brain/
conceptartificial-intelligence

The three camps of AI: superintelligence-quest / sober-researchers / value-capture

Notes

The three camps of AI: superintelligence-quest / sober-researchers / value-capture

Vintage: Dec 2025. Ali Ghodsi's three-camp framing on the Bg2 Pod December 23, 2025. Useful as a map of which camp's framing applies to a given debate (timelines, capex, regulation) rather than as a claim about which camp is right.

One-line summary: ali-ghodsi in late 2025 partitioned AI thinking into three camps with very different operational assumptions: (1) the superintelligence-quest camp at the frontier labs, pursuing recursive-self-improvement through scaling-laws-driven compute build-out; (2) the sober-researchers camp (Sutton, LeCun, etc.) who say current LLM architecture is the wrong path and AGI is 20 years out; (3) the value-capture camp (Databricks, Glean, most enterprise vendors) who say "we already have AGI" and the work is to deploy it inside organizations. Conversations across camps frequently talk past each other because the camps have different success criteria, capital intensities, and bottlenecks.

The three camps

Camp 1 — Superintelligence quest

Membership: The frontier labs (OpenAI, Anthropic, xAI, Google DeepMind, Meta AI). Approach: scaling laws + larger compute clusters + RL. Success criterion: solve harder benchmarks (Math Olympiad, Physics Olympiad, programming contests) on the way to recursive self-improvement that "cures cancer and 10× GDP."

ali-ghodsi in 2025-12-23-bg2-databricks-glean-enterprise-ai: "Any of your cost equations are going to pale in comparison to the economic value that this thing is going to provide. So that's like one camp. And the way they're developing it is bigger and bigger clusters, more and more energy, and that's how they're going about it. And that's where most of the capital is going to."

Capital intensity: extreme. This is where the $200B+ Nvidia + half-trillion CapEx debate lives. The camp's bet is that the economic upside makes the math work if recursive-self-improvement materializes — and bankrupts everyone if it doesn't.

Cross-link: autoresearch-recursive-self-improvement tracks the first first-person evidence (Karpathy on his nanochat model) that this camp's premise is moving from "speculative" to "demonstrably operating, even if small-scale."

Camp 2 — Sober researchers

Membership: The Turing Award winners and senior academic figures who invented the underlying techniques — Rich Sutton (RL), Yann LeCun (CNN / self-supervised). Approach: argue current LLM architecture is fundamentally wrong for AGI. Success criterion: a different paradigm that learns the way animals / humans do.

ali-ghodsi in 2025-12-23-bg2-databricks-glean-enterprise-ai: "Rich Sutton, who created reinforcement learning... You have Yann Lecun, who was one of the three founding fathers... they've been saying that first camp is not going to. That's not even the right approach, is their view. They're like, no, that's just like autoregressive next token prediction... that's not how humans learn. That's not how animals learn. We operate in a different way... they say it's 20 years out."

Capital intensity: low. The camp is doing research, not building products. Ghodsi notes they're "the ones that arrived, unfortunately" — implying their underlying intuitions may turn out to be right while having little near-term operational consequence.

Camp 3 — Value capture inside enterprises

Membership: Databricks, Glean, most enterprise AI vendors, most application-layer startups. Approach: take current LLMs (acknowledged-to-be-commoditizing — see llm-as-commodity-thesis) and integrate them into proprietary data and workflow patterns to extract economic value. Success criterion: move the ai-coding-productivity-paradox "5% of projects working" to 10%, 20%, 30%.

ali-ghodsi in 2025-12-23-bg2-databricks-glean-enterprise-ai: "I don't think we need super intelligence. Like, I don't think we need that super intelligence right now. Maybe they'll get there. That's awesome if they do. But I think we have AGI. I think we have artificial general intelligence. We really have it... if we already have AGI, we just need to make it useful inside the enterprise. We need to just expand that 5% to be 10%, 20%, 30%."

Capital intensity: moderate. Operator-grade product development. Ghodsi explicitly notes: "We're not in a bubble in a sense that we're not spending huge amounts of capital on what we are doing. We're just trying to get actual economic value inside of these organizations."

Why it matters to this thread

The thread accumulates evidence from sources in different camps and the framing helps decide which body of evidence applies to a given question:

  • "How fast will AI capability advance?" — Camp 1 says weeks (compute scaling); Camp 2 says decades (paradigm shift needed); Camp 3 says "we already have what we need, the question is deployment skill." The ../../_meta/AI_CAPABILITY_TRACKING discipline implicitly tracks Camp 1's view through frontier-lab releases and Camp 3's view through practitioner adoption signals.
  • "Is the AI capex bubble real?" — Camp 1 is the bubble (potentially); Camp 3 is not. Ghodsi: "There is a bubble. There is a super intelligence quest camp. I would be very worried there... we're not spending huge amounts of capital on what we are doing." Sax's anthropic $44B-ARR-by-April framing is a Camp 1 datapoint; Salesforce's $300M/yr anthropic spend (Benioff) is a Camp 3 datapoint.
  • "What does AGI mean?" — Camp 3 says we have it (the 2009-Berkeley definition is satisfied; we've moved the goalposts). Camp 2 says we don't and won't soon. Camp 1 says we're approaching superintelligence past AGI. Disagreement is largely about definition, not capability data.

Evidence

The full camp framing

  • ali-ghodsi in 2025-12-23-bg2-databricks-glean-enterprise-ai: "I think there's like three paradigms or three kind of camps. Let's start with the first camp. I think the first camp is this quest for super intelligence camp... it's really still being a lot of it comes from the scaling laws mentality... There's a second camp, which are the people that created the original technology. The scientists who created the technology... Rich Sutton... Yann Lecun, who was one of the three founding fathers... they say it's 20 years out... Third camp, which is, I think what we are in... I think we have AGI."

Tension with the "physics problem" framing

  • apoorv-agrawal in 2025-12-23-bg2-databricks-glean-enterprise-ai: "On the semi side of things, assuming that is just 50% of the CapEx, you're spending about half a trillion on CapEx. And then you've got to earn about a trillion dollars of AI revenue for all of this capex to be worth it... This seems like a physics problem at this point." — frames the Camp 1 bet's existential math problem.
  • arvind-jain in 2025-12-23-bg2-databricks-glean-enterprise-ai: "AI is not actually extending software in a marginal way. It's a different product and in fact it's actually going to grab a lot of revenue that actually today is in services industry which is 25 times larger than software industry." — partial answer: the trillion comes from services-industry disruption (cross-link ai-as-services-disruption), not from extending the software TAM.

Camp 3 operator framing on Camp 1

  • ali-ghodsi in 2025-12-23-bg2-databricks-glean-enterprise-ai: "How do they know that they're succeeding? They're not just like, oh, trust us, they're very smart people working on this... they're throwing all the most intellectually challenging [problems]." Recognition that Camp 1 has internally-coherent success criteria even if Camp 3 doesn't share them.

What this is not

The camp framing is a map, not a verdict. Each camp has internally-coherent assumptions and operational consequences. The wiki should not collapse a Camp 1 forecast into a Camp 3 deployment claim, or vice versa. When sources from different camps appear to contradict each other on AGI timelines or capability levels, the contradiction is often about which camp's success criterion the speaker is using rather than about disagreement on capability data.

Related

Referenced by
Concepts
ChatGPT super-assistant vision: chat → actions + proactivityLLMs are a commodity; the moat is proprietary data + workflow integration
Entities
Ali Ghodsi