06-reference

innermost loop singularity horizon

2026-06-28·reference·source: Innermost Loop·by Alex Wissner-Gross

"Welcome to June 28, 2026" — @Alex Wissner-Gross

Why this is in the vault

Wissner-Gross's framing that the AI singularity has "stopped having a ceiling and started having a horizon" — with GPT-5.5 able to think for weeks before benchmarks flatten — is the clearest statement yet that the plateau narrative has inverted into an open-horizon race, with direct implications for how RDCO calibrates model-tier decisions and agent architecture.

The core argument

The June 28 issue is a dense weekly curation across five fronts: (1) frontier model capability, (2) cyber-AI as geopolitical lever, (3) AI cost engineering, (4) compute infrastructure, and (5) physical-world automation and scientific discovery.

Frontier capability. Noam Brown (OpenAI) says the plateau is "really far out these days" — a well-scaffolded GPT-5.5 can sustain useful reasoning for weeks. Nous Research exposed mixture-of-agent presets claiming to beat Opus 4.8 by 8% and GPT-5.5 by 11%. DeepSeek's MIT-licensed V4-Pro-DSpark packs 1.6T parameters and a 1M-context window while using a tenth the KV cache via speculative decoding. Shopify's CTO pairs GPT-5.6 (broad experiments) with Fable 5 (code writing), treating them as complementary rather than competing.

Cyber-AI geopolitics. Anthropic is reportedly set to restore Fable 5 within the week — framed as a "backhanded tribute to real cyber teeth." Chinese systems are matching Mythos in cybersecurity evals; China's 360 and Tokyo's Sakana are shipping Mythos-shaped alternatives. The NSA's own cyber benchmark (due August) is described as the gate that clears Fable 5 and GPT-5.6 for public deployment.

Cost engineering. Coinbase nearly halved AI spend while usage soared — achieved via cheaper model defaults, smarter routing, and lifting cache hit rate from 5% to 60%. Gartner projects AI coding costs will outrun the average developer's salary by 2028. The "loop engineering" paradigm is explicit: Claude Code's creator reportedly no longer writes his own prompts — Claude does.

Compute and power. IBM is committing $11B toward quantum (including a 2029 Starling machine). Google throttled Meta's Gemini access due to scarcity. General Fusion tripled plasma to 8.4M degrees via mechanical squeeze. Renewables hit 30% of US generation, with solar overtaking wind.

Automation and discovery. AGIBOT has shipped its 15,000th humanoid robot with a 39% market share. A $300 ForceBand wristband trains robot hands from human muscle signals at 87% accuracy on pick/squeeze/place. Erdős problem #870 fell to GPT-5.5-Pro, verified across 180,000 lines of Lean 4. A lone PhD synthesized PAC-832 (first selective GalR1 Alzheimer's antagonist) in a garage lab running Claude Code and a liquid-handling robot.

Mapping against Ray Data Co

Reinforces (strong) — loop engineering is the right bet. The explicit framing of Claude Code's creator not writing his own prompts — "Claude does" — validates the RDCO agent-unhobbling thesis. The phData DSA role depends on the COO agent having the toolset and visibility to do compounding work; this issue confirms the frontier is actively building toward that pattern, not away from it.

Reinforces (medium) — open-weights tier as real alternative. DeepSeek V4-Pro-DSpark at a tenth of KV cache cost extends the pattern from the April 26 AlphaSignal note. For RDCO client recommendations, the 18-month gap between frontier and open-weights is narrowing in cost terms even if capability gap holds.

Operational watch item — cache hit rate is a lever. Coinbase's 5%→60% cache lift is a concrete, repeatable cost-engineering move. RDCO's agent workloads (vault ingestion, newsletter processing, check-board loops) likely have exploitable cache patterns — worth auditing once phData workload stabilizes.

Tension — Gartner 2028 cost projection. If AI coding costs outrun average developer salary by 2028, the solo-founder cost structure for RDCO agents needs to scale with revenue, not just capability. Currently budget-controlled per founder directive; no action needed but worth tracking annually.

Related