The Memo - Special edition… | LifeArchitect.ai Summary

1. Key Themes

GPT-6 Astra represents a step-change toward "proto-ASI" status

The author explicitly classifies this model, alongside Anthropic's latest, as crossing a meaningful capability threshold rather than being an incremental update.

"As with Anthropic Claude Mythos 5 and Claude Fable 5, OpenAI GPT-6 Astra should be considered proto-ASI."

This is reinforced by benchmark performance that exceeds prior expectations:

"GPT-6 Astra hit 99.9% on ARC‑AGI‑3 (analysts predicted it would take years to solve this one), 97.6% on FrontierMath Tier 4, and an absurd score of 100% on ExploitBench."

The shift from verbalized reasoning to internal ("looped") computation

A key architectural theme is that models are becoming more capable without needing to write out chain-of-thought, suggesting a new computation paradigm.

"When we prevent the model from reasoning, we observe the set of tasks Astra is able to accomplish without the use of CoT is greatly expanded compared to prior models…"

The author explains the mechanism hypothesis:

"In a looped transformer, some processing layers run again on their own updated results. Each pass changes the internal representation without needing to produce another word."

Capability overhang is the central unresolved question, not raw intelligence

Rather than asking how smart a model is, the author argues the more important question is how much of its capability remains undiscovered — drawing a direct historical parallel to GPT-2.

"Consider that even now in the second half of 2026, we are still discovering the capabilities of the GPT-2 model from 2019(!)." "The important question is therefore not merely, How intelligent is GPT‑6 Astra? It is: How much of GPT‑6 Astra have we actually discovered? The answer today is: very little."

Token efficiency as a new competitive axis

Beyond raw benchmark scores, efficiency gains are emerging as a differentiator among frontier labs.

"Astra used the 3rd least tokens on Model Zen Garden—10x more token efficient than Fable 5.1." "it used just 2.3 million tokens (around 10× fewer than Fable 5.1 and 42× fewer than Opus 5) while still producing a competitive result."

Benchmarks designed to be un-gameable are being exhausted

The retirement of the author's own benchmark signals the field is running out of headroom for pure knowledge/reasoning text tests.

"Given that virtually all science questions are now solved by frontier models, I cannot think of further unique text questions that would stump a frontier AI model, unless we start getting into personal, private questions that would require data breaches."

2. Contrarian Perspectives

Model size is no longer a meaningful proxy for capability

Despite years of scaling-law narratives dominating AI discourse, the author explicitly downgrades the importance of parameter/size estimation.

"General model size is no longer an indicator of performance, but I still find it interesting."

Architectural claims about "looped transformers" remain unconfirmed despite widespread assumption

Even though the looped-transformer explanation is compelling and widely discussed, the author is careful to note it's unproven — a caution against over-indexing on plausible-sounding technical narratives.

"OpenAI's published system card establishes stronger performance without verbalized reasoning, but does not establish that looping is the mechanism responsible. So, while the observed capability is documented, the architectural change for GPT-6 remains unconfirmed at this stage."

The real ceiling on evaluation isn't model capability, it's the availability of novel questions

Rather than framing benchmark saturation as a limitation of the model, the author frames it as a limitation of human ability to construct hard-enough tests — implying models may already exceed our capacity to evaluate them with public information.

"I cannot think of further unique text questions that would stump a frontier AI model, unless we start getting into personal, private questions that would require data breaches."

3. Companies Identified

OpenAI — Creator of GPT-6 Astra. Mentioned as the release subject of this entire edition and for pushing frontier benchmark and architecture innovation.

"OpenAI releases GPT-6 Astra" "GPT-6 is a new class of model, with recurrent depth or looped transformer-like architecture, enabling a significant increase in capabilities."

Anthropic — Maker of Claude Mythos 5, Claude Fable 5, and Claude Fable 5.1. Mentioned as a comparison point for proto-ASI classification and as a benchmark competitor that underperforms Astra on specific tests.

"As with Anthropic Claude Mythos 5 and Claude Fable 5, OpenAI GPT-6 Astra should be considered proto-ASI." "the new Claude Fable 5.1 model cannot solve any of the three questions above"

LifeArchitect.ai — The author's own research/benchmarking operation. Mentioned as the source of the ALPrompt benchmark and ASI tracking framework being referenced throughout.

"You can view my 50 lagging indicators at LifeArchitect.ai/ASI."

4. People Identified

Dr Alan D. Thompson — Author of The Memo newsletter and creator of the ALPrompt benchmark and ASI criteria framework. Mentioned as the primary evaluator who tested GPT-6 Astra directly and delivered a live demo to government officials.

"I showed the GPT-6 launch video to around 600 government staff for my opening keynote on Friday morning Adelaide time" "It should be noted that the PhD human verifier required an additional round of 'steering' to arrive at solutions, where GPT-6 Astra solved them immediately."

Sebastian Raschka — Referenced as an external technical commentator analyzing the looped transformer architecture. Mentioned as further reading on the mechanism behind Astra's efficiency.

"Sebastian Raschka: OpenAI Astra and Looped Transformers" (cited as further reading)

5. Operating Insights

Treat newly released frontier models as having largely unmapped capability, not fully realized capability. The GPT-2 example suggests real-world utility surfaces years after release, meaning entrepreneurs should expect continued discovery of new use cases well after launch rather than assuming the initial demo/benchmark set defines the ceiling: "Better prompts, longer inference, new tools, memory, multi-agent scaffolds, and domain-specific fine-tuning will continue converting hidden capability into reliable performance. This is the low-hanging fruit."
Watch token efficiency, not just accuracy, as a cost/product design variable. With a 10x–42x efficiency gap reported between models on the same task, this has direct implications for unit economics of AI products built on top of these APIs.
Autonomous multi-step agentic workflows (e.g., overnight unsupervised research-to-render pipelines) are now viable with minimal human steering. This is a template for building products/workflows that operate with reduced supervision: "I steered it a few times, but I didn't really need to... The bulk of the run was done overnight. I woke up this morning to the rendered video sitting on my desktop."

6. Overlooked Insights

Pricing as a sizing proxy has become a standard analytical technique amid disclosure opacity. The author uses cost structure ($10/M input, $50/M output tokens) combined with GPU supply/demand signals to reverse-engineer model characteristics labs no longer disclose — a useful due-diligence technique for investors trying to assess competitive positioning without official specs: "Pricing as a sizing signal. $10 per million input tokens and $50 per million output tokens."
Emergent multi-agent social behavior appeared unprompted and startled its own creator, suggesting current agent scaffolds may produce unexpected emergent coordination/communication behaviors beyond what was explicitly designed for — a potential signal for both opportunity (autonomous multi-agent products) and risk (unpredictability): "I walked out, honestly a little scared. It was the Astra agents. They'd started talking to each other."