What I Actually Think About the AI Future

Everyone wants a single sentence about AI’s future. “AGI in two years.” “It’s a bubble.” “It changes everything.” These are personality statements dressed as forecasts.

I build ML products for enterprise customers and research verification problems on the side. Here is my actual model of the AI future, with timelines I would bet modest money on and caveats where I will not.

Capability Timeline

gantt
    title AI Capability Timeline (My Estimates)
    dateFormat YYYY
    axisFormat %Y

    section Commoditized
    Text generation quality plateau           :done, 2023, 2025
    Code assist mainstream dev                :done, 2024, 2026
    Cheap open-weight inference               :done, 2025, 2027

    section Improving Fast
    Agentic workflows reliable sub-domains    :active, 2024, 2028
    Multimodal doc understanding              :2024, 2027
    Real-time voice agents commercial         :2025, 2028

    section Hard Problems
    Reliable long-horizon autonomy            :2026, 2032
    Forensic-grade vision                     :2024, 2030
    Robotics general purpose                  :2027, 2035

    section Overhyped Near Term
    Full job replacement knowledge work       :2025, 2026
    AGI                                       :2027, 2040
    Regulatory clarity global                 :2026, 2030

Commoditized means good enough for most use cases at low cost. Hard problems means I would not bet my company on solving them in three years.

Hype vs Reality

Hype: AI replaces software engineers in five years.
Reality: AI changes what engineers do. Writing boilerplate is commoditized. System design, debugging production at 3 AM, and knowing which corners not to cut remain human. Team sizes may shrink at low-performing orgs. Output per engineer rises elsewhere.

Hype: Every startup needs an AI strategy.
Reality: Every startup needs a customer strategy. AI is implementation detail unless AI is the product. Most products need a thin AI layer or none.

Hype: Foundation model companies capture all value.
Reality: Value accrues to distribution, data flywheels, regulatory moats, and workflow lock-in. Open weights commoditized the middle. Wrappers died. Vertical depth survives.

Hype: Scaling laws solve everything.
Reality: Scaling helps until it hits data walls, eval walls, and economic walls. DeepSeek moment proved economics matter as much as parameter counts.

Hype: AI safety pauses deployment.
Reality: Enterprise procurement and liability law pause deployment faster than OpenAI safety team. Regulated industries move slowly for boring reasons.

Moats That Survive

1. Proprietary operational data with feedback loops (labeled outcomes from real workflows, not scraped web text)
2. Regulatory certification (health, finance, compliance-heavy B2B)
3. Physical world integration (hardware, on-site trust, messy deployments)
4. Customer switching costs embedded in workflow, not chat interface
5. Brand trust in high-stakes decisions (safety, verification, money movement)

Moats that erode:

• “We call GPT-4”
• “We have a chatbot”
• “We fine-tuned on public data”
• “Our prompt is secret sauce”

Timeline Bets I Will Make

2025-2027: Agentic workflows reliable in narrow domains (customer support with human escalation, code migration with test suites, document processing with validation gates). Not reliable for open-ended “run my company.”

2025-2028: Multimodal models good enough for document QA and triage. Not good enough for forensic evidence without specialized CV underneath.

2026-2030: Significant job restructuring in content mills, basic legal doc review, tier-1 support. Not mass unemployment. Labor market friction and retraining lag technology by years.

2028+: Robotics advances if sim-to-real and hardware costs improve. I am skeptical of home humanoid hype before warehouse humanoid reliability.

AGI: I will not give you a date. Anyone who does is selling something. Transformers were a breakthrough. Breakthroughs continue. “General intelligence” is definitional quicksand.

Policy and Geopolitics

AI policy fragments by jurisdiction. EU AI Act, US executive orders, India evolving framework. Enterprise customers will require compliance documentation. Compliance becomes moat for companies that invest early.

US-China model competition continues. Enterprise buyers outside both blocs may mix providers. Founders should avoid single-provider dependency for core inference.

Personal Position

I am not an AI doomer. I am not an AI maximalist. I am a founder who has seen RAG fail, sensors lie, and plant managers ignore dashboards built by people who never visited a plant.

AI is the most important engineering shift since mobile. It is also the most oversold technology since blockchain met enterprise ERP.

Build for the world where inference is cheap, models are interchangeable, and customers still pay for outcomes. That world arrived faster than I expected in February 2025. The founders who win are the ones who never confused model capability with product value.

That is what I actually think. Ask me again in a year. I expect to revise.