16) Inga Strümke om hvor AI er på vei, den outsourcede problemløsningen og datamanifolden

Truls Jørgensen og Audun Fauchald Strand — Heis Podcast · podcast · 2026-05-04 source ↗ Reader 🖥️

theoretical · mixed — Predominantly theoretical (Strümke arguing for a particular understanding of where AI is and what it can/cannot do), but laced with practical/normative claims about Norwegian public-sector adoption, the chatbot-as-interface trap, and what KI Norge should do. The single tightest theoretical anchor is the data-manifold framing of why neural-net approximations may converge with reality for practical purposes but never coincide with it. The practical edge is a critique of 'outsourced problem-solving' and the Norwegian risk-aversion that fuels it.

Structure

Key terms

• datamanifold (the data manifold)
  Strümke (2026-05-04, ~36min): all data points that exist in reality form a 'teppe' (carpet/blanket) — possibly with holes — constrained by physical and logical rules; we move around the manifold via interventions; ML models, when trained on samples from it, build an internal approximation of that manifold. The critical property: the model's internal manifold is discontinuous (full of holes), so it cannot smoothly traverse reality, even if it approximates it well on sampled points.
• outsourcet problemløsning Insourcing & In-House Capability
  Strümke: if you cannot build the system that solves your problem yourself, then using it does not mean 'you use AI' — it means you have outsourced problem-solving to the vendor of that system. She proposes this as a regulatory/narrative reframing for KI Norge: organizations should be required to say 'we outsourced to vendor X' rather than 'we use AI'.
• modningsreise (maturation journey) Insourcing & In-House Capability
  Strümke: the slow organizational/national journey toward genuine AI capability — proper IT infrastructure, data control, legal clarity, problem specification, in-house developers — which she claims ChatGPT has 'spent bein på oss' (tripped us up) by replacing it with 'buy Copilot licenses and force employees to use it 10x/day'.
• forbrukerteknologi (consumer technology)
  Strümke: the moment a technology becomes something end-users buy and use, mass adoption explodes — and so does the false belief that 'using it' equals 'understanding it' (her examples: laptops, mobile phones, cameras, and now chatbots).
• nevrosymbolsk AI (neurosymbolic AI) Neurosymbolic AI
  Strümke: the combination of hardcoded deterministic rules and probabilistic neural models; she frames it as an open academic debate (do we need neurosymbolic, or are neural nets enough?), and observes that Claude Code's use of regex for the literal string 'continue' instead of routing to the language model is an implicit production-level vote for neurosymbolic.
• forklarbar AI (XAI, explainable AI) Explainable AI
  Strümke: the research field (her own) concerned with rendering neural-net behavior meaningful to humans; she notes Google Scholar publications since 2017 have grown exponentially, so the Norwegian public sector's assumption that explainability is unattainable is out of date.
• universelle approksimasjonsbevis (universal approximation theorems)
  Strümke: the theoretical result that, given sufficient width/depth and data, a neural network can approximate any function — used as the 'for' side of the argument that nets can in principle match reality's manifold.
• AI-imperativet (the AI imperative)
  Strümke: the cultural/organizational pressure that 'we must use AI to become more efficient', which she frames as the wrong type of motivation — efficiency-via-tokens treated as the goal rather than as one tool among many.
• udetekterbare bakdører (undetectable backdoors) / one-shot detection AI Alignment & Safety
  Strümke: a planted behavior in a neural network's parameter space that triggers on specific input, embedded so that detection requires solving a hypothesis test against an unknown distribution; her research bets on whether these are provably undetectable or whether a one-shot method exists.
• Project Glasswing / Mythos Mythos
  Strümke: Anthropic's most-capable vulnerability-finding model (Mythos) released only to vetted partners (Project Glasswing); she argues this is theater, because she and Riegler showed that a swarm of five small 1.2B-parameter models on a MacBook can already find similar vulnerabilities — so Anthropic's restraint is neither a security mechanism nor sustainable.
• deterministisk vs probabilistisk Neurosymbolic AI
  Strümke: the stack from transistors → bits → assembly → high-level language has always been deterministic; the language-model layer on top is probabilistic — which is architecturally novel and risky as a translation layer.

