{
  "slug": "australian-frontier-dependence",
  "title": "Can Australia benefit from AI without controlling the frontier?",
  "topic": "Australia and the frontier",
  "answer": "Australia can gain from imported AI while remaining dependent on the people who control it. My concern is that AI-assisted research could make that dependence harder to escape. Cheaper access would help businesses, but would not settle ownership, continuity or bargaining power.",
  "card_summary": "Imported AI can create Australian gains while leaving frontier dependence intact. AI-assisted research could compound the leaders’ advantage, but the size and persistence of that effect remain unsettled. Useful adoption, local continuity and assured frontier access need separate tests.",
  "scope": "Australian access, capability and bargaining power as AI development becomes increasingly AI-assisted",
  "standing": "Personal assessment · conditional outlook",
  "confidence": "High that useful adoption and frontier control are different; lower on the future concentration and pace of frontier progress",
  "body": "## My current answer\nYes, Australia can benefit from AI without owning the leading models. Better services, cheaper inputs and more capable local businesses are real possibilities. I remain committed to helping people use these tools well.\n\nBut I am increasingly doubtful that waiting for cheaper models provides an adequate national strategy. If the best AI helps its owners build the next generation faster, a frontier lead can become an advantage in producing further advances.\n\nMy expectation is that recursive self-improvement will happen. In the broader sense of AI helping improve the systems behind AI, it is already happening. Fully autonomous development of successive frontier models remains a stronger, unproven claim. I expect further progress towards it; that expectation should be tested rather than presented as an established result.\n\nThe question for Australia is whether we can capture substantial benefits while preserving meaningful choices about access, control and the income generated.\n\n## The feedback loop deserves attention\nThe mechanism is straightforward: a better model helps with research and engineering; those improvements help produce a better model. It can matter well before humans disappear from the process.\n\nGoogle DeepMind reports that AlphaEvolve improved a kernel used in Gemini training by 23%, reducing overall training time by 1%. That is a concrete contribution to AI development, with a much smaller system-wide gain than the headline kernel improvement. [AlphaEvolve report](https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/).\n\nAnthropic reports large increases in code production and growing use of AI in research. Its eightfold code-output measure is not an eightfold productivity estimate, and employee estimates of faster work are self-reports. The company still identifies research judgement as a major gap. [When AI builds itself](https://www.anthropic.com/institute/recursive-self-improvement).\n\nMETR’s evaluation of Claude Opus 5.5 supports likely acceleration of parts of AI research, while judging full research-and-development automation unlikely. It does not establish a trend in the rate of acceleration. Anthropic had the opportunity to review and edit the report, and METR signed off on the final text. [METR evaluation](https://metr.org/blog/2026-09-22-claude-opus-5-5/).\n\nThese findings make the feedback mechanism credible. They leave its eventual strength and limits unresolved.\n\n## Why a small lead might matter more\nA model that reliably completes a research workflow its competitor cannot complete may unlock a different way of working. Extra capability can cross a practical threshold, allowing more experiments, quicker implementation or less supervision.\n\nMy concern is that the leading laboratories could repeatedly cross those thresholds first. Their advantage could compound, potentially very quickly, while followers face worsening catch-up prospects.\n\nThere is an important qualification. Exponential improvement alone does not prove an ever-widening relative gap. Two systems following the same exponential path a fixed time apart can retain a constant ratio, even as their absolute difference grows. A claim of widening relative advantage needs evidence of unequal research acceleration or a growing catch-up lag. Benchmark scores are not measurements of intelligence on a ratio scale.\n\nI would watch research output across successive generations, reliable completion of difficult work and the tools laboratories use internally. A public leaderboard cannot reveal the full advantage of an unreleased model used to build its successor.\n\n## Catch-up is becoming a harder assumption\nDistillation can transfer capability by training on another model’s outputs. Leading US laboratories are actively trying to make unauthorised extraction harder.\n\nAnthropic reports that its defences reduced reasoning tokens leaked to attackers by roughly 10 to 100 times on its own platform. Their durability against highly sophisticated, persistent adversaries remains uncertain. OpenAI describes defensive information-sharing through the Frontier Model Forum and coordination with cloud providers. These are efforts to restrict one catch-up route, rather than proof that every route is closed. [Anthropic’s defences](https://www.anthropic.com/research/distillation-defenses), [OpenAI’s account](https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign/).\n\nThese reports do not establish how much the defences have affected Chinese model development or caused the current capability gap.