiring are important, the other key tool critical in deterring unscrupulous behavior is lawsuits. Who can be sued if AI (un)intentionally synthesizes information - the end user, model developer, researcher behind the paper…? For example, the landmark New York Times suit against OpenAI/Microsoft.?

Answering the above questions has nothing to do with technical expertise, but more relationship-building, communication, influencing, leading changes… This can be firm/industry-specific, but once you get over that bump, these are some general steps/principles you can take to smooth out the adoption.

• Understand the variance in the scenarios. High variance = higher uncertainties, so phase the roll-out in stages, start with small-scale pilot and refinements, then launch on a large scale only when we are comfortable.?

• Pre-empt the worst-case scenarios, e.g. types of decisions/responses and discuss the following:- Is the team comfortable with having to deal with the aftermath? In some cases, the reputational risk might not justify this.?- If the team is comfortable, plan in advance & automate the coping mechanism so that things can be addressed swiftly and damage contained- Build a feedback loop into business processes to reduce errors moving forward- If the circumstances are very different from the data used to train the model, it is recommended to give a canned ‘I do not have sufficient information’ response or suggest a next best answer and clarify the constraints, instead of giving a very wrong answer.?

• Categorize key decisions into this 2x2 matrix of frequency/volume of decisions vs. the stakes at risk if things do not go well.

• Frequency/volume: The higher the frequency/volume, the more can AI play a role. 3 drivers at play here: (i) Sometimes it is just physically impossible to not automate the work (ii) Almost impossible to ensure that decisions will be consistent across the different people doing it at different times (iii) High data volume means that the algorithm can be trained on more diverse scenarios, hence more robust
• Stakes: The lower the stakes, the more likely that AI has a role. Lower stakes decisions tend to have higher tolerance of failure, hence perfect for AI when the underlying model is still being perfected. AI plays the leading role, while humans handle escalations

• Some management teams might salivate at the prospects of ‘less people doing more things’. While that is true, it's also important to not expect everyone/thing to be as efficient as before. When s*** happens, the time needed to address these outliers/thorniest issues will be much longer than expected. Especially if your organization is so streamlined that those expertise might have been lost and you need to engage external help. Hence, when doing re-organizations,?- Restrain from being highly precise/surgical- Build some buffer- Have some of these experts on speed dial in case things go very wrong.

Bring everyone along, even for people who might be impacted. Up-skill people who could be impacted in advance for new skills required

Examples of lower-hanging use cases which might be soon disrupted:?

• Consumer service which requires high customization (e.g. travel, shopping for goods & services)
• Internal-facing analytics/consulting/support
• Workloads that are highly cyclical, requiring massive expansion & contraction of workforce in the past

Anyone who has tried swimming knows this - it’s almost impossible to jump out of the pool, unless your feet can touch the ground. Even then, the water friction is so huge that one can barely move up an inch. Hence, the only time to jump out of a boiling pot is before the water level has risen substantially. Because no matter how slow the trickle is now, as long as the water keeps flowing in, it will eventually be high enough that one cannot jump out.?

Unlike not getting on board when Uber/AirBnB could not give you a satisfactory match and only coming back later when everyone else is using it, for some of us, not consciously optimizing for GenAI now is the difference between the frog who jumped out of the pot and the ones left behind, getting all toasty. Even if we might only be using GenAI in our personal lives or something trivial in work, DO NOT brush that aside. Keep learning!

Which frog do you want to be?

An addendum exploring if Human & AI will be in a cycle of symbiosis here