Propositions

• explicitGenerative chatbots becoming AI's cultural moment was a sociological accident, not a technical inevitability — Strümke expected reinforcement learning / regression to be the breakthrough, with AlphaGo beating Lee Sedol in 2016 as the real watershed.conf: high
• explicitOnce a technology becomes consumer technology, mass adoption follows, and with it the false sense that 'using it' equals 'understanding it' — chatbots have made everyone think they understand AI.conf: high
• explicitChatGPT 'sabotaged' (spente bein på) Norway's AI maturation journey: the pre-ChatGPT path required IT infrastructure, data control, legal clarity, specifying problems, and putting developers on them; the post-ChatGPT path is buying Copilot licenses and forcing employees to use them 10x/day.conf: high
• explicitcorroboratedUsing natural language as the top interface to computing is the realization of a long-standing dream, but it introduces a probabilistic step on top of an entirely deterministic stack — and that asymmetry is a fundamental architectural problem, not a routine new abstraction layer.conf: high
  supports ←Presentation: The Ironies of A^2 I^2 (Ironies of Automation and Artificial Intelligence in Incidents) · p25 — Both treat non-determinism as a structural problem: a probabilistic layer undermines the interpredictability that coordination with deterministic systems relies on.
• explicitClaude Code's use of a regex to detect literal 'continue' rather than routing to the LLM is a frontier-product proof that neurosymbolic AI has already won in practice, regardless of the academic debate.conf: high
  supportsThe biggest advance in AI since the LLM · p2 — Strümke cites the same Claude Code deterministic-regex evidence as proof that neurosymbolic AI has already won in practice.
• explicitAn ML model's learned manifold can asymptotically approximate reality's manifold but is provably discontinuous (full of holes), so the two will never coincide — though for many practical purposes the difference may be undetectable.conf: high
• explicitAll models are wrong but some are useful — and the risk now is that they become 'hur useful som helst' (arbitrarily useful) and we stop remembering the difference.conf: high
• explicitLearning is physically painful — new neural connections must be made — so if AI agents do the work, junior developers will end up with worse intuition ('magefølelse'), worse tacit knowledge of code, and fewer of the 'physical' learning experiences that build judgment.conf: high
• explicitcorroboratedYou can only judge whether 'I should do this or the agent should' if you yourself are capable of doing it — so blanket AI delegation destroys the very faculty needed to delegate well.conf: high
  supports ←Presentation: The Ironies of A^2 I^2 (Ironies of Automation and Artificial Intelligence in Incidents) · p2 — Skill deterioration from disuse supplies the mechanism for why blanket delegation destroys the very faculty needed to delegate well.
• explicitIf you cannot build the system that solves your problem, you have not 'used AI' — you have outsourced your problem-solving to the vendor of that system. KI Norge's most useful intervention would be to enforce this distinction at the level of narrative and reporting.conf: high
• explicitClassical regression/classification ML models are tiny, deterministic, cheap to run (50 øre/day), explainable, and well-suited to Norwegian public-sector problems like sykefravær estimation — yet are being skipped in favor of expensive foreign LLM tokens.conf: high
• explicitExplainable-AI research has grown exponentially since 2017, so the Norwegian public sector's habit of treating explainability as a permanent blocker — derived from the 2021 Datatilsynet ruling on the NAV sandbox sykefravær case — is empirically out of date.conf: high
• explicitNorway's risk-aversion ('cover your ass' culture) and the Datatilsynet stance combine to push organizations toward buying Copilot licenses — which is actually higher risk geopolitically than developing in-house ML capability.conf: high
• explicitUS economic growth's lead over Europe over the past 10–20 years can be 'fremskrives' (attributed) to Big Tech's growth; by buying Big Tech products, Europe sponsors the very dynamic widening the gap.conf: high
• explicitcontestedAnthropic withholding Mythos via Project Glasswing is not a security mechanism — a swarm of five 1.2B-parameter models on a MacBook can already find similar zero-days; vendor self-restraint is structurally insufficient and unsustainable in a competitive market.conf: high
  contradicts ←This Week in AI: Rethinking the Agent Harness · p18 — Claims restricted access is justified for defensively valuable frontier capability, against Strumke's claim that vendor withholding is not a security mechanism at all.
• explicitNorwegian research-grant funding cycles are too slow for the speed of capability emergence; Norway needs coordination structures that mobilize research faster than the Forskningsrådet utlysningssyklus.conf: high
• explicitThe right Norwegian strategy is to drop overordnede visjoner and pick concrete cases (e.g. the NAV sykefravær case) where the state takes the risk on behalf of agencies that don't dare interpret the law — solving problem by problem, rinse and repeat.conf: high
• explicitStrümke's own modus is utilitarian (the working solution is the output), whereas the host Truls's modus is aesthetic (the elegance of his own code matters), and the two coexist within a single codebase depending on what's load-bearing — she will use agents for front-end fun, but open Vim for data-model and migration code.conf: high
• explicitStrümke is afraid of agents — they might delete files if you prompt wrong — and her default for serious work is Vim with deterministic muscle memory from age 19; sandboxes give 'deterministic protection around the probabilistic model'.conf: high
  supportsAgents don't know what good looks like. And that's exactly the problem. · p5 — Sandboxes as 'deterministic protection around the probabilistic model' restate the deterministic-guardrails imperative.
• explicitPublic engagement about ML in Norway should distinguish 'avklaringsspørsmål' (problems we already have methods for, just needing agreement, e.g. consumer-facing explanations that hold up in court) from 'forskningsspørsmål' (true blue-sky research, e.g. defining the cognitive limits of an agent).conf: high
• implicitUnder the right framing, an ML model can support a 'mer informert dialog' with users (e.g. NAV beneficiaries) than rule-based coding can — because the model can answer 'minste endringen du må gjøre for å endre dette' (smallest counterfactual change), not just 'why was the score high'.conf: high

