Generative AI: Get on board, Evolve and Succeed

Dec 18

Article written by Carlos Dangelo and Philippe Ciampossin

This article is an overview of AI’s Generative Models reflecting personal views of the technology and opportunity for advanced value creation in your day-to-day job.

EXECUTIVE SUMMARY

Business leaders and professionals must become proficient at interacting with generative AI to stay competitive. This brief summary provides key questions to consider when evaluating foundational large language models for decision-making.

While current GenAI includes reinforcement learning as part of training, we will skip it here given the field's rapid evolution. This brief focuses on:

The rise of human-AI hybrid workers
Understanding the basic elements underlying generative AI technology
The Art and Science of Prompting Generative AI
Evaluating Current Generative AI Capabilities and where it provides value for business
Offering tentative conclusions and guidelines for navigating these tools

The goal is equipping professionals to interact productively with generative AI as collaborators, while applying thoughtful oversight. With hands-on learning, leaders can unlock creativity and productivity gains from AI collaboration.

The rise of the hybrid worker

As a professional in today's data-driven business landscape, adopting generative AI is becoming essential to stay competitive. While these technologies hold great promise, it's wise to approach them with open eyes regarding potential pitfalls.

We are entering an era of unprecedented human-machine collaboration. The symbiotic brain relationship with technology really began with smartphones (who can remember days before mobiles) and is expanding as AI tools are further woven into daily life through the technology. Now we are at a stage where it does really start to complement us in acquisition of knowledge and our reasoning abilities.

Generative AI can accelerate acquiring new insights and magnify thinking by interacting with a system that occasionally seems eerily human. However, fully realizing the potential of generative AI requires recognizing these technologies as aids for collaboration, not as oracles. The core mechanism aggregates information and knowledge in a way so that business professionals avoid needing to ingest the entire corpus of data, instead allowing targeted querying and iterative content refinement through added context.

To maximize the potential of generative AI, we must use it in a closed-loop fashion with human oversight – the AI tools may never be ready for autonomous thinking or worse, making decisions without human oversight. While generative AI provides intriguing information and insights, you must refine, fact-check, and validate that the output makes sense for your specific problem. The key is to augment your reasoning capabilities.

Experts recommend professionals to explore hands-on testing to determine how best to integrate generative AI into day-to-day work. If you haven't already opened a dialogue with management about these tools, those conversations will likely come sooner than later.

Let’s look at the basics on how it works to understand the Gen-AI engine behind the prompts

Learning

Learning is derived from input Data, Information, Context, human inputs all based on their Probability Distributions over time and space. Foundational LLM ( Large Language Model) is the foundation of Generative AI. It learns from its Data and dynamically refines it within the context of conversational sessions.
Input data are tokens (words,sub-words), numbers, pixels. Internally, algorithms convert everything into floating point numbers for internal model representation. Time series or word sequences are examples of data in context. They are represented internally by vector embedding. Data sets are vast collections of examples of relevant data out there.
LLMs are usually pre-trained over a large set of strings or text sequences, optimized over a large number of learning parameters (Neural Net weights). Training from scratch is computationally expensive and time consuming. Once done, it is re-used and augmented.

Inference

A learning step usually predicts an outcome by computing an error w.r.t. input sequences and then uses optimization algorithms (typically gradient computations) to minimize such errors. Inferences of outcomes are derived from learning.
Probabilistic scores are used to compute and infer the most probable outcome conditioned to given user inputs.
Generative, pretrained LLMs commercially available for deployment (2023):GPT3.5, GPT4, GPT4-Vision, LLamas, Bard, Claude, plus others are available as open-source or proprietary models. Proprietary models do not disclose training data and detailed algorithms.

A word about Model Bias

AI Models are trained from “sampled data” available from a universe of Big Data. The probability distribution of data is often unknown but can be skewed before and after sampling.

Moreover, because of the explosive growth, data is compressed and aggregated, reduced to numbers for processing. Both steps introduce potential skewing, biases, violations of ethical rules and/or censoring of free-speech. Countries (Europe, US, China) have different standards for bias, ethics and free-speech. Standards are evolving. While the technological advances in AI are exhilarating, these issues have put significant pressure on global governance, prompting another AI race for governments to regulate AI tools.

For some use cases (e.g., money lending decisions) there are clear guideline laws for what field of data are to be excluded: gender, race, location, etc. Also, guidelines exist for use of copyrighted material, use of national security data, etc. Pre-processing, curating and anonymizing data can alleviate obvious biases but they are not enough to deal with privacy issues, potential data leaks, etc.

The Art and Science of Prompting Generative AI

While community guidelines exist, prompt engineering remains more art than defined science at present. Generated outputs often require refinement via user feedback given through prompts.