\n\nDeepSeek remains impressive. The September V4.1 Flash release was materially behind GPT-6 Astra on several demanding tasks in the independent comparison. Artificial Analysis’s comparison of the September releases supports substantial gaps on terminal work and GDP.pdf, alongside some Flash strengths. This is a task-dependent assessment, not a universal ranking of intelligence. [Independent comparison](https://artificialanalysis.ai/models/comparisons/deepseek-v4-1-flash-vs-gpt-6-astra).\n\nThe 17 October Flash update reports improvements on DeepSeek’s own evaluations. That comparison cannot be treated as a matched test of that new checkpoint. Fresh independent testing could change the picture. [DeepSeek’s model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017).\n\nI am watching the next few weeks closely for a major Chinese release that challenges this assessment. Without a convincing response, I would become more concerned that countries outside the leading providers face a choice between paying for access and accepting a capability disadvantage. Release timing is an indicator, not a scientific deadline after which catch-up becomes impossible.\n\n## Cheap adoption helps, but leaves another question\nIn the IMF discussion, India’s chief economic adviser V. Anantha Nageswaran made the case for benefiting as a later adopter, avoiding some early capital risk while models become cheaper. That is an argument about India. Applying it to Australia is my comparison. [Discussion recording, around 67:43](https://www.youtube.com/watch?v=Rqt9kWoTJbc&t=4063s).\n\nI can see the adoption case. Businesses often benefit from technologies they did not invent. But I regard the stronger hope that cheaper later adoption will resolve frontier dependence as wishful thinking.\n\nAccess to a service does not usually provide the model weights, training process, internal research tools or right to keep operating independently. Competition among US suppliers could drive prices down while Australia remains dependent on their decisions and the laws governing them.\n\nThe distinction matters for my existing proposal for [Australian-controlled inference](/articles/australian-control/). A legally and operationally credible local option could protect continuity for useful work. Maintaining parity with the latest frontier would be an additional and much harder test.\n\n## What would change this answer\nThe strongest challenge to my concern would be repeated evidence that followers close consequential gaps despite weaker models and extraction defences. Open models that remain competitive across generations, independent discoveries that overturn an incumbent advantage, or research bottlenecks that stronger AI does little to relieve would all matter.\n\nEvidence that frontier advances add little value for most Australian work would also change the practical stakes. So would durable access arrangements, low switching costs and substantial gains retained by Australian workers, businesses and consumers. Dependence can coexist with worthwhile benefits.\n\nConversely, reliable AI research improving faster inside a few laboratories, widening gaps on comparable tasks and restrictive access conditions would strengthen my concern. Prices alone would not settle it.\n\n## What Australia should do meanwhile\nI would keep working on adoption, skills and useful local applications. Waiting for national frontier ownership before helping a small business would sacrifice benefits available now. [Practical capability](/articles/capability-and-adoption/) remains essential.\n\nAlongside that work, procurement and infrastructure decisions should examine continuity rights, portability, dependencies and the cost of switching. Public investment should specify the capability it secures and how that capability stays useful as models advance.\n\nForeign model profits do not by themselves establish an Australian loss. Imported AI can support domestic expertise, better services and consumer gains. The [national benefit assessment](/articles/australian-net-benefits/) still needs retained income, transition costs, resource costs and distribution, without counting the same cost twice.\n\nMy position is to pursue those benefits while taking frontier dependence seriously. The possibility of a compounding lead makes that question more urgent. It does not yet give us a measured answer about how far the lead will grow, or a blank cheque for any project labelled sovereign.\n\n## Limits of this edition\nThis is my current judgement, not a forecast model or a claim of inevitable permanent US dominance. Laboratory reports contain interested-party evidence; benchmark comparisons are dated and task-specific. The IMF passage was checked against automatically generated captions, with speaker attribution inferred from context rather than separately audio-verified. New releases and independent evaluations may warrant a prompt revision.\n",
  "sources": [
    "alphaevolve",
    "ai-builds-itself",
    "metr-opus55",
    "distillation-defences",
    "distillation-coordination",
    "flash-astra-comparison",
    "flash1017",
    "imf-later-adoption"
  ],
  "published": "2026-10-18",
  "evidence_through": "2026-10-18",
  "version": "v1",
  "release": "2026-10-18-v1",
  "source_records": [
    {
      "id": "alphaevolve",
      "title": "AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms",
      "author": "Google DeepMind",
      "published": "2025-05-14",
      "url": "https://deepmind.google/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/",
      "kind": "Laboratory report of deployed technical improvements",
