Context
The way we design and execute work is being challenged by the advent of LLMs. However, their widespread adoption is being hampered by their limited relevance, with exploiting their full potential proving more complex than expected.
Added to this inertia is an illusion on the part of the user, who confuses the objectives of the model providers (the quest for ASI) with the real capacities of the LLMs.
The promises of autonomy do nothing to improve this illusion, generating paradoxical expectations and irrational behaviors, shared between the passivity inherent in the fear of a planned replacement, and the desire to hasten the advent of the ASI to fill the disillusionment, thus denying the immediate, yet very real and concrete, utility of the LLM.
Lack of relevance and disproportionate promises convey an immediately erroneous perception of the usefulness of the models ofIA transformers.
So, from the very first uses, GPT models seem to us to be of little use compared to the efforts required to extract the equivalent of our expectations from them.
However, to perceive the immediate extent of the potential of LLMs, it is necessary to change the perspective, starting by remembering that they are only tools to support tasks, with humans remaining at the heart of value production.
The replacement will indeed take place, but it will be done under the impulse of the human who knows how to exploit the potential of the GPT models, and not by the AI itself, by the sole factor of intelligence, because without the intention of the human, theartificial intelligence remains inert.
The challenge of successful adoption therefore lies first in understanding the limits of LLMs, in order to grasp the extent of their potential from the right angle, and in learning to exploit it fully.
Understanding the limits
The problem with LLMs is their one-size-fits-all approach to language processing.
They produce answers that are too generic, diluting specificity, precision, originality, even going so far as to distort the truth to conform the characteristics of relevance to a weighted average value, statistically the most probable.
This phenomenon, intrinsic to the functioning of artificial neural networks, degrades the relevance perceived by the user and causes a loss of adequacy between expectations and the results provided.
The transition effort then seems too great to adapt the solutions to the unique needs of users, notably through the prism of another untenable promise, that of fine-tuning.
Complex, costly, time-consuming and not adapted to the perpetual changes in the professional environment, fine-tuning only imperfectly solves the problem of systematic generalization.
However, to counter this tendency towards generalization, one thing is enough: segment the tasks, then entrust them to hyper-specialized AIs, dedicated and focused on unique objectives, so that the attention paid to the tasks is optimal through hyper-contextualization.
By assigning micro-tasks to dedicated artificial intelligences, specifically programmed for these particular micro-tasks, it becomes possible to process each element with increased speed and expertise, without special technical developments, without reduction of perceived relevance and without absolutely resorting to fine-tuning.
This segmentation guarantees more relevant results from every point of view, thus improving the quality of the results generated and user satisfaction.
On the other hand, unlike fine-tuning, this approach allows continuous, flexible and easy-to-implement adjustments to refine the capabilities of specialized AIs, and strengthen their behavior according to the scenarios encountered during their operation.
It is still necessary to know how to segment by considering the variability of the objectives and intentions of the applicants, while including the prism of interpretation of AI.
The ESP method constitutes the answer to these problems, by approaching relevance through all the dimensions it touches, via a fine division of professions, disciplines and tasks into micro-tasks.
This approach goes beyond simple process improvement to ushering in a new era of efficiency and relevance in the AI-enabled world of work.
Definition of the method
The ESP method, acronym for “Exponential Segmentation Process”, is a strategic segmentation approach aimed at optimizing organizational and operational performance exponentially through cumulative effects in which each task amplified by AI amplifies the others.
The method is based on the principle of breaking down processes, disciplines and tasks into micro-tasks and sub-disciplines, allowing efficiency gains to be accumulated.
Each component of this method plays an essential role through synergy and accumulation, focusing on the systematic optimization of the smallest step of a process and the deep ultra-segmentation of each component of the business disciplines.
The “E” for “Exponential” highlights the importance of increasing performance exponentially through the accumulation effects of optimizations. For each step in a task process, organizations can achieve significant cumulative time savings. This dynamic promotes rapid and sustainable growth, essential in a competitive business environment.
The “S” for “Segmentation” refers to the systematic decomposition of projects, processes, departments, jobs, tasks and contextualized subtasks, aimed at eliminating the conflicts of intentions and objectives inherent in the generalization induced by GPTs. This approach allows for a better distribution of resources, greater acuity of execution and more efficient management of projects.