Prompts are tokenized input sequences that, combined with the LLM's learned internal representations, yield probabilistic textual completions.. Successful user interactions depend on how one frames questions plus context to elicit high quality answers. Mis-interpretation and spurious (made up) answers is not uncommon.

Emergent prompt "recipes" offer starting points, but the model evolves dynamically as its parameters shift during continued fine-tuning on new usage data. LLM uses ‘transformer’ architecture to encode token sequences into a ‘number vectors’ for parallelized execution of processing in NN ‘attention’ layers to arrive at the most probable next token in input sequences. Reinforcement Learning from human feedback is deployed to better ground the Model prior to release to the public.

Transformer methods have made inference fast enough for human interactions but expensive to process due to costly GPU hardware. Periodically, Training becomes obsolete and needs to be re-done as new information in NN ‘attention layers’.

Prompt chaining should be seen as a way to implement AI workflows. By reformulating prompts interactions with the model you can steer outputs by providing additional context. You need to drive the interaction to the level you need it to be. As a business professional you will also need to own the AI output as your responsibility, not as the machine’s fault.

ChatGPT also offers the capability of defining ‘custom persona’ as another way of "warming up" the system pre-prompt phase by configuring some attributes defining the optic you are looking for and the type of output you are expecting.

We seem to already have a new version of « the dog ate my homework » going around in lieu of someone claiming that the AI made a mistake. At work it will not be received well, work so our main advice is don’t ever use the first output, carefully review and refine it before you deliver anything (code, report, analysis)

Evaluating Current Generative AI Capabilities and where it provides value for business

Based on our recent hands-on experience, today's generative AI shows promising capabilities but still requires human guidance and verification. We have experimented with ChatGPT for various personal interests including AI market research, technical paper summarization, code generation, and exploring hypothetical business scenarios.

The results have been surprisingly good but imperfect. Even open-ended prompts yield some intriguing insights, though success is not guaranteed. For wide adoption of AI, Trust built from examples of ‘good (plausible)’ responses of an LLM is important. Responses along human Q/A sessions need to be checked before acceptance. Natural Language descriptions are full of ambiguity, redundancy, logical inconsistencies, etc. It only plays against us.

Current real-world applications of natural language processing include (but are not limited to) :

Text summarization of documents, meetings, and recordings - extracting key insights from complex content.
Content creation such as websites, advertisements, blogs and books - AI-assisted drafting can amplify human creativity.
Small-scale code generation in languages like Python and Java - helping programmers with rote coding tasks. Large scale self generated software that directly results in actual products is still not there. Also, there are issues about IP infringement and who really owns the generated code. One of my teams successfully automatically translated old perl test code into python with some success so we can see the potential is here.
Workflow generation (from data flow diagrams), documentation and summarization while distilling processes and contexts.
Small scale data analytics based on built-in algorithm and datasets on the smaller side

Remember that this technology doesn’t really create new content but more or less summarize and “assemble” the answer it gives based on a probabilistic outcome of the next most probable token suggested from its ‘knowledge base’ and state of its neural network. Sometimes it will be right, others totally off the mark.

So using it beyond its capabilities is interesting but not that good in terms of open ended questions, asking to solve complex problems or asking for opinions on sensitive topics as GenAI is not a reasoning tool per se.

Generative AI should be viewed as a tool to augment human capabilities, not an oracle that can replace human expertise and judgment. These large language models amplify mental models through conversational interactions. Leveraging this technology generates valuable insights when incorporating domain-specific examples mimicking real-world problems.

Conclusion and guidelines for navigating these tools

It is now a business imperative for professionals at all levels to educate and start to incorporate the usage of such technology in our day to day activities to stay competitive but the key is to understand that you must interact with it and drive the direction of its output to get the best results.

Generative AI works very well but the noise to signal ratio is still high. You need to use your experience (and brain) to make sure that the information you are leveraging is accurate and ethical. The burden is always on you to trust but verify what you get and somehow get ‘good enough’ closure for the problem at hand. Upon closure, then use, cut and paste the results.

Using the tool is not hard. It’s currently more an art than science. We are now in a world in which AI is fed more and more information and someone may already wonder who is the best authoritative source, human or machine ? Algorithms running in the machine do not include common sense, rules of logic, etc. Only pattern matching approximations. So while it might look very good you need to remember that it is a probabilistic technology and non deterministic no matter how good the approximation is.

LLM creates economic value when speed of response is key and the knowledge base to be leveraged is larger than you can acquire by yourself and you fully use your experience and judgment to use it to your advantage… You need to get started on the GenAI Learning curve by getting on board, adapt your mental model and skills to succeed by discovering what works! Add your own common sense, logic , rules and ethics as you go along in the journey to make sure that you are not delivering something that could crater your career at your company.

philippe ciampossin