FORUM IP Talks: Sebastian Goebel on advantages of local Generative AI in patents
Jean-Claude Alexandre Ho, LL.M.
IP Knowledge Enhancer, I help people to increase their IP knowledge and skills | Developer and Organiser of mostly IP training (also mining law and gambling law) | IP Manager | Lawyer
Jean-Claude: Sebastian, it seems as if locally-run generative artificial intelligence is on the verge of a breakthrough since late 2024. For those of the readers not that well into generative AI: What are the particularities of locally-run language models?
Sebastian: Local models are basically the same as cloud models like ChatGPT, but smaller. The key distinction of locally-run language models is that they operate entirely on your own hardware - whether that's your computer or a local server - rather than in the cloud. As a patent attorney, I see the main benefit without doubt in the processing of confidential data. All data processing happens on your device, which is crucial for the use for confidential patent drafts and analyses. This makes the use of generative AI (GenAI) for our work simply much more carefree.
What many don't know: Even the technologically highly praised but for data protection criticized cloud model “DeepSeek R1” can be used locally without any concerns. Because the data remains on your own computer and is not sent over the internet. This opens a whole new level of language processing quality for local applications, which until now has only been available in expensive models.
Another significant benefit is the better cost-efficiency of local models. Additionally, you maintain full control: You can modify and customize the model for specific patent-related tasks.?
Where do we stand at the moment with the most powerful locally-run language models?
The landscape has evolved significantly in 2024. DeepSeek's R1 demonstrating that locally-run models can now compete with cloud-based solutions in specific features like reasoning. LLaMA 3.1's release with 405B parameters showing that large-scale models can be optimized for local running. This is all very promising for use in patent law firms and patent departments.
On the other hand, cloud models are also getting better, and it would be a mistake to rely on just one solution. Furthermore, there are now additional regulatory aspects. The EU AI Act's implications for local models, especially in commercial settings, need careful consideration.
What is your take on xLMST (Long Short-Term Memory)? Sepp Hochreiter, one of its creators, recently talked about the third phase of AI. Which potential do you see for patent work?
As a patent attorney actively working with AI technologies, I find xLSTM particularly intriguing, especially in the context of what Hochreiter calls the "third phase" of AI. Let me share my perspective on its potential for patent work. What particularly interests me from my patent practice perspective is xLSTM's claimed efficiency advantages - faster processing, better memory usage, and linear runtime. This aligns perfectly with the needs of patent practitioners. If these claims are true, xLSTM could enable more sophisticated AI tools to run locally, addressing the confidentiality concerns I highlighted earlier. However, I remain cautiously optimistic. While xLSTM shows great promise, we need to see how it performs with larger datasets and in real-world applications. Without doubts, we are now transitioning into the third phase, the industrial application of AI, where AI methods are being adapted to real-world applications in robotics, life and earth sciences, engineering, and large-scale simulations – and patent work of course.
"The golden rule is to use local models to process information that is already available in the context."
Which are the best use cases for locally-run generative artificial intelligence in the field of patent drafting?
The obvious application is certainly in the drafting of the description, and to a limited extent also in the drafting of claims. But the use of GenAI in the evaluation of patent documents or the querying of invention disclosures is at least as important. However, most of these activities require a high level of confidentiality.
The advantage of using locally executed models is that you usually don't have to worry about the confidentiality of the data. Particularly when querying content from documents, sometimes marketed as ‘ask-your-data’, local models result in great efficiency gains and hardly any restrictions compared to cloud solutions. Local models are particularly useful when the knowledge is provided by the user as context and therefore no knowledge base needs to be present in the model itself.
When creating the description, the efficiency gains are slightly lower when using only local models. This is because often I want the model to define technical terms, describe technical relationships, and explain advantages and effects. This already requires a lot of ‘knowledge’ (i.e., training with extensive training data) that small local models simply don't have. If knowledge is queried that is not in the training data, there is also a higher risk of hallucinations, i. e. false facts output.
The golden rule is to use local models to process information that is already available in the context (e.g. the invention disclosure). For everything else, use larger models, but ideally only cloud-based for non-secret information. And in any case, every GenAI output must always be strictly rechecked.
In my workshop, I therefore explain how to address these issues and discuss how local models can be used more efficiently by automation and combined with cloud models and special prompting techniques, while of course maintaining confidentiality.
Which are the best use cases for locally-run generative artificial intelligence in the field of patent portfolio management?
In patent portfolio management, there is a much larger number of activities that do not require confidentiality. Consider, for example, a landscape or patent portfolio analysis of granted and thus already published patents. Powerful cloud models can be used here without concern. I recommend focusing on cloud solutions that not only use a powerful large language model (LLM) but also provide the LLM with a handful of tools to perform internet searches or create impressive visualisations, for example. It should also be noted that language models per se are not initially able to evaluate large amounts of data or to calculate statistical analyses. Advanced models are therefore integrated into a framework in which they use tools or generate code, which then analyses the data. These are, of course, all advanced techniques that require targeted user training. It will probably take some time before GenAI can be operated completely intuitively without AI or prompt engineering knowledge and still deliver reliable results.
The European Patent Institute has recently adopted guidelines for the use of generative AI. What are the advantages of locally-run generative AI?
The advantages of locally-run generative AI are particularly compelling when viewed through the lens of the new epi Guidelines. The primary benefit is the enhanced confidentiality it offers - a crucial consideration given Guidelines 2a and 2b's emphasis on data confidentiality. When running AI locally, sensitive information never leaves your own infrastructure, giving us much better control over data security. This directly addresses one of the main concerns about confidentiality of training datasets and prompts that the Guidelines highlight.
Which are the most important consequences you see for patent attorneys arising from the epi Guidelines?
The epi Guidelines define something that should be self-evident for patent attorneys: they are expected to handle their clients' data responsibly. However, GenAI initiated a fundamental shift in how we must approach the integration of technology in our practice, which is acknowledged by the Guidelines as well. In case we want to use GenAI, we're now required to actively inform ourselves about AI models' characteristics, we must understand how language models work, while remaining fully responsible for any AI-generated work. This means implementing rigorous checking processes and extensive AI training for users.
In essence, these Guidelines require us to treat AI as a professional tool that demands the same level of diligence and ethical consideration as any other aspect of our practice. This aligns perfectly with what I discussed in our last interview about the importance of understanding AI technology to use it effectively and responsibly.
Sebastian, thank you very?much for the interview!
About the interviewee:
Sebastian Goebel is a European and German Patent Attorney and a founding partner of B?sherz Goebel, a decentralised, highly digitalised boutique patent law firm with patent experts throughout Germany. His professional focus is on the protection of digital technologies. Sebastian is passionate about AI and software development, both now and through his previous work as an electrical and computer engineer. Other areas of his expertise include medical engineering, electrical engineering, electronics, and signal processing. He is also a co-chair of the I3PM's IP & AI Committee.
Sebastian is going to speak at: