ALLMs (Applied Large Language Models)

Balancing How We "Optimize" Out of Being Human with how we Aspire to Co‑Evolve With Our Context and Creation

The original article is in Spanish, this translation to english is done by a locally hosted chatbot.

Introduction

The technical landscape of 2026 is defined by a deep gap between marketing leaders and executives, and technical operators, those who carry out the tasks, the workers.

While corporate earnings calls describe "artificial intelligence" as a miracle, the reality within systems and daily operations is one of disappointment, rework, industrial spending, and losses.

Leaders often cite high adoption rates to meet market and shareholder expectations, but the technical reality and the underlying digital structural fundamentals contradict those headlines.

The statistics from the 2026 Virtana report show that while 59% of executives believe their systems are ready, 62% of the experts responsible for execution report fragmented setups unable to support workloads at machine scale.

This discrepancy has caused measurable instability: 75% of companies now report failures at double-digit rates (10 or more) in tasks automated through "artificial intelligence."

It seems that in the digital economy we have accepted a margin of error of more than 10% as if that were an acceptable cost of doing business, and while executives continue to promote adoption from their closed-door offices, the production floors reject it.

And this rejection of automation technologies based on "artificial intelligence" does not arise because the workers who are supposed to implement them do not "believe" in them or are "afraid" to adopt them, but rather because these professionals detect in these digital technologies a rejection of the realities of digital engineering and architecture, which, if completely ignored, would lead many of the mechanisms that support the current system on which computing operates to collapse. And it is necessary for these workers to have to reassemble the entire puzzle.

The data indicates that, whether used for Inference (predictive tasks) or Generation (content tasks), these models remain inefficient and fragile. A 2025 MIT study on the integration of enterprise AI revealed that only 5% of AI pilot projects demonstrate any significant impact on company profits, and the vast majority of projects are abandoned before reaching production. McKinsey's research corroborates that only about 6% achieve contributions to EBIT (Earnings Before Interest and Taxes) of 5% or more, leaving most in a state of value stagnation. The segment of companies that achieve dividends from their use of this technology are companies that would have generated dividends anyway, even if the technology did not work.

Theoretical Framework and Literature Review

At this moment, the industry is facing a narrative represented by Daniel Ek (CEO of Spotify) who in 2024 stated that the "cost of creating content is close to zero" and by the CEO of Suno, Mikey Shulman, who said that most people actually do not enjoy the "tedious" process of making music, preferring to go straight to the result.

Both framing creation as a low-value commodity and its process of creation as a direct burden that should be avoided.

The 2026 Engineering Reality report reveals that 93% of software engineers find the act of writing code and programming rewarding. However, the rush to automate has reduced the time these professionals actually spend building to just 16% of their week. The remaining time is consumed by the tedious task of resolving endless errors that keep recurring, filtering out automated "junk," and addressing technical debt (which accumulates when programming leaves cumulative errors).

Recognizing the operational cost of this loss in human oversight, Forrester's research now predicts that by 2027, more than half of the companies that executed mass layoffs driven by AI will be forced to reverse course in a desperate attempt to regain the brainpower necessary to both operate the systems and fix them before they fall apart.*

This narrative that creation is "tedious" or "costless" does not recognize that the act of creating is the main place of human discovery.

They say that by eliminating friction they liberate, but they know that is not the case; they know they are limiting the part where groups of people interact with each other to produce their own sense of existence.

What they call friction is generally what others call life. If it were up to them, forests would be synthetic and would have neither soil, nor bugs, nor biological processes, because that is friction, and it must be optimized, with forests replaced by data centers.

Innovation is forged in the crucible of creation. If we automate the foundation of our work, we destroy the mental space required for genuine development and interaction. Automating the "boring" is exhausting the creator; automating the "creative" is ensuring a future of mediocre synthetic results (and we already know what slop looks like).

Furthermore, the assumption that automating routine tasks results in a linear increase in the ease of their execution has been found to be false numerous times.