Arguments

Narratives

• Strümke recoiling when gym-goers approach her mid-set to monologue about how they use ChatGPT — and the consolation that public visibility has at least stopped men from explaining to her what she already knows.
  → Establishes the chatbot-as-cultural-event phenomenon at street level, and her embodied skepticism toward the 'everyone now thinks they understand AI' dynamic.
• AlphaGo defeating Lee Sedol in 2016 as Strümke's personal 'now the future starts' moment — and her surprise that the wider world waited for a chatbot in 2022.
  → Reframes what 'where AI is going' means by relocating the technical breakthrough to a domain (reinforcement learning) that never became consumer technology.
• Google's data centers using only 3% of input electricity on CPU because ML-optimized airflow and cooling — pre-chatbot AI that nobody noticed.
  → Counter-evidence to the 'AI = chatbot' equation; supports the 'AI is much more than chatbot, but try giving an example' aside.
• The Claude Code source-code regex for the literal lowercase 'continue' — instead of routing the keystroke through the LLM — which Strümke posted on Instagram and was misread as 'doesn't understand regex uses less power than LLMs'.
  → Concrete proof-by-example that the frontier has already chosen neurosymbolic AI, and a barbed comment on how naturally-language framings prevent people from seeing the technical irony.
• The jazz student who must play one song for the entire first year before being allowed to move on — and only then discovers the principles generalize to all songs.
  → Analogy for 'depth-first' learning the host Audun describes for junior developers; supports the inference that mastery requires depth before breadth even (or especially) in an AI-assisted era.
• Strümke buying one of the first XPS gamer laptops as an NTNU student because she 'didn't want to learn garbage collection' — and bracing for the boys at Programvareverkstedet to throw her out.
  → Personal reminder that the 'this abstraction makes you worse' anxiety has been raised at every layer of the stack from manual memory management onward — and that being right about that history does not entail the current case is benign.
• Inga writing a book shortly after ChatGPT launched and refusing to let 'a wretched language model into my work' — but happily letting it draft sections of reports where the product is the goal, not the expression of her thinking.
  → Personal demonstration of the modal distinction she argues developers must learn to make between high-load-bearing and low-load-bearing artifacts.
• Standing at the table with cups during the recording, illustrating the data manifold: she can pick a cup up; she cannot make it hover; if she lets go it falls at 9.1 m/s² — and chatbots correctly predict that, which is why we have intuition that their language manifold tracks reality for everyday physics, even though we lack any such intuition for programming languages.
  → The narrative core of the entire 'datamanifold' frame — visualized in lieu of a whiteboard.
• Mikael Riegler at SimulaMet building a swarm-of-five-tiny-models scaffold on a MacBook that found vulnerabilities comparable to those Anthropic's Mythos found.
  → Carries the entire 'Glasswing is theater' argument and the broader claim that capability is no longer gated by vendor frontier access.
• Strümke's two PhD students working in opposite directions: one proving undetectable backdoors are undetectable; the other developing a one-shot detection method 'he is just as smart, one year younger'.
  → Personal stake / cliffhanger — and a structural illustration that AI security is still settling foundational mathematical questions, not just engineering ones.

