AI Fundamentals with Mihail Eric, Head of AI Monaco

Main ideas & lessons

• AI progress is not “new”: Modern AI—especially LLMs/agents—is the result of multiple waves over 10–15+ years, culminating in today’s capabilities.

Key historical inflection points

• Deep learning wave (approx. 2012–2016): Neural networks “reborn” due to improved training methods, GPU compute, and large datasets (notably ImageNet) enabling major gains in vision, speech, and translation.
• Transformer wave (around 2017): The Transformer architecture (from “Attention Is All You Need”) enabled modern LLM backbones via next-token prediction trained at scale on text.
• LLM product wave (around 2022): ChatGPT popularized large language models through an effective product interface and post-training + RLHF (reinforcement learning from human feedback).
• Reasoning / agent wave (around 2024): “Reasoning models” (e.g., OpenAI o1) enabled more complex multi-step behavior and helped form the foundation for agents, aided by ideas like test-time compute and verifiable rewards / test-time RL.

Modern AI “model recipe” (from neural nets to useful agents)

Systems typically follow this progression:

• Pre-training → next-token prediction on web-scale corpora.
• Post-training (supervised fine-tuning) → learn task-aligned input/output behavior from examples.
• Alignment → incorporate human preference data to match what humans want.
• RL / reasoning RL with agentic environments → train models to perform sequences of actions and improve via rewards from controlled environments.

Agent architecture is more than the LLM

• An agent = LLM + instructions + tools + an execution loop/harness.
• The harness orchestrates tool use, skills, and context assembly to achieve the user’s goal.
• Skills are packaged procedures (often prompt-based), such as a SKILL.md file defining behavior and required tool usage.

Business value patterns

• Agentic AI value is often highest when paired with domain workflows—i.e., “vertical” applications (sales, coding, legal, customer support).
• Repeated playbook:
  1. capture workflows,
  2. codify them into an agent/harness,
  3. connect them to the model(s),
  4. deliver measurable outcomes.

New business metrics & pricing models

• Tokens become a central metric for cost/complexity/speed (“token maxing”).
• Pricing evolves from:
  • seat-based SaaS
  • → usage/token-based
  • → outcome-based pricing (charge for completed results).

Other AI modalities beyond text

• Visual generation (diffusion + transformers)
• Audio generation (transformers conditioned on rich speech inputs; output decoded to waveforms)
• Robotics (vision-language-action models; data + cost bottlenecks, with early progress)

Infrastructure is a major bottleneck and opportunity

• “Picks and shovels” layers include:
  • data providers
  • RL environments/data
  • neoclouds (AI-optimized compute)
  • inference providers for low-latency serving
• The discussion includes open-weight vs closed-weight models, and why companies invest in open-weight (customization, speed/cost potential, avoiding monopolies), despite a current performance gap.

Open problems

• Model vs harness: where the real gains come from (model improvements vs workflow codification).
• Memory / continual learning: how agents learn from past interactions correctly.
• Security, misinformation, accountability for agentic systems (including cross-system “tunneling” risks).
• Evals/verifications: needing better benchmarks and RL environments to know when agents succeed.
• Energy/data-center scale-up: training/serving demands could require massive electricity growth and trillions in infrastructure spending.

GTM and entrepreneurship framing

• Emphasizes an “age of the builder”: non-technical/non-CS people can still build value via domain expertise and workflow-specific systems.
• Mentions forward-deployed engineering (FDE): embedded teams close to customers to learn workflows and improve products.

---

Methodologies / instructions (detailed bullets)

1) Modern AI / agent “model recipe” (training pipeline)

Step 1: Pre-training

• Input: web-scale text corpora (e.g., Wikipedia, books, “Common Crawl”)
• Objective: next-token prediction
  • Learn to predict the next token from prior tokens.

Step 2: Post-training (Supervised Fine-Tuning, SFT)

• Input/Output: example pairs for the target task (prompt → desired completion/output)
• Goal: mimic correct behavior (e.g., produce an investment memo style output)

Step 3: Post-training (Alignment)

• Add “preference” signals so outputs match human goals.
• Mechanism: human preference comparisons (conceptually similar to RLHF).
• Result: outputs become more appropriate to user intent and audience.

Step 4: RL / Reasoning RL with agentic environments

• Create agentic environments where an agent can:
  • take actions,
  • observe outcomes,
  • earn rewards defined by the system designer.
• Typical environment behaviors:
  • long-horizon tasks (multiple steps),
  • tool use (e.g., web research / databases / document actions),
  • iterative refinement toward task completion.

2) Agent design “building blocks”

• User goal: prompt/instruction to start the task.
• Harness (or orchestration layer) that manages execution:
  • tool calls (e.g., search, code execution, plotting, integrations),
  • system prompt (behavior directives, safety/style constraints, model instructions),
  • skills (reusable procedures/prompted behaviors packaged for invocation),
  • context engineering (assemble the right combination of user instructions, skills, and prior context),
  • model execution loop (LLM “intelligence layer”).

Tools + skills + harness coordination

• The harness decides:
  • when to call a skill,
  • what tools to use,
  • how to integrate results into the next model step.

Skill example pattern (as described)

• Skill stored as a text file (e.g., SKILL.md)
• Contains:
  • a description of behavior/persona,
  • instructions for how to perform the procedure,
  • tool expectations for that procedure.

3) Vertical agent application “business playbook”

• Step A: Capture domain workflows

  • Observe or collect real usage patterns in a domain (sales, legal, support, coding).
  • Build a platform even before intelligence is perfect—focus on workflow capture.

• Step B: Codify workflows into an agent harness

  • Convert observed workflows into:
    • tools,
    • skills,
    • prompts/system instructions,
    • orchestration logic.

• Step C: Attach the intelligence layer (LMs)

  • Use LLM(s) to drive decisions, drafting, analysis, and multi-step work.

• Step D: Deliver business outcomes

  • Aim for measurable impact such as:
    • ticket resolution,
    • revenue workflows,
    • legal task completion,
    • coding task throughput.

• Step E (implied): Iterate toward better ROI

  • Improve harness/environment/data so agents reliably complete tasks end-to-end.

---

Speakers / sources featured (as mentioned)

Featured people

• Mihail Eric — Head of AI Monaco (main lecturer/speaker)
• Joyce — audience member (MBA2)
• Nir — audience member
• Esther — audience member
• Ahmed — audience member
• Host / Moderator — named only as [MODERATOR] (exact name not provided)

Referenced institutions / authors / organizations

• Fei-Fei Li — creator of the ImageNet dataset (Stanford professor)
• Google researchers (2017) — authors of “Attention Is All You Need”
• OpenAI — referenced for ChatGPT and o1 reasoning model
• Anthropic — referenced for models (e.g., “Claude” family) and “Reasoning models” competition
• NVIDIA — referenced in the context of GPU history/compute enabling deep learning
• Hugging Face — referenced for open-weight model ecosystem and a security anecdote