For companies, this is more than just a technical nuance. The limitations of LLMs directly impact how AI can be utilized in medium-sized enterprises, what decisions make sense, and where specialized AI services become necessary. HighPots supports companies precisely at this interface: between language-driven AI, real business logic, and applications that reliably deliver results.
LLMs are powerful – but they remain language models.
Current large language models are primarily based on the transformer architecture. They operate on discrete symbols extracted from language or images. This allows for high performance in text processing, analysis, automation, and AI agents for businesses.
But this very strength also marks the boundary:
- LLMs do not learn real physical states.
- They do not model dynamics, causality, physics.
- They generate plausible answers – but do not create world models.
AI pioneer Yann LeCun succinctly summarizes:
LLMs are “wordsmiths in the dark” – they operate in the dark, without understanding what happens in the real space.
Even alternative architectures like state-space models (SSM) or RWKV still do not address the core problem. They are efficient sequence models but remain language- or token-based, not world-based.
For companies, this means:
Limits of LLMs: An LLM can interpret your data, but cannot comprehend physical reality.
It does not know forces, objects in space, or the consequences of an action – apart from what is described in text.
Why language models cannot process real-time sensor data.
For applications related to the physical world – production facilities, logistics, navigation, automation – AI services based on an LLM are not sufficient.
There are four technical reasons for this:
1. The data formats do not match
Sensor streams and video data are continuous, high-dimensional, and time-critical.
However, LLMs:
- process tokens instead of raw signals,
- lose continuous transitions,
- and cannot represent spatial structures.
A production robot does not operate in tokens. It works in Newton meters, pixel clouds, distances, forces, and rotational speeds.
2. There is a lack of a world model
LLMs do not have an internal representation of:
- space
- time
- dynamics
- cause and effect
They cannot simulate how a machine behaves, how an object falls, or how a sensor value will evolve in the next milliseconds.
Without such a world model, every control task remains uncertain.
3. Latency and architecture
Real-time applications allow for very few milliseconds of delay.
On the other hand, LLMs require large computational blocks because the entire sequence is processed – even optimized variants remain comparatively sluggish.
For control systems, it holds:
Too late is synonymous with wrong.
4. Multimodality ≠ understanding of the world
Multimodal models such as GPT-4o or Gemini can analyze videos – but these data are also internally broken down into tokens. This is pattern recognition, not true physics simulation.
What companies should derive from these limitations of LLMs
For decision-makers, it is crucial:
For processes that need to represent or control real-world conditions, more than an LLM is required.
However, LLMs remain extremely valuable in companies – when used correctly:
- for analysis, research, and knowledge automation
- for AI agents that structure internal processes
- for RAG systems that leverage corporate knowledge
- for automating communication, documentation, support, quality assurance
- as an interface between humans, systems, and business logic
The key lies in using LLMs where language and decisions dominate – and specialized models where physics, dynamics, or real-time are required.
Why World Models will be the next major development
For true machine intelligence, systems need models that:
- understand continuous state spaces,
- map dynamics over time,
- learn causality,
- be able to simulate scenarios.
Such World Models currently exist only in early research stages. Nevertheless, it is foreseeable that they will become the foundation for future autonomous systems – from robotics to manufacturing to transportation.
However, large technology companies primarily rely on LLMs because they scale commercially faster. This creates a gap where Europe can leverage technology.
An opportunity: rethinking architecture, fundamental research, and industrial applications.
HighPots is already accompanying this development today: with modern AI services, RAG-AI, agent-based systems, and solutions that reliably model real business processes.
What HighPots can already offer companies today
As a European AI service provider focused on feasibility, HighPots combines three competencies:
1. AI architecture and RAG systems for data-driven decisions
We develop systems that securely utilize your company knowledge – On-Premises or in the Cloud.
2. AI agents for operational processes
We implement agent-based systems for support, documentation, processing, and automation – with stable interfaces into existing IT landscapes.
3. Software development & middleware for real business operations
Whether SAP Concur, WooCommerce, WordPress, or custom backend systems: We integrate AI directly into your existing processes and systems.
This clearly positions HighPots:
as a partner for companies that want to make AI truly productive.
Limits of LLMs: How companies can sensibly implement AI
A successful AI project does not follow trends, but the requirements of the business:
- What tasks are language-based? → LLMs + RAG + AI agents.
- What tasks are based on measurement data or physical interaction? → Specialized models & traditional ML methods.
- What systems need to be integrated? → Middleware & API engineering.
- Where are data protection, security, and availability critical? → On-prem solutions.
HighPots supports medium-sized companies in each of these phases – from concept to operation.
