AI's Decade-Long Ascent: From Deep Learning to Agents

Mihail Eric, Head of AI at Monaco, challenges the common perception that AI is a new phenomenon, stating that its development spans at least 10 to 15 years.

The progress in artificial intelligence has been marked by successive waves of technological advancement, including the emergence of deep learning.

Significant evolutionary steps, such as the introduction of Transformers and advanced reasoning models, demonstrate this long-term, incremental development rather than an abrupt arrival.

This innovation became the backbone for all modern Large Language Models (LLMs), moving away from earlier recurrent neural networks.

The Transformer architecture refined the training paradigm by employing next-token prediction, enabling models to generate coherent text by forecasting subsequent words or tokens.

This method involves training models on internet-scale datasets to learn patterns for completing text sequences effectively.

The widespread success of Large Language Models (LLMs) like ChatGPT was not solely due to model intelligence but significantly driven by accessible, chat-based interfaces.

Prior to ChatGPT's launch in 2022, models such as GPT-3.5 already existed but lacked a user-friendly way for the general public to interact with them.

Simple interface designs, exemplified by ChatGPT, dramatically lowered the barrier to entry, making powerful AI capabilities accessible to a broader audience.

This highlights that product design and the ease of user adoption are as critical as the underlying model's sophistication for AI to achieve mainstream success.

The second phase, post-training, involves supervised fine-tuning, where input-output pairs are created to refine the model's utility and align its behavior with desired outcomes.

Alignment, a critical part of post-training, leverages human preference data to ensure that model outputs are consistent with human expectations and values.

The final step introduces Reinforcement Learning (RL), specifically reasoning RL, which places models in agentic environments where they learn from rewards based on task execution.

A critical component, termed the 'harness,' orchestrates the various steps required to fulfill user objectives by managing the interplay of instructions and tools.

System prompts provide the overarching behavioral directives that guide the model, ensuring its actions are aligned with desired outcomes and constraints.

These tools allow agents to connect with applications such as Slack, Figma, or Notion, extending their capabilities beyond pure language generation.

Within the 'harness' framework, tools are categorized and managed to facilitate efficient and dynamic execution of complex tasks.

The access and configuration of these tools are customizable, ensuring that agents can be equipped with the precise functionalities required for specific user demands, with product management plugins serving as a practical example.

The initial step involves building a platform designed to capture and understand domain-specific usage and workflows.

Subsequently, these observed workflows are codified into an intelligent agent, leveraging the insights gained from the platform's initial use.

This approach combines the 'harness'—encompassing tools, skills, and prompts—with the core intelligence layer provided by Large Language Models to deliver effective solutions.

Usage-based pricing, which charges customers based on their consumption, such as the number of tokens used, is becoming a standard metric in the industry.

Outcome-based pricing is gaining traction, particularly among vertical companies, where charges are applied for specific completed tasks, such as successfully deflecting a customer support ticket.

This model aligns costs directly with tangible results, offering a new approach to valuing software services beyond mere access.

Mihail Eric indicates that the field of robotics is on an upward trend, yet it still requires foundational technological advancements to achieve widespread adoption.

The primary barriers preventing consumer-scale robotics are significant challenges related to data accessibility and the high costs associated with hardware.

Although some experimental systems, such as those developed by Google, demonstrate promising capabilities in performing household chores, the robotics industry remains in its nascent stages of development.

A growing number of companies, including Scale AI, Reka, and Surge, are generating substantial revenue by providing high-quality, task-specific datasets.

The demand for these specialized data solutions spans a wide range of applications, from solving complex math problems to processing intricate legal documents.

Consequently, creating tailored environments for Reinforcement Learning (RL) has become a top priority for developers focused on training advanced models.

Their core focus is on optimizing the serving layer to ensure low latency, cost-effectiveness, and the ability to deploy models that exceed the capabilities of consumer hardware.

These providers play a crucial role in the AI ecosystem by hosting and making accessible both closed and open-weight models, democratizing access to powerful AI capabilities.

The core intelligence layer of AI is embodied by the 'model,' while the 'harness' represents the surrounding framework of skills, tools, and prompts.

Mihail Eric argues that many vertical AI companies derive substantial value not just from the model itself, but from the harness, by codifying specific workflows.

This dynamic creates a constant strategic tension between investing in enhancing model intelligence and developing robust workflow harnesses.

He describes this as a seesaw, where companies continuously weigh the benefits of deepening the core intelligence against improving the scaffolding that enables practical application.

The estimated cost for constructing the necessary data centers to meet these demands could reach upwards of $5 trillion.

Mihail Eric compared this projected infrastructure investment to the massive railroad expansion seen in the 1800s, highlighting its unprecedented scale.

A pressing concern is that the pace of AI model development is currently outstripping the construction of data centers needed to support them.

To stay current in the rapidly changing AI ecosystem, Mihail Eric advises subscribing to a select set of high-quality newsletters from trusted experts.

Integrating these carefully chosen information sources into daily workflows is crucial for consistently staying informed about significant developments.

He also suggests focusing on core AI principles and foundational concepts rather than attempting to track every minor or incremental technological advancement.

This approach helps to cut through the information overload and concentrate on insights that truly matter for understanding the field's trajectory.

FDEs are tasked with extracting critical insights from client pain points, which they then use to either develop bespoke solutions or contribute to improvements in the core product.

This role is gaining prominence because it offers a faster and more effective method for customizing workflows through direct customer collaboration, rather than solely relying on base model learning.

The trend underscores a shift towards more hands-on, client-centric development, where engineers work closely with users to address specific needs and drive product evolution.

Current AI services are heavily subsidized, a business strategy akin to those used by category-defining companies like Uber or Lyft to rapidly gain market share.

However, this subsidization model is expected to face a reckoning under public market scrutiny, which will likely result in increased costs for consumers.

Mihail Eric notes that model providers inherently lack 'stickiness'—the ability to retain customers easily due to integrated workflows.

To overcome this, model providers may pursue acquisitions of verticalized players, such as Harvey, to build deeper, more integrated workflow solutions and enhance customer loyalty.