Finally, the “P” for “Process” refers to the establishment of standardized procedures to ensure consistent execution of processes, aligned with frameworks of intentions and objectives, allowing for a better understanding of incremental effects and increased control over them, thus admitting the full exploitation of cumulative effects.
The impact of the ESP method on operational efficiency is significant, allowing the AI efficiency rate to be optimized by up to 70%.
Which not only reduces execution time, but also errors, while increasing the quality of results and user satisfaction, each of these aspects having an influence on performance and turnover.
Principles in action
The ESP method allows to extract the driving substance of LLMs through four principles and approaches that act in synergy: anthropomorphism, decomposition, depth and subjectivity.
Anthropomorphism
We humans do not intuitively perceive the finesse of what a job includes in terms of disciplines, tasks and associated subtasks. We do know, however, what types of general needs a job (or service) meets.
Thus, we (humans) must base our professional approach to AI through the prism of the profession, because it is an intuitive benchmark that eliminates a large part of the questions concerning what can be asked of one profession rather than another.
But we must not forget that with AI, the methods of implementing a profession are defined by ourselves, which implies that we are both actors and spectators of the result.
As an actor manufacturing the relevance of a result, we must define with extreme precision what the AI in charge of a mission must receive as information. This is so that it can access our requests as a spectator, considering that this perspective of relevance is variable depending on the circumstances of solicitation.
This is where an anthropomorphic approach makes sense, because, in truth, as actors in the outcome, we are incapable of the level of precision required to satisfy the demands of our role as spectators.
On the other hand, as a spectator of the generated result, we can measure the relevance and therefore adapt our role as an actor.
Indeed, as recipients of an AI-generated response of which we are also the architects, we can both arbitrate the adequacy of our expectations in the face of the implications that we assume of a profession, and orient this adequacy so that the generalization is directed in the direction that we give it.
Thus, business anthropomorphism is above all a benchmark, for us more than for AI, allowing us to play different roles relating to business interactions, in the construction of the expected relevance.
Business anthropomorphism thus induces a pre-orientation of intentions and general objectives, essential for the contextualization of the variable requests of the applicant, while maintaining an adequacy between its objectives and those of the requested anthropomorphic AI.
Business anthropomorphism therefore constitutes an essential first step in the ESP method, allowing us to align our expectations with results whose relevance is anticipated by intuitive identification with business implications, without necessarily knowing the ramifications.
However, business anthropomorphism is not sufficient in itself, because it does not indicate in any way specific processing methods inherent to requests, although they are business-oriented.
The principle of decomposition
The embodiment of a profession by AI induces underlying behaviors that appear coherent, but the results produced by simple anthropomorphism remain too general and therefore unusable.
Another key to relevance then lies in extracting the deep substance of the professions, namely, the disciplines, skills and associated knowledge, thus making it possible to integrate the precision required for the implementation of an anthropomorphized AI.
To achieve this level of precision, it is first necessary to clarify the differences between the profession and the disciplines that it incorporates, while considering the AI's own benchmarks and its own interpretation criteria, in order to prevent it from weighing up the specificities that we think we introduce through precision.
The profession according to AI
A profession involves the variable application of multiple skills, each echoing several spectrums of knowledge whose importance is also variable. (Skills relating to the field of the applicable, knowledge relating to the dynamics of interpretation concerning the skills to be put into practice, in direct relation to the context of execution).
Discipline according to AI
A discipline, on the other hand, is intrinsically constituted by specific skills, each of which also echoes specific knowledge, but put into interactions within a reduced framework (the exclusive universe of a discipline). This prevents the generalist weighting induced by the sum of the business components, because all the details extracted from the nature of a discipline are consubstantial with it.
These notions and their differences are crucial, because without limitations of orientation and width of interpretation spectrum having a behavioral impact, AI systematically generalizes and arbitrates the characteristics of relevance for us.
Indeed, in all circumstances, it interprets our intentions, and assumes our expectations within a general framework, which, here, is that of a profession, itself composed of disciplines, and therefore of multiple underlying skills and spectrums of knowledge whose characteristics at this contextual scale are already a weighted average which does not allow depth.