Questions addressed

• ✅ Where is AI actually heading?
  Toward language-modeled chatbots and agents (the two are converging) as the dominant interface, because of natural-language affinity — but the deeper trajectory is whether neural-net manifolds can approximate reality well enough to feel indistinguishable from it, which is an open research question.
• ⭕ Will LLMs become 'good enough' to write all kinds of software, or are there fundamental limits?
  Strümke translates the question into the data-manifold framing: whether the language sub-manifold (training data) can carry enough information to represent the full target programming-and-system manifold is genuinely open. Humans lack intuition for programming-language semantics in the way they have intuition for falling cups, so verification methods are weaker here.
• ✅ What should KI Norge actually do?
  Enforce a narrative distinction between 'AI use' and 'outsourced problem-solving'; pick concrete cases (sykefravær) and have the state absorb risk that individual agencies refuse; reduce the over-reliance on facilitation/conferences; coordinate research faster than the Forskningsråd cycle. Strümke admits she doesn't fully understand KI Norge's current mandate.
• ⭕ How does Norway escape Big Tech dependency without isolationism?
  Strümke names it as 'grisevanskelig' and 'et politisk spørsmål'; we cannot pause, cannot match US datacenter scale, but Norway is the richest country in the world and if we cannot do it, no one can. The honest answer is: we don't have one yet.
• ✅ Is the felt erosion of programmer skill under AI assistance just the same fear we had about garbage collection?
  Strümke acknowledges yes (historical pattern recurs) but argues 'something is not historically new does not mean it is not dangerous' — we cannot see 30 years forward. The discussion ends in genuine open empirical uncertainty, with the host Audun's optimistic note (more programmers, lower barrier) sitting against Strümke's pessimism about tacit-knowledge erosion.
• ✅ Is Anthropic's Glasswing a meaningful security mechanism?
  No — Strümke and Riegler showed that a swarm of small open models can already find similar vulnerabilities. Vendor restraint is theatrical and unsustainable.
• ⭕ Can undetectable backdoors in neural nets be planted?
  Open — two of Strümke's PhD students are working on opposite hypotheses. She admits she believes both, which is logically impossible, and is excited to see which proves correct.
• ✅ Is XAI mature enough to support real public-sector ML deployment in Norway?
  Yes — the field has grown exponentially since 2017; the Datatilsynet 2021 stance on the NAV sykefravær case is empirically out of date.