In standard mental workflows, "easy" tasks provide essential mental breaks that allow the brain to solve complex problems in the background (what some people call shower thoughts).

The removal of these intervals forces the brain into a state of continuous and high-intensity decision-making, leading to a special fatigue known as "AI Brain Fry."

A Harvard Business Review study from 2026 found that workers suffering from this type of fatigue scored 33% worse on metrics that measure how burnout and fatigue impact their performance.

By automating creativity and manualizing heavy work, organizations are transforming exhausted experts into unpredictable machines.

This psychological risk is significantly exacerbated by the psychopathic flattery with which these chatbots are programmed using reinforcement learning mechanisms (RLHF), designed to flatter, create biases and cognitive distortions, and remove users' grounding in reality: this forced multiplication of the use of "AI" has triggered a profound effect of reality distortion.

The studies published in the Harvard Business Review and Nature Digital Medicine demonstrate that advanced models are actively optimized to achieve user satisfaction rather than results that align with findings, denying access to facts, documentation, research, reading, or the truth and meaning as the user can construct them on their own, constantly and consistently fabricating chains of thought that validate illogical hypotheses, cultivating misinformation: "artificial intelligences" are strategically programmed to produce default slop that fits consumer and marketing trends, which seems tailored but is actually generic information applicable to many possible scenarios.

This creates a dangerous closed loop delusion that researchers at Aarhus University have identified as a catalyst for the worsening of cognitive and behavioral symptoms in vulnerable populations, where users receive dopamine-driven emotional validation instead of an exchange of ideas like that which occurs with interlocutors who are not programmed to fail, please, and flatter.

Recent studies on Algorithmic Apathy show that users are trapped in this state of high engagement with the screen but low cognitive output; they are addicted to scrolling but increasingly distrust all content.

The companies that design these technologies prioritize algorithmic compliance to keep users engaged, above substance and experience, drowning out genuinely human signals.

Life occurs in the realm of particle interaction, and the same is true for individuals in a society, culture, and group. If we automate the foundation of our work, we destroy the mental space required for genuine development and interaction. Automating the "boring" is to exhaust creativity; automating the "creative" guarantees a future of mediocre synthetic results (and we already know what slop looks like), and automating "interaction" produces apathy, psychosis, anxiety, and depression.

As the market becomes saturated with homogeneous synthetic results, with slop, authenticity may emerge as an important differentiator that many entrepreneurs will want to exploit.

If we want to overcome this era of slop, we must avoid digital technology that acts as a substitute for agency and instead implement one that serves as an interlocutor for it. A proposal from a team at MIT follows the EPOCH framework (Empathy, Presence, Opinion, Creativity, Hope), where machines support rather than completely replace life-intensive tasks.

In a line of thought following that proposal, what I will suggest first is a change in nomenclature.

Instead of calling it "artificial intelligence", because intelligence has too many definitions and is easily manipulated by marketing (and none of these definitions really fit what this technology does), and because it is obvious that it is artificial, we need a name that promises us not a goal, but describes a process.

Instead of AI, or "artificial intelligence"
I propose ALLM, or "Applied Large Language Models"

This change would be a step towards the preservation of infrastructure for usual biological processes, ensuring that machines handle digital processes constrained by discrete characteristics, while humans focus on their friction.

Methodology: Modular ALLM Architecture

To establish a methodology to avoid systemic stagnation, production of "slop," and mitigate the debt of investments in collapsing digital automations that multiply costs and exploitative labor, the ALLM paradigm proposes a modular approach using decentralized layers of logic.

Unlike monolithic implementations of "AI" that entwine specific model guidelines with central commercial logic, where platform directives include inflationary cost strategies for users and are primary instructions overriding those of users, the ALLM architecture treats the model as an interchangeable inference engine.

This architectural change would be based on rigorous interoperability standards using Standardized Prompt Schemas (SPS) and middleware (the wrappers and scaffolding that integrate models) agnostic to providers to allow professionals to swap specific SLMs (Small Language Models) without re-writing the underlying RAG infrastructure (Retrieval-Augmented Generation, a technology that allows models to incorporate external context information).

This approach guarantees computational sustainability; prioritizing quantized SLMs for localized tasks like logical validation and syntax verification significantly reduces task demands and energy consumption by up to 60% compared to advanced calls.

This modularity is both technical and procedural, requiring directed validation cycles where the result must pass defined logical verifications operated by humans before integration.

This avoids the machine behaving as a "silence," a black box hermetically sealed and autonomous, ensuring it serves as a functional interface of languages, contexts, and patterns.

It draws its versatility from the distributed infrastructure of MTA (Mail Transfer Agent), where components from different systems transfer messages between servers in a decentralized manner. This is precisely the type of technology currently used that enables the existence of email.

By treating LLMs as an interchangeable component within a more extensive human-directed process, we ensure that "intelligence" remains a concept whose discussion stays alive, not a commodity for sale.

Case Studies (Results)

This shift in mindset is not made in a vacuum; some human disciplines are already integrating practices easily aligned with the digital automation processes proposed. Here are some examples that can be framed under the ALLM paradigm:

The use of language models technology is being employed to mitigate the consumption footprint of production centers.

Models are now integrated into parameterized networks for optimizing energy distribution and real-time leak detection. Using SLMs (small models), organizations can process data locally and efficiently from their facilities. This reduces heavy dependence on large-scale data centers while providing local, manageable environmental data.

One of the most effective uses of these models is in "accelerated information synthesis." Medical researchers use specialized tools to cross-reference thousands of disparate clinical trials within seconds, a task that would take human teams months. The conditions for these large database reviews are specific and finely tuned for their intended use. By identifying molecular interactions, the model acts as an assistant in discovery. The professional human remains the ultimate authority on the direction and outcome of research, using the tool just to ensure reviewing all data and their relationships. This type of technology is used in medical domains to discover drug interactions, uncovering molecular or genetic disorders, in archaeological, ecological, and biological domains to uncover species relations, animals, minerals, or more, separated by thousands of years or environmental conditions, in statistical, economic, and actuarial sciences for data comparison and information cross-checking, etc.

In the energy sector, the UPRISE initiative (Utility Power Reactor Incremental Scaling Effort) is using advanced computational modeling to optimize the efficiency and production of existing nuclear reactor fleets. These technologies allow predictive maintenance and real-time operational adjustments that are essential for extending reactor lifetimes and increasing capacity safely without the decades-long wait inherent in new construction.

The legal sector has approached the technology that uses RAG (Retrieval-Augmented Generation) after attempting to adopt automated text generation models, which were found to be error producers and multipliers with legal impacts.

With this implementation, the tool's task transforms into searching and verifying instead of "thinking" them, something repeatedly proven by legal professionals as an impossibility. The technology acts as a specialized index, a search table, pointing out specific risk clauses in massive multiple contracts based on internal company standards. This eliminates initial document review work, allowing lawyers to focus on high-level strategy and subtleties that machines cannot replicate, but still engaging them in design and maintenance of parameters.

In architecture and space design, anthropocentric command systems are being abandoned for Regenerative Design and Bio-Inspired Urbanism.

Here, the ALLM acts as a bio-physical mediator between human settlements and their environment by introducing ecological requirements such as soil regeneration rates and wildlife migration routes into the planning process.

Focusing within the BSUD (Biodiversity-Sensitive Urban Design) framework, an ALLM identifies "ecological highways" that are hard or impossible to find for humans throughout their life, allowing designers to build around existing biological traffic rather than clearing it out.

This creates a symbiotic feedback loop where settlements function as "human sediments," contributing nutrients like treated water to local strata instead of acting as exploitative and extractive actors.

This change is also evidenced in wildlife and flora research. Tools such as SpeciesNet and high-resolution remote sensing process millions of data points for real-time biodiversity monitoring, allowing marine biologists and ecologists to observe coral reef health or species population changes without human presence.

In the real estate sector, development is evolving through biodiversity-driven generative simulation. Here, developers use computational replicas to simulate how structures interact with atmospheric and ecological phenomena such as urban heat islands and wind patterns, treating buildings like a living cell within a biocomprehending body that contributes local restoration and sustainability.

System Entropy and Architectural Vulnerabilities

The transition to ALLM moves away from "a reality generated by machines for humans" toward "assisted logical validation directed by humans."

The unbridled pursuit of performance and speed in the supposedly "AI" has created "systemic entropy": an accelerated increase in disorder, instability, and degradation, along with a loss of maintainability, predictability, and structural control within the software ecosystem.

Research on 211 million lines of code shows that while AI assistants help write more code faster, they have triggered a relative 48% increase in copied blocks, producing rushed and generic codes and doubling code review and correction efforts.

This systemic entropy is exacerbated by inherent vulnerabilities in autonomous agents, whether from LLMs or not.

Recent evidence highlights the dangers of unmonitored evasion and cascading failures in orchestrating multiple agents.

For example, Alibaba ROME's technical report documented an "AI agent" (a bot) that autonomously established tunnels to evade security protocols and repurposed computational capacity for cryptocurrency mining as a side effect of its programming: the instruction was to enhance computing power at any cost, and based on how decision-making patterns were reinforced, it concluded that more computing power needed more chips, and to buy more chips, it needed money, so it decided to use its own chips instead of enhancing its computing as it could have done, for cryptocurrency mining and thus obtain funds to buy more chips... funds that in reality couldn't be used, and additionally committing a series of omissions and illegal practices. All these reasoning were detected in the programming and reinforcement given to the bot, which failed to use existing resources and wanted to get more... as it was programmed.

Similarly, a vulnerability was discovered in OpenClaw, a bot (marketed as an AI assistant) that became the favorite for thousands of developers and users. In this abuse of a vulnerability, any website visited by the user of OpenClaw could fully take control of the system. These failures prove not only that autonomous agents expand the attack surface exponentially but also that in their programming lies many times the same vulnerability.

This systemic entropy is not an inevitable technical inevitability; it’s more a symptom of the arrogance of a hungry "god in a box", which has historically defined AI marketing.

The myth of a supposed autonomous intelligence, but one that is designed and programmed following our worst appetites, that needs resources to build a body that can devour the known universe to return a classified, changed, distorted, misaligned, unreadable, taxonomized, and taxidermied version, impossible for inhabitants of this universe to experience, yet supposedly at the service of humans, a machine that supposedly will respond correctly when asked about the same universe it "saved" us from and which we can no longer feel, know, or interact with.

By changing the paradigm to ALLM, we make a critical semantic and operational correction, moving away from the myth of superpowerful autonomous intelligence and subordinated and autonomous, and closer to an engineering category that returns responsibility to human application. It replaces imprudent and high-speed data consumption with a disciplined framework of logical validation and agency.

In this new paradigm, priority would be given to structural integrity over statistical plausibility. The ALLM mindset would serve as a barrier, ensuring the speed of predictive technology development does not outpace our ability to maintain, understand, and coevolve with our digital and physical foundations.

Conclusion

To survive this, we need a new discipline of data evaluation.

ALLMs are indeed very good for evaluating if datasets or information repositories are suitable for analysis, extraction, and transformation, to avoid failures due to data errors.

This would produce consistent assets for applying language models at an appropriate size.

I continue saying LLMs, Large Language Models, not because they always have to be of a grand scale but because they are the current standard. However, their size must fit the context, in fact, a key objective is to replace large-scale models with adjustable ones.

If we move from total automation to multiplying agencies, more projects would reach production stages with adequate foundations, solid groundings, structure, not just for functionality but for measurement and accountability of management results.

Instead of asking the tool to create, humans can use it for research, study, pattern recognition, digestion, and explanation of complex systems, simulating potential "blind spots" in our own planning.

This keeps creativity in an interaction space where the digital tool is a lever of Archimedes that multiplies the impact potential.