How do we 10x GPT-5?
How do we 10x GPT-5?
On the long road ahead to having true digital coworkers
Welcome to Fully Distributed, a newsletter about AI and its impact on business. Join our growing community by subscribing here:
We've heard Sam Altman say that GPT-5 will be to GPT-4 what GPT-4 was to GPT-3. But what does a 10x improvement from here actually look like? What about another 10x for GPT-6? In recent conversations, I've noticed people struggling to envisage such a leap - after all, our current models are already pretty damn good. How much better can they possibly get?
My answer: a LOT better. We are only getting started.
Let me explain.
The Hierarchy of Work: From Tasks to Jobs
To understand where AI is going, we need to zoom out and look at the bigger picture of how work is structured. Let's break it down:
Tasks: Single, discrete actions
Marketing: Writing a cold email copy
Finance: Analyzing a company's quarterly earnings report
Medicine: Interpreting a patient's X-ray
Workflows: Series of related tasks
Marketing: Planning, creating, launching, and monitoring an entire email outbound campaign
Finance: Conducting a comprehensive company valuation, including financial modeling and competitor analysis
Medicine: Managing a patient's treatment plan from diagnosis to recovery, including tests, medication, and follow-ups
Jobs: Series of workflows
Marketing: A marketing manager responsible for multiple campaigns across various channels
Finance: An investment analyst covering a specific industry sector
Medicine: A primary care physician managing multiple patients with various conditions
This (imperfect) hierarchy illustrates how work complexity increases dramatically as we move from tasks to jobs, presenting unique challenges for AI automation.
The Compounding Problem
As we move from tasks to workflows to jobs, AI faces a compounding accuracy challenge. Each step in a process introduces a chance for error, and these errors multiply as the number of tasks increases.
If an AI is 90% accurate in understanding a task (i.e., the user's "instructions") and 90% accurate in executing it (final output matches what the user intended), the overall accuracy for that single task is 81% (0.9 * 0.9). While this may seem acceptable for a single task, it becomes problematic when applied to complex workflows or entire jobs.
This compounding effect explains why AI agents have seen limited adoption in enterprises so far. Current AI agents lack the reliability needed for complex real-world processes. Yet, this also means that even minor improvements in AI accuracy could yield outsized impacts, especially in multi-stage workflows.
At OffDeal, we've faced these challenges firsthand when building AI agents for M&A strategic fit analysis. Despite using advanced models, we still rely on meticulously crafted prompts (which take many hours to develop A/B test) and strictly formatted data inputs. As AI model capabilities evolve, we anticipate the need for such clunky workarounds to gradually diminish.
The Path to 10x Improvement: From Tasks to Jobs
We're in the early stages of AI automation, currently focused on individual task optimization. The future, however, promises far more:
Near Future - Workflow Automation: Complex workflows will be increasingly automated, introducing true "digital coworkers" capable of handling significant workloads.
Medium Term - Entry-Level Job Automation: AI agents will begin autonomously handling entire roles, likely starting with junior or outsourced positions.
Long Term - Advanced Job Automation: AI will progress to managerial-level decision-making, orchestrating intricate, interconnected workflows that demand high-level strategic thinking.
Each step represents a massive leap in AI capabilities, far beyond what most people envision when thinking about "10x improvement."
It's important to note that smarter models alone won't suffice. Substantial infrastructure development is necessary for agents to transition from data processing to action-taking. Additionally, hardware advancements are crucial, as exponential growth in token generation demand will require significantly expanded compute capacity. These elements are expected to evolve concurrently, potentially at comparable rates.
Conclusion
As we stand here in 2024 at the base of the AI S-curve, we're ought to ponder
How will humans work alongside increasingly capable machines?
Will there be resistance to this change, and how might it manifest?
Is it possible to "future-proof" yourself in a rapidly evolving job market?
How will regulation and politics shape the adoption and development of AI?
If you're interested in learning more about how OffDeal employs AI agents to transform SMB M&A, or simply want to chat about AI Agents x Finance don't hesitate to DM.
As always, pushback and feedback always welcome. Twitter DMs always open @leveredvlad
If you enjoyed reading this, subscribe to my newsletter! I regularly write essays about AI and business.