Chamath's Singularity Loop Is Half Right

Capability is accelerating on longer-horizon tasks, but options flow and compute markets show no recursive AGI loop yet

Chamath says we are in the loop. The tape doesn't show it yet.

Chamath posted today that the AI singularity loop is active: humans build AGI, AGI does AI research, designs smarter AI, cycle repeats faster. He says the next 18 months will be wild and marginal costs go to ~$0.

The claim has two parts: capability is accelerating, and the market is positioning for it. One is visible in public results. The other is not visible in options flow.

What options flow shows: no singularity bet

If traders believed a recursive takeoff was starting, you would expect unusual call volume, expanding long-dated call open interest, and short-gamma positioning where dealers amplify moves.

We see the opposite on Aug 3:

• No unusual volume flag on any name.
• Long-gamma means dealers dampen moves, price pins — opposite of a gamma squeeze.

LEAPS (dte >=365) conviction detail:

No large new call positions. The tape is pinned, not chasing.

What public capability data shows: old benchmarks are dead

Public leaderboards can no longer show acceleration. AIME, GPQA Diamond, SWE-bench Verified, MMLU are all at or near 95-100% for frontier models. Only a handful of harder evals like Humanity's Last Exam and FrontierMath still have headroom.

That is why lab results over the past few weeks feel like a step-change to practitioners but look flat on public leaderboards. The field has moved to longer-horizon, agentic tasks that are not captured by those saturated benchmarks.

This supports Chamath's "look at results from labs" intuition directionally, but it does not yet prove recursive self-improvement where AGI designs smarter AI.

What would confirm a real loop in compute and benchmarks?

A true recursive loop would have to show up in three places that are publicly observable over the next 18 months:

1. Compute allocation shifts. Labs would pull inference capacity offline to train the next self-improved model. You would see cloud GPU spot availability tighten sharply and on-demand pricing rise, plus hyperscaler capex guidance revised up mid-year for training.

2. Benchmarks change. Saturated benchmarks stay saturated. The signal moves to how long an agent can work autonomously without failing. If doubling time for that horizon collapses from months to weeks, that is a loop signal.

3. Costs. Chamath says marginal costs go to ~$0. That is already happening for open-source inference because of competition and efficiency, not because of AGI self-improvement. The tell for a loop would be frontier closed-model prices dropping 5-10x in a quarter while capability jumps, not the steady 20-30% cuts we see now.

Right now we have steady inference price declines driven by open-source competition and human-driven post-training improvements, not a supply crunch from recursive training.

Bottom line

Capability is accelerating on longer-horizon tasks, which makes the next 18 months likely to be transformative as Chamath argues. But options positioning in NVDA, AMD, and GOOGL shows no large speculative call chase, and compute markets are balanced, not acute.

If the loop were active, you would expect unusual call volume, expanding LEAPS call OI, and tightening GPU availability. None are flashing yet.

What to watch: the next frontier model releases and whether hyperscalers revise capex up outside their normal cycle, plus whether long-dated call OI in NVDA and GOOGL starts expanding with net positive OI change week over week.