Under these conditions, the result cannot be relevant, because its generation involves countless approximations, each of which is smoothed to obtain the statistically most probable mean.
Which can only lead to generality and the absence of specificity inherent in the expected relevance.
This results in many reworks, and often more time-consuming adjustments than doing it yourself, without AI.
Thus, the decomposition of professions into disciplines is of capital importance, because it restricts the possibilities of interpretation while narrowing the focus of attention of the AI, thereby increasing the relevance produced.
This decomposition is a prerequisite for implementing anthropomorphic business AI, but it is not enough, because this decomposition only presents a first level of depth: a discipline of a profession.
The principle of depth
For the relevance of the results to be optimal and systematic, each extracted discipline must in turn be segmented, this time in terms of skills and knowledge, leading to the extraction of tasks with single objectives.
This further subdivision in turn narrows the execution frames, thereby reducing the possibilities of misinterpreting intentions and objectives, while further narrowing the focus of attention of the AI.
A single-objective task no longer involving conflicts of interpretation or objective can then be divided in turn into hyper-specialized micro-tasks, resulting in a notion of optimal relevance, characterized by a focus of attention entirely dedicated to each micro-task individually.
Thus, each task systematically broken down into several micro tasks, receives a maximum level of attention, enhancing the relevance of interpretation of the AI by depth effect and the perceived relevance by focalization effect. Each of these effects amplifying each other.
This also enhances the ability to manufacture relevance when we have the role of actor of this relevance, because the deep decomposition reveals to us the smallest components of these micro-spots. Components that AI considers to be landmarks, while these are imperceptible to us.
As an actor of relevance
Extracting the details of the elements that make up a profession, the disciplines, skills, knowledge, tasks and related microtasks, allows us to specify our expectations and to configure anthropomorphic AIs in an optimal manner.
As a spectator
These segmented details and extractions reveal the potential expectations of the end user (who can be ourselves), allowing us to better define the framing of intentions so that the requests are perfectly aligned with the capabilities of the AI made available, contributing to effective pre-orientation on the one hand, and to the depth of interactions on the other.
Subjectivity
Finally, it is appropriate to return to the way in which a computer-assisted task is understood, without AI, in order to better perceive the graduation of depth levels and their influential adjacent factors.
Humans intuitively allocate their attention according to their priorities and objectives, using their cognitive abilities, time, skills, knowledge, and energy to establish the depth of processing as well as its level of detail according to the nature of the immediate need.
Humans constantly arbitrate the hierarchy of attention priorities, the level of concentration, the depth required, which itself depends on objectives, the latter also being dependent on interpretation factors.
This cognitive and instinctive work is constant. It cannot be reproduced by AI, because, in all cases, the judgment of relevance is sovereign to us.
Indeed, the characteristics of relevance are subjective. They are therefore subject to our judgment by essence.
In other words, any result that appears to be relevant, but whose relevance criteria retained by the AI escape us, cannot present a level of reliability sufficient for its use.
Because reliability is exclusively defined by our judgment.
Hence the need to carefully extract the characteristics of this subjective relevance in order to master all the aspects that would escape us in a non-segmented approach, which is by nature generalist.
Because it is neither more nor less than trust given to results generated in variable frameworks, involving judgments and decisions whose reliability must be based on the arbitration of the human. It is up to him to define the graduation of depth with full knowledge of the facts, ranging from the profession to the microtask.
This, without forgetting that the frames and contexts must be as restricted as possible to ensure the specificity of the depth on the one hand, and maximum attention oriented towards what we have previously defined in light of the segmentations on the other hand.
This contributes to confidence in the results produced, because, by the detailed extraction of the smallest components down to indivisible scales, we master the foundations of reliability and relevance.
This is achieved through the consecutive segmentation: Profession > disciplines > | skills, knowledge | > tasks > micro-tasks, constituting the depth.
At the same time, this fine framing makes it possible to pre-establish subjective principles of interpretation in all the dimensions involved in a task, without the result being diluted in the expected relevance in terms of specificities and depth.
The ESP method thus ensures optimal exploitation of the potential of LLMs, by transposing, ultimately, human cognitive processes into segmented procedures implemented by humans.
The latter then becomes an operator of AI ecosystems. It pilots, commands, controls and puts into action AIs specifically created and programmed to accomplish particular tasks, for particular objectives, which it then aggregates in order to contribute to a broader objective of which it remains the project manager.