      "limitations": "Reports a 23% improvement to a Gemini training kernel, translating to a 1% overall training-time reduction. A concrete AI contribution to AI development, not autonomous successor creation, accelerating returns across generations or an economy-wide productivity effect.",
      "checked": "2026-10-18"
    },
    {
      "id": "ai-builds-itself",
      "title": "When AI builds itself",
      "author": "Anthropic Institute",
      "published": "2026-09-18",
      "url": "https://www.anthropic.com/institute/recursive-self-improvement",
      "kind": "Laboratory evidence and interpretation",
      "limitations": "Updated 18/09/2026. Eightfold code-output growth is an imperfect productivity proxy; employee-reported gains may overstate effects. The report identifies research judgement and other bottlenecks and says fully autonomous successor development has not been achieved. Future scenarios are possibilities, not measured outcomes.",
      "checked": "2026-10-18"
    },
    {
      "id": "metr-opus55",
      "title": "Summary of METR’s predeployment evaluation of Claude Opus 5.5",
      "author": "METR",
      "published": "2026-09-22",
      "url": "https://metr.org/blog/2026-09-22-claude-opus-5-5/",
      "kind": "Preliminary capability and research-acceleration assessment",
      "limitations": "Limited-duration API testing and additional information with access restrictions. Likely researcher acceleration and limited R&D automation do not establish full automation or an acceleration trend. Anthropic could review and edit the text; METR signed off on the final version.",
      "checked": "2026-10-18"
    },
    {
      "id": "distillation-defences",
      "title": "Strengthening our defenses against distillation attacks",
      "author": "Anthropic",
      "published": "2026-10-12",
      "url": "https://www.anthropic.com/research/distillation-defenses",
      "kind": "Provider account of defensive measures",
      "limitations": "Reports roughly 10–100-fold lower reasoning-token extraction on its platform. Durability against sophisticated persistent circumvention remains uncertain. This does not measure the effect on Chinese model development or establish an impenetrable frontier.",
      "checked": "2026-10-18"
    },
    {
      "id": "distillation-coordination",
      "title": "Disrupting a coordinated model-distillation campaign",
      "author": "OpenAI",
      "published": "2026-09-30",
      "url": "https://openai.com/index/disrupting-a-coordinated-model-distillation-campaign/",
      "kind": "Provider account of defensive coordination",
      "limitations": "Describes defensive information-sharing through the Frontier Model Forum and work with cloud providers. It does not establish shared frontier R&D, a permanent closed boundary, a measured effect on the capability gap or the elimination of independent catch-up routes.",
      "checked": "2026-10-18"
    },
    {
      "id": "flash-astra-comparison",
      "title": "DeepSeek V4.1 Flash (Max) versus GPT-6 Astra (Max)",
      "author": "Artificial Analysis",
      "published": "Live comparison; checked 18/10/2026",
      "url": "https://artificialanalysis.ai/models/comparisons/deepseek-v4-1-flash-vs-gpt-6-astra",
      "kind": "Independent benchmark aggregation",
      "limitations": "September releases under Intelligence Index v4.3.2: Astra leads on the aggregate and several difficult tasks; Flash has strengths on other measures. Scores are not intelligence ratios. This is not a matched independent evaluation of DeepSeek’s 17 October update.",
      "checked": "2026-10-18"
    },
    {
      "id": "flash1017",
      "title": "DeepSeek-V4.1-Flash-1017",
      "author": "DeepSeek",
      "published": "2026-10-17",
      "url": "https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017",
      "kind": "Official model card and vendor benchmarks",
      "limitations": "Vendor-reported improvements for the 17 October checkpoint, including Terminal-Bench 4.0 and DeepSWE. Its comparison set does not settle performance against the latest US frontier releases. Older independent scores must not be attributed to this checkpoint.",
      "checked": "2026-10-18"
    },
    {
      "id": "imf-later-adoption",
      "title": "IMF discussion: later adoption and AI economics",
      "author": "IMF",
      "published": "2026-10-17",
      "url": "https://www.youtube.com/watch?v=Rqt9kWoTJbc&t=4063s",
      "kind": "Recorded discussion; automated-caption verification",
      "limitations": "Passage around 67:43 checked against full automated captions in the parent’s research. Speaker attribution inferred from context, not separately audio-verified. The India-to-Australia comparison and criticism of later adoption as a sufficient frontier strategy are Alex’s analysis.",
      "checked": "2026-10-18"
    }
  ],
  "change": {
    "old": "No earlier site edition.",
    "new": "New personal assessment of AI-assisted research, catch-up constraints and Australian frontier dependence.",
    "why": "Requested on 18/10/2026. Sets out the concern that AI-assisted research may compound frontier advantage, distinguishes that outlook from demonstrated results, and separates useful adoption from ownership, operational continuity and assured frontier access.",
    "confidence": "High on the distinction between adoption and control; lower on future concentration and pace",
    "practical_effect": "Pursue useful adoption while testing continuity, portability and dependency. Includes evidence that would weaken or strengthen the assessment; no inevitable permanent dominance or measured national loss is claimed."
  }
}