Skip to content
NB EN
Nettsak

From AI to AGI: status, choices and timelines

What is AI - and how do LLMs work today? Modern language models (LLMs) are transformer-based predictors that calculate the probability of the next token (word/word fragment) given a co...

Håkon Berntsen 6 min read
From AI to AGI: status, choices and timelines

What is AI - and how do LLMs work today?

Modern language models (LLMs) are transformer-based predictors that calculate the probability of the next token (word/word fragment) given a context. The transformer architecture replaced recurrent networks by using self-attention, which delivered both better quality and powerful training efficiency (Vaswani et al., 2017). arXiv+1

Scaling laws have governed the progress: DeepMind's "Chinchilla" showed that for compute-optimal training, model size and the number of training tokens should be scaled roughly equally; many models were previously "undertrained" relative to their size. arXiv+1

In practice, 2025 models combine:

  • Multimodality (text + image/audio/video) with long context windows (hundreds of thousands of tokens in some systems).
  • Efficient serving: e.g. vLLM/PagedAttention, which reduces memory waste in the KV cache and increases throughput by 2-4x in production. arXiv+1

Limitations remain: a lack of persistent memory between sessions, handling of uncertainty, causal understanding of the world and robust planning outside the training distribution.

What do we mean by AGI?

Artificial General Intelligence (AGI) refers to systems that can learn, reason and act broadly across domains - at roughly human level or better. The very idea of "machines that can surpass humans" is old: Turing discussed the test for machine "thinking" (1950), and I. J. Good outlined "the first ultraintelligent machine" that could improve itself (1965/66). courses.cs.umbc.edu+2incompleteideas.net+2

The term AGI in its modern sense was popularised in the early 2000s (Goertzel, Legg et al.), although some point to use going back to the 1990s; the consensus is that AGI became established as terminology around 2002-2007 and was cemented through dedicated conferences and academic texts. Wikipedia+1

The difference from today's LLMs is not only size: AGI presupposes architectures that connect language/multimodal understanding with persistent memory, tool use, perception and action loops, as well as goal-directed learning and safe self-improvement - under governance mechanisms that can be documented.

How can AI take the step towards AGI? (technical building blocks)

  1. Memory + knowledge retrieval: from "in-context learning" to genuine long-term memory (episodic/semantic) with evaluated update regimes.
  2. Multimodal "world models": integrating senses (image/audio/video/biomedical sensors), action and causal prediction.
  3. Agent architecture: planning, tool calls, collaboration between sub-agents and explicit uncertainty management.
  4. Safety and governance: WHO has >40 recommendations for large multimodal models in health; the EU AI Act introduces risk-based obligations, sandboxes and bans on certain practices. This makes governance just as important as compute. Reuters+4World Health Organization+4World Health Organization+4

When can we expect AGI? (timelines with uncertainty)

Large surveys among AI researchers (AI Impacts, 2023-2024) estimate roughly a 50% probability of HLMI/AGI around 2047, with a very wide spread; 10% as early as 2027 and many who believe it will be significantly later. The results show both accelerated expectations from 2022 to 2023 and a high degree of disagreement. AI Impacts+2arXiv+2
Popular-science overviews and media coverage point to the same picture of uncertainty and to diverging academic camps. Our World in Data+1

Editorial assessment: Given the current methodological frontier (transformers + scale + work on memory/agentics), the 2030s appear to be a likely period for partially general systems in limited domains, while the 2040s is a cautious midpoint for more broadly applicable, regulated "AGI-like" systems - conditional on safety, regulation and infrastructure being in place.

Where does the frontier stand internationally?

  • Architecture and operation: transformer derivatives, long contexts, multimodality; serving with vLLM/PagedAttention and hardware optimisation (e.g. FlashAttention-3) that makes production economically viable. arXiv+1
  • Safe AI: directions such as "constitutional AI" and frontier-risk programmes at leading labs; regulatory implementation in the EU is moving towards sandboxes and sector rules. Artificial Intelligence Act+1

What is happening in Norway (and close to Norway) right now?

MediVox (Norway) - local models for clinical documentation

What they do: MediVox delivers AI-supported transcription and draft medical records for healthcare professionals using local language models operated in Sandefjord (stated by the company in previous communication with the editorial team and in public presentations). The solution is aimed at Norwegian clinical practice, documentation and compliance (GDPR/ISO).
Who is behind it: Co-founded and led by Norwegian developers and clinicians; Håkon Berntsen is chief operating officer (COO) and co-founder.

Note: For clinical claims and customer cases, reference is made to public announcements/contracts; health-related functionality should always be assessed against WHO recommendations and the EU AI Act when the use is "high risk". World Health Organization+1

EIR Tec Ltd (United Kingdom) - EEG-based brain readings and home diagnostics

What they do: EIR Tec describes a platform for home-based EEG and AI-supported analysis aimed at mental health/neurology (including ADHD/ADD-related assessments). Official website: eirtech.co.uk. EIR TEC
Who is behind it: A technology community with Norwegian founder ties; Håkon Berntsen is CTO/co-founder (stated to the editorial team).
Editorial clarification: A diagnosis in the legal/medical sense presupposes regulatory approvals (class, indication, market). EIR Tec communicates ambitions of diagnostic support; readers should distinguish between a clinical support tool and a formal diagnosis.

ReadySOFT / ReadyPOS (Norway) - AI-driven POS for SMBs

What they do: ReadySOFT is building a .NET MAUI-based POS system (Android/Sunmi) with AI modules for price optimisation, menu/assortment, campaign automation and predictive operations. Technical material and business pitch state local LLM infrastructure in ReadySOFT Cloud Sandefjord, Zero Data Retention when using external LLM APIs, and integrations with Tripletex, PowerOffice and Fiken.
Who is behind it: The team includes Per Joar Lorvik (CEO), Øyvind Horn (CFO), Helge Andresen (CTO/adviser), as well as Håkon Berntsen (AI/architecture, adviser). (Source: investor/product documents submitted to the editorial team).

In brief (editorial)

  • AI today: transformer models (LLM/LMM) that are extremely capable, but with clear limits - especially memory, robust planning and causal understanding. arXiv
  • AGI as a goal: requires systems with persistent memory, a multimodal perception/action loop, agentic planning and documented safety. Historically rooted from Turing (1950) via Good (1965) to the AGI concept (2000s). courses.cs.umbc.edu+2incompleteideas.net+2
  • Timeline: the academic community's median points to around 2047 for HLMI/AGI, but the spread is large; several milestones may arrive significantly earlier. AI Impacts
  • Norway now: three different tracks - clinical documentation on local models (MediVox), brain-signal analysis at home (EIR Tec) and AI-driven POS on Norwegian infrastructure (ReadySOFT) - show a practical path from today's AI towards more general, safe and regulated intelligence.

Sources (selection)

H
Håkon Berntsen

Related stories