It is in this simultaneous competition of principles and approaches that the relevance of the use of AI lies, increasing tenfold the speed of execution and quality, reducing error rates and having a major impact on turnover, in particular through the economies of scale achieved thanks to the excess capacities with which each human is endowed.
Implementation of the ESP method
Implementing the ESP method requires prompt engineering expertise and appropriate tools.
Prerequisites:
Have an account on neuraking.com, in order to be able to create professional anthropomorphic AIs.
To have prompt engineer capabilities valid to exploit the full potential of ESP AI on which the implementation of the method is based.
Procedures via IA ESP:
Visit the ESP AI here: https://neuraking.com/ia-assistants/esp/.
Step 1: Business Identification with ESP AI
Every profession involves sectorizations that should be identified prior to any segmentation, in order to ensure that our definition of the implications corresponds to the idea that the transformative models that will be requested have of them.
Here we will use the example “accounting".
a: Extract or verify the name of the profession inherent to the need
b: Identify the professions to be transposed into anthropomorphic AIs
The ESP AI will suggest a list of professions associated with the term entered (here: accountant).
c: Select and classify professions
At this point, the ESP AI response tells us an organizational structure that will need to be replicated, with each job being an AI to be created, like this:
For accounting:
- A general accountant = an AI
- An assistant accountant = an AI
- An internal auditor = an AI
- Etc.
And so on for each profession which must subsequently be segmented.
Please take note of this to submit each of the trades to this procedure.
Step 2: Extracting disciplines with an ESP AI teammate
a: Extract disciplines with Marco C.3.02.11.24
From one of the professions provided by the ESP AI, we will proceed to extract the disciplines with the AI “Marco C.3.02.11.24”, ESP's first teammate.
For the demonstration, we will use “tax expert“, specifying “in an accounting firm”. This clarification is necessary, because we started from the term “accountant”, but this angle is not known to Marco C.3.02.11.24.
b: Identify the disciplines to be transposed into anthropomorphic team departments
The IA Marco C.3.02.11.24 will propose a list of disciplines associated with the function/job, with initial explicit notions on the possibilities of IA LLM intervention (here: tax specialist in an accounting firm).
Here, the usage possibilities mentioned refer to the skills and knowledge that will be necessary to exploit the associated AIs, but this does not come directly into play at this stage of the extraction process.
At this point, the response from AI Marco C.3.02.11.24 tells us first of all the organizational structure of the teams' departments that should ideally support the AI Tax Specialist, like this:
Main AI:
- Tax AI
AI Team Departments:
- Tax analysis and forecasting
- Specialized crew members (unknown at this stage)
- Regulatory conformity
- Specialized crew members (unknown at this stage)
- Tax planning
- Specialized crew members (unknown at this stage)
- Fraud detection
- Specialized crew members (unknown at this stage)
- Data entry and report generation
- Specialized crew members (unknown at this stage)
- Tax consultation
- Specialized crew members (unknown at this stage)
- Documentation management
- Specialized crew members (unknown at this stage)
Transposing itself at the end of the procedure in this way:
c: Select and classify departments
Record the listed departments in order to supply them with crew members later.
At this stage we cannot yet define these, but it is necessary to establish the organization that will integrate them.
Step 3: Extract action verbs limited to LLM capabilities
Echoing skills and knowledge, action verbs and implicit actions are the elements that will characterize the orientations of objectives and tasks, without the establishment of knowledge and skills being necessary.
To this end, Marco C.3.02.11.24's response provides guidance on the possibilities of AI assistance, making it possible both to limit the approach to the scope of generative AI capabilities and to guide the selection choices in this procedure.
To extract action verbs and non-verbal implicit action elements, select one of Marco C.3.02.11.24's answers to pass to Laurent F.1.01.24, for example: “Tax analysis and forecasting: AI can help analyze large amounts of tax data to identify trends and predict tax outcomes.”
These elements are also important for specifying the task segmentation guidelines contributing to the proper conduct of this procedure.
Step 4: Extract tasks associated with a discipline
a: Extract tasks
From one of the disciplines provided by the AI Marco C.3.02.11.24 and the elements provided by Laurent F.1.01.24, we will proceed to extract the tasks associated with the third ESP teammate: Sienna B.2.02.11.24.
