AI investor + engineer discuss the current state of AI

Overview

The discussion is a wide-ranging update on the AI ecosystem, with a heavy focus on:

• AI coding agents
• AI infrastructure stability
• How these trends affect startups, markets, and product strategy

---

1) “Coding wars” are huge—and still in an exploratory phase

• OpenAI and Anthropic are described as making coding a top priority, with Cursor also cited as a major player.
• The coding market is portrayed as having grown explosively in the last ~year, with multiple players reaching multi-billion ARR-level scale (figures cited include):
  • Anthropic: ~“2.5B from Cloud Code”
  • OpenAI: ~“2B”
  • Cursor: “rumored ~2B”
• Despite big growth, the hosts argue the market is still in “capability exploration”, not an efficiency/optimization plateau—meaning:
  • Experimentation and aggressive iteration (including higher token usage) can be rewarded.

---

2) Agent trend: 2025 was coding agents; 2026 is “agents breaking containment”

• Core thesis: coding agents will expand beyond writing code to doing broader tasks—“breaking containment to do everything else.”
• Belief: software creation becomes the mechanism that “eats the world” because agents generate software, and software permeates other industries.

---

3) Infrastructure is stabilizing, but specific agent components will keep changing

• A key debate: has AI infrastructure reached stability? (Referenced via Harrison Chase / LangChain viewpoints.)
• Counterpoint: stability may be emerging at the “minimal viable” layer, such as a standardized “skills” format for tool use.

• However, agent behavior around:

  • real-time execution
  • sub-agents
  • memory
  • and related disciplines will continue evolving.

• Episode contrast:

  • Applications: can discard code faster because end users validate and drive iteration.
  • Infrastructure: faces higher switching costs and developer churn risk, making frequent reinvention harder defensively.

---

4) Selling to agents: treat “agent experience” like “developer experience”

• Infra companies increasingly target agents as the primary customers, not humans.
• Agents are described as:
  • prompt-injectable
  • systems that heavily exploit what’s already installed and winning
• Practical guidance:
  • If an interface isn’t available via an API agents can call, it won’t “exist” for agent workflows.
  • Build like you would for developers (docs, stable APIs, statelessness, progressive disclosure/search), with extra emphasis on:
    • CLI
    • automation surfaces

---

5) Models vs startups: foundation models may disrupt some categories, but opportunity remains

• Investor perspective: mid-size infrastructure startups face consolidation risk as foundation models and standard tooling mature.
• However, very early/micro startups are viewed as having limited “being eaten” risk because outcomes may include:
  • acquisition
  • becoming talent pipelines
• Biggest pressure is suggested to be on traditional low-NPS SaaS, where AI reduces the need for large parts of conventional software workflows.

---

6) Training your own models, RL, and chips: a nuanced “agent lab playbook”

“Agent lab playbook” framing

• Start with large foundation models.
• Specialize for the domain.
• Once you have enough user workload and high-quality data, train your own models to improve:
  • cost
  • latency
  • and potentially differentiation

What’s viewed as clearly valuable

• Domain-specific models, especially for search-like tasks.

What’s less clear

• DIY RL:
  • may improve quality
  • but uncertainty remains around whether it’s efficient versus alternatives

Chips and inference speed

• Custom/alternative chips (e.g., Cerebras) are discussed as increasingly important because inference speed improvements can unlock new application patterns.
• Investment cycle is considered multi-year and hard to predict.

---

7) Open models: sentiment is shifting toward more open rather than less

• A “changed mind” segment argues openness has improved in practice:
  • even if capability gaps remain, open models show meaningful adoption (e.g., open-router usage, accounting for discounting)
• The host distinguishes the top cohort from the broader market:
  • the “top 20%” of agent/model builders behave differently than average wrapper/startup users.
• Open models are also linked to:
  • better economics (speed/cost)
  • scaling-workload dynamics that make custom fine-tuning / post-training more viable

---

8) Market structure outlook: likely “two big players + long tail,” unless major shocks occur

• Likely end state for coding:
  • two dominant labs
  • plus a long tail of niche players for use cases the big two don’t prioritize

• Structural change would require major shifts in:

  • economics
  • brand-building
  • distribution (Examples raised include Microsoft expanding beyond Copilot via GitHub.)

• Enterprises are still believed to want dedicated partners (last-mile implementation and orchestration), not just raw model access.

---

9) Strategic “AI coding psychosis” and dark factories

• The episode introduces an SDLC inversion concept (“dark factories” / DM-Simon Willison reference):
  • move toward zero human written code
  • and even zero human review in some pipelines
• Implication: massively higher software throughput via automated testing/verification and process changes.
• The argument: quantity can be leveraged not just for slop, but to accelerate:
  • experimentation
  • and quality improvements over time

---

10) Next frontiers: memory/personalization and world models

• Biggest “next frontier” candidates:
  • Memory and personalization, including better systems than simple recency/frequency
  • World models / spatial intelligence, framed as improving “intelligence itself” (understanding physical realities and how the world works)

---

Presenters / contributors

• Jacob Effron (host; investor at Red Point; presenter for “Unsupervised Learning”)
• Swyx (co-host/contributor from Latent Space)
• Harrison Chase (LangChain CEO; referenced)
• Matt Billman (Netlify; referenced)
• Malte Ubl (Vercel CTO; referenced)
• Bret Taylor (Sierra; referenced)
• Max (Lagora) (referenced)
• Alex Wang (referenced via “breakfast discussion”)
• Fei-Fei Li (referenced via world models essay)
• Geoffrey Hinton (referenced via bio-safety comment)
• Geoffrey Hinton is mentioned in the conversation context; also “Ryan LePopolo” and “Ankur Goyal” appear as prior podcast guests (referenced)