Is the LLM winter coming?
Reports of delays and problems are piling up in the AI world. The big names in the industry - Amazon, Meta, OpenAI, Apple - are postponing their ambitious projects. Is this the beginning of the end of the AI boom? It's worth taking a closer look.
Amazon's ambitious plans for a generative AI upgrade of Alexa have been repeatedly postponed. The reason? Reliability problems - even for simple queries, the new version often provides incorrect answers. Amazon has wisely decided to prioritise quality over speed, which is now jeopardising the launch originally planned for 2025.
Meta AI's path to Europe proved difficult. After months of delays due to data protection concerns, the text-based chat was finally launched on 19 March - albeit with significant restrictions. The multimodal functions available in the USA, such as image and video generation, are completely missing. It remains to be seen when European users will receive the same functionality.
We see similar patterns with OpenAI,who had to postpone its promising ‘Operator’ by a few weeks at the beginning of the year due to safety concerns.
And very prominent Apple,where the integration of AI functions in Siri has even been postponed until 2026, which has actually earned Apple a class action lawsuit in the US for misleading advertisingas these features have been prominently advertised in commercials since September 2024. Quality and safety problems are also the reason here; it is suspected that a lack of robustness against prompt injections plays an important role.
So the reasons are always the same: quality and safety. And the standard answer of recent years, the scaling of models, apparently no longer brings any corresponding improvements in terms of quality and safety.
The end of a hype?
Do these problems mean the end of the AI hype? Are we facing an ‘LLM winter’, similar to the AI winters of past decades?
I don't think so. The ability of modern AI systems to communicate in natural language remains revolutionary. It is exactly the piece of the puzzle that the development of truly intelligent applications has always been missing.
However, we need to put this ability into perspective: Applications that are supposed to be able to do more than interpret complex language and formulate eloquently also need other skills, for example:
- Reproduce facts correctly
- Distinguishing company-specific information from general knowledge
- Active conversation, correct conversation behaviour
- Resistance to manipulation attempts, e.g. through prompt injections
We find approaches for many of these capabilities in the LLMs. These approaches arise quasi incidentally and statistically from the enormous size of the models - but are not their core competences.
The path to truly productive AI applications
We want to develop sophisticated applications - AI that researches for us, navigates the web, fills out forms and makes bookings. AI that really does the work for us, automates processes and talks to customers. AI on our smartphone that gives us comprehensive advice and understands our personal context, but also treats it with care. Applications like these are necessary to turn AI from an eternal experiment into a truly productive economic factor.
But the current challenges show: We need more than just a strong language model. Where exactly we need to upgrade depends very much on our application and the associated expectations. For example, do we want an interlocutor who responds to the situation and also keeps the company's interests in mind, or do we actually expect deterministic behaviour as we know it from programmed software - with a natural language interface?
The good news is that the delays at the major providers demonstrate that even companies with their own AI models are struggling with similar challenges. They are not fundamentally privileged over other market players when it comes to developing productive applications. For them too, LLM is ultimately a black box. Quality and security do not come about by themselves, but require careful development and safeguarding.
Conclusion: A phase of consolidation
What we are currently experiencing is not an LLM winter, but a necessary phase of consolidation and transition. From exaggerated expectations to more realistic assessments of what this technology can already do today and where we still need to do development work.
The communications revolution of LLMs is here to stay. What is changing is our understanding of how we can transform them into reliable, secure and truly useful applications.
Klaus Reichenberger