For the demonstration, we will use “Tax analysis and forecasting“, specifying the action verbs and implicit actions provided by Laurent F.1.01.24.
At this stage, the depth of the tasks begins to appear with a chronological notion of processing and underlying actions (similar to skills and knowledge to be introduced later as a parameter).
b: Identify and select
Identify the tasks that best fit your AI assistance expectations, then make note of them for later decomposition of each of these tasks into processing steps.
Step 5: Breaking down tasks into steps
Starting from one of the tasks provided by the AI Sienna B.2.02.11.24, we will proceed to the decomposition into processing steps with Hugo D.1.02.11.24.
This AI will not only indicate what the human will have to do, but will also establish the associated microtasks that can be segmented with other team members, business AI or departments.
For the demonstration, we will use “Gather relevant tax data from previous periods".
This example is deliberately chosen for its counter-intuitive aspect, because the term “gather” leads us to believe that this task (gathering) is the responsibility of humans, and therefore that it has no place.
However, it is crucial to delve deeper into this basis in order to extract the fine substance of the considerations of transformative models.
Hugo D.1.02.11.24's response will help us understand the essence of each task by shedding light on how to proceed both in terms of procedural relevance and for optimal decomposition into appropriate AIs and team members.
At this point, the response from AI Hugo D.1.02.11.24 tells us the details of the organizational structure of the team members who should ideally be integrated into the departments associated with the AI Tax Specialist, as previously established by the response from Marco C.3.02.11.24.
Like this :
Main AI:
- Tax AI
AI Team Departments:
The attributions are now known, producing a segmentation of team members in the “Tax Analysis and Forecasting” department like this:
- Tax analysis and forecasting (Assimilated department)
Team members:- Define the periods to be covered
- Identify data sources
- Collect the necessary documents
- Organize data
- Extract relevant information
- Check the accuracy of the data
- Compile the data
- Analyse the data
- Keep data
- Prepare a report or summary
Transposing itself at the end of the procedure in this way:
And so on for the other departments determined by Marco C.3.02.11.24's response:
- Business department: Regulatory conformity
- Teammates
- Business department: Tax planning
- Teammates
- Business department: Fraud detection
- Teammates
- Business department: Data entry and report generation
- Teammates
- Business department: Tax consultation
- Teammates
- Business department: Documentation management
- Teammates
Step 6: Extracting utility implications and dependencies
At this point, the steps identified and listed by Hugo D.1.02.11.24 may lack meaning to us and, again, seem counter-intuitive to us.
It is then appropriate to restrict their interpretation to the limits of the LLM while extracting the action elements again, but with Émilie C.1.01.24 this time.
Continuing the example, we will use “Define the periods to be covered: Identify the specific years or quarters for which tax data needs to be gathered.".
Emilie C.1.01.24 will identify the related actions and their conditions, which will make it possible to identify the usefulness of the team member, and on which data the latter will have to rely.
At this point, it is possible to create a first anthropomorphic AI, because we know the job and its ramifications.
In fact, we extracted a particular discipline from it, then action elements allowing us to extract tasks.
These have been divided into stages, the usefulness and conditions of which have been analyzed, thus making it possible to provide information on some of the objectives induced by the profession.
Induced objectives:
This will allow the teams self-constructing anthropomorphic AI profiles to take over in order to optimize each of the parameters of each of these AIs that will need to be built for each department of a profession.
Step 6a: Extraction of key elements
Finally, without this being obligatory, it is possible to extract the key elements of each step, which will help to better guide the configuration, and will make the task easier for the self-construction team members.
To extract the key elements, use Nathan H.2.02.11.24. This AI allows you to have an overview of the essential factors in the processing of tasks related to the profession, but also related to disciplines, tasks and micro-tasks.
Useful for checking the adequacy of the organizational elements defined by ESP with your perception.
Useful for novice engineers who are unfamiliar with the business at hand, and who need an overview of the utility levers.
Conclusion
The ESP method is an essential process, prior to the design of relevant anthropomorphic artificial intelligence ecosystems through their fine segmentation and their attribution of unique objectives within professions and departments.