LLMs won’t take us to AGI – why lack of continuous learning matters
A recent post on Reddit argues that today’s large language models (LLMs) won’t get us to artificial general intelligence (AGI) because they don’t learn continuously after training. The poster links to a new paper that frames the gap in cognitive terms and points to “world models” and autonomous, experience-driven learning as the missing pieces.
You can read the discussion here: LLMs won’t take us to AGI and this paper explains, and the paper referenced is on arXiv: https://arxiv.org/abs/2603.15381.
The core claim: LLMs are powerful, but they don’t learn after training
The post’s central point is straightforward: once an LLM is trained, its parameters are effectively frozen. Prompts, fine-tuning, and RAG can help it behave better in context, but they don’t make the model itself acquire new knowledge autonomously from experience.
“They don’t update themselves from real world experience… They don’t have autonomous continuous learning.”
Quick definitions to keep us aligned:
- LLM (large language model) – a neural network, usually a transformer, trained to predict the next token in text.
- Fine-tuning – additional training on curated data to specialise a model for a task or tone.
- RAG (retrieval-augmented generation) – connecting a model to an external knowledge base so it can “look up” facts at generation time.
- Context window – the span of text the model can “see” when generating a response.
These techniques improve outputs without changing how the model fundamentally learns. The Reddit author argues that this is why LLMs can appear brilliant one moment and brittle the next: they match patterns seen in training rather than building an evolving, grounded understanding of the world.
Continuous learning and world models: what the paper argues
According to the post, the paper ties current model limits to cognitive science: intelligence emerges from systems that acquire knowledge through ongoing interaction, not just by static next-token prediction. That’s where “world models” come in – internal models that represent how the world works, so an agent can plan, experiment, and update its beliefs over time.
The poster notes that Yann LeCun is involved in this line of work and mentions significant funding around world models. The funding detail is not disclosed in the paper and is not independently verified here, but the broader point stands: there is growing interest in architectures that go beyond passive text prediction.
Are we hitting a scaling wall with LLMs?
The post’s stance is that scaling LLMs will take us far, but not all the way to AGI. Two realities can be true at once:
- Benefits of scale – bigger models with more compute and data keep getting better at language, coding patterns, and multi-step reasoning when scaffolded by prompts and tools.
- Limits of the paradigm – without mechanisms to update themselves from interaction, models remain pattern learners. They rely on static weights and contextual hints rather than forming persistent, grounded knowledge that improves through experience.
In other words, better prompts and larger context windows help “steer” the fixed system. But they’re not the same as a system that can autonomously learn from its own actions, successes, and mistakes.
Implications for UK developers, teams, and decision-makers
Whether you agree with the AGI prognosis or not, the near-term takeaways are pragmatic:
- Plan for model brittleness – LLMs can excel at familiar patterns yet fail unpredictably out of distribution. Add verification, guardrails, and human oversight for high-stakes use.
- Favour retrieval over risky “learning” – use RAG to bring in current, authoritative information rather than relying on a model’s latent memory.
- Keep compliance front and centre – for UK organisations operating under GDPR and sector rules (finance, health, public sector), ensure data minimisation, auditability, and clear data flows. Avoid sending personal or commercially sensitive data to services that can’t guarantee UK/EU processing where required.
- Evaluate total cost of ownership – LLMs remain compute-hungry. Budget for inference costs, latency, and monitoring rather than assuming “free learning” over time.
- Prefer repeatable pipelines over ad hoc prompting – design workflows that are testable, observable, and robust to model updates.
Practical approach: get value today, design for tomorrow
What to build with current LLMs
- Task-specific assistants with RAG – keep knowledge external and up to date; log citations; show sources.
- Tool-augmented agents – call APIs or databases for calculations and actions, but gate critical steps with rules and approvals.
- Evaluation and feedback loops – capture failures and iterate prompts, data, or fine-tunes offline. Don’t expect the model to self-correct over time without retraining.
If you’re integrating LLMs into everyday tooling, here’s a practical guide to streamline common tasks: How to connect ChatGPT and Google Sheets.
What to watch as the field evolves
- Memory and autonomy – research that adds persistent memory, planning, and safe forms of online adaptation.
- World models – agents that build internal representations of environments and update them through interaction.
- Safety and governance – methods that allow learning from experience without compromising reliability, privacy, or legal compliance.
A balanced view: are LLMs enough if we just keep scaling?
The Reddit post says no: without autonomous continuous learning, LLMs will hit a ceiling. A more optimistic counterpoint is that clever scaffolding (tools, retrieval, memory, simulators) might extend the current paradigm much further than expected.
Both positions agree on one thing: today’s models are not self-improving agents. If you need systems that learn from your environment in real time, you’ll need explicit feedback and retraining processes, not just bigger prompts.
Why this matters
“Scaling LLMs will take us far but not all the way.”
For UK organisations, the message is practical rather than speculative. LLMs are already valuable for productivity, drafting, coding assistance, and structured data tasks. Treat them as highly capable pattern models augmented by retrieval and tools, not as autonomous learners. Build for auditability and compliance, and keep humans in the loop where it counts.
If the paper’s thesis holds, the next breakthroughs will pair strong language capabilities with systems that learn continuously from the world. Until then, the winning strategy is clear: deliver value with today’s models, design your stack so it can absorb tomorrow’s.
Sources and further reading
- Reddit discussion: LLMs won’t take us to AGI and this paper explains why
- Paper cited in the post: arXiv:2603.15381