Gaps

• incompleteThe 'outsourcet problemløsning' framing is sharp but Strümke does not develop the policy mechanics: what counts as 'building it yourself' on a continuum? OSS-finetuned-from-an-open-base, Norway-hosted closed-weight, fully in-house from scratch — the framing presupposes a binary that procurement policy would have to operationalize.
  Read next: Cuijpers / EU AI Act interpretation on capability sovereignty; Anu Bradford on Brussels effect; OpenEuroLLM / Silo AI literature on sovereign-capability tiers; recent NSM/Nkom positions on cloud sovereignty thresholds.
• incompleteThe data-manifold frame is presented as Strümke's primary lens but she explicitly says 'we don't have access to either manifold' — so the entire empirical apparatus that would let us know how close approximation has gotten remains undeveloped in the conversation. The host doesn't push on what an operational proxy would even look like.
  Read next: Yamins & DiCarlo on neural-net-vs-cortex representational geometry; Bengio on the manifold hypothesis as ML's organizing assumption; recent interpretability work (Anthropic's circuit-tracing, sparse autoencoders) as one empirical handle on the model side of the question.
• incompleteStrümke argues for neurosymbolic AI and points at Claude Code's regex as proof, but doesn't address the harder question: where does the symbolic boundary go? Claude Code's choice is for a single literal token; the production design pattern (when LLM, when regex, when AST, when proof) is unexplored.
  Read next: Gary Marcus's neurosymbolic writings; DeepMind's AlphaProof/AlphaGeometry as advanced symbolic-neural hybrids; the harness-engineering literature already in the user's library (Anthropic, Boris Cherny) for production-design patterns.
• incompleteThe Norway/Europe-vs-Big-Tech argument is geopolitically sharp but treats the US bloc as monolithic. The conversation does not engage with the question of whether dependency on US open-weight models (Llama, Mistral-Mistral-Mistral, etc.) is qualitatively different from dependency on closed-API products like Copilot — even though that distinction may be the most actionable lever Norway has.
  Read next: Together AI / Mistral / EuroLLM positioning papers; Sam Altman's geopolitical AI writings as counter-foil; Helen Toner / RAND on AI compute geopolitics; recent open-weight-vs-closed-API analyses from CSET.
  closed byAI Sovereignty: Bargaining with Big Tech and the Promise of Full Stack Open Source AI — Doesn't resolve the specific Llama/Mistral-vs-Copilot question, but reframes it: the piece argues the open-weight-vs-closed-API binary is the wrong lever entirely, since sovereignty depends on which layers of the whole stack (harnesses, inference, protocols, storage, observability) stay open, not on model weights alone.
• incompleteThe conversation reaches the 'cover your ass' diagnosis of Norwegian public-sector risk-aversion but does not investigate the structural causes — why has the last ~10 years been 'feigere'? Datatilsynet stance is named, the Nikolai Tangen episode is named, but the institutional feedback loop producing the pattern is not unpacked.
  Read next: John Cutler on institutional overload and accountability culture (already in the user's reading); recent NSM/Nkom assessments on resilience; Norwegian public-administration literature (NIBR / NIFU) on risk-distribution between politikere and direktorater; Annie Duke / Daniel Kahneman on outcome-bias driving risk-aversion.
• incompleteStrümke names two interesting categories — 'avklaringsspørsmål' (we have the methods, we just need agreement) and 'forskningsspørsmål' (genuinely open) — but doesn't extend the taxonomy. The third missing category is presumably 'capability-building-spørsmål' (we know what we need to build but lack the people / infra to build it) — which is arguably the dominant Norwegian bottleneck.
  Read next: Donella Meadows on system leverage points (where in this taxonomy is the highest leverage); KI Norge mandate documents from Digdir; recent NTNU/SimulaMet hiring and capacity reports.
• uninformedStrümke is open about her own production-coding skepticism ('jeg merker når jeg sitter og koder med agentene, så er sånn tulleprosjektet'), and the hosts (Audun from NAV) push back gently with the more developed practitioner view that high-stakes work warrants Vim while front-end can be agent-led. Strümke's reading of agentic coding may underweight what's currently possible with proper harness, evals, and sandboxing — though she does flag the macOS sandbox feature.
  Read next: User's existing thread on harness engineering (Anthropic videos, Boris Cherny, Bockeler/Fowler static sensors); Beane on Precision Proactivity; the user's earlier Heis Anders Sveen / Vilde Opsal extractions; Greger Wedel; concrete production-engineering reports from Builder.io, Cursor, and Claude Code teams.
• incompleteThe closing 'undetectable backdoors' research preview is fascinating but isolated from the rest of the conversation. If undetectable backdoors are plantable, the entire 'outsource problem-solving to Copilot' diagnosis acquires a much sharper national-security dimension that the conversation does not draw.
  Read next: Goldwasser et al. on undetectable backdoors (the foundational paper); recent NeurIPS / ICML papers on neural-net backdoor detection; supply-chain security literature (SLSA, sigstore) extended to ML model artifacts.

Closes gaps elsewhere

• Skill Issue: Andrej Karpathy on Code Agents, AutoResearch, and the Loopy Era of AI · g5 — Provides the critical perspective on agent dependency the gap seeks: learning requires effortful practice, so agent delegation erodes junior developers' tacit knowledge and destroys the judgment needed to delegate well.
• Radar Trends to Watch: May 2026 · g1 — Answers the game-theoretic objection to unilateral restriction: equivalent zero-day-finding capability already exists via swarms of small open models, so vendor self-restraint is structurally insufficient in a competitive market.

Related documents

• Agents don't know what good looks like. And that's exactly the problem.
• 7) Stine Haugseth: Insourcing, retning og strategi på en ølbrikke
  Data Sovereignty Insourcing & In-House Capability Public-Sector Digitalization
• The best response to AI slop, infinite advice, and online noise is from Robin Williams
  Cognitive Surrender LLM Limitations
• AI Is Undermining Leaders’ Judgment. Here’s What to Do About It.
  Cognitive Surrender Explainable AI
• All of AI's New Models and Tools
  AI Alignment & Safety Mythos
• Anthropic Can Now Read Claude's Mind
  AI Alignment & Safety Explainable AI
• A global workspace in language models
  AI Alignment & Safety Explainable AI
• Assessing Claude Mythos Preview's cybersecurity capabilities
  AI Alignment & Safety Mythos
• Fragments: February 4
  Cognitive Surrender LLM Limitations
• How governments and organizations are leveraging Google's AI breakthroughs for crisis resilience
  Data Sovereignty Public-Sector Digitalization