One year ChatGPT - is the Honaimoon over?

Exactly one year after blowing away all of us with opening the GPT large language model (LLM) to the public, OpenAI seems in tectonic shifts. Just one week after the first big developer conference, the access to the "plus" services was restricted, and surprisingly co-found Sam Altman was forced out of the company. There are many rumors why this could have happened (see an exhaustive overview here and the timeline of events here), but we will know for sure in the next days.

Before we dive into this week's edition, let's talk first about words to make sure we have the same understanding. I asked GPT4 to define the most important ones:

• Machine Learning (ML): A subset of AI that involves the development of algorithms that allow computers to learn and make predictions or decisions based on data, rather than following only explicitly programmed instructions.
• Deep Learning: A subset of machine learning involving neural networks with many layers (deep neural networks). It's particularly powerful for handling large amounts of unstructured data like images and text.
• Transformer Model: A type of deep learning model that's particularly effective for tasks involving natural language understanding and generation. It uses mechanisms called attention and self-attention to weigh the significance of different parts of the input data.
• LLM (Large Language Model): A type of artificial intelligence model designed to understand, generate, and respond to human language. These models are 'large' due to their substantial size, often containing billions of parameters, allowing them to process and generate complex and nuanced text.
• GenAI (Generative Artificial Intelligence): This term refers to AI systems that can generate new content, whether it be text, images, music, or other forms of media. These systems don't just analyze data but can create entirely new outputs based on learned patterns and examples.
• GPT (Generative Pre-trained Transformer): A series of language processing AI models developed by OpenAI, designed for tasks like translation, question-answering, and text generation. Each version of GPT is more advanced than the last, with GPT-4 being one of the latest and most sophisticated.

So while the press and public opinion was focussing a lot on the so called "generative AI" and their chatbot interfaces, the research and industry communities have a much broader focus on development and implementation of "AI".

Do you really know ChatGPT?

I'm pretty sure that one year ago, you tried out the OpenAI webpage and chatted with #ChatGPT. Asked some questions, let your children tell jokes (and the homework for school). You were surprised about the ease of that communication, and you were not surprised that it started to hallucinate, told wrong "facts" about your CV and come pretty soon to its limits when you tried it for business jobs.

As with any new toy, ChatGPT ended up in a corner for many. And the verdict has been reached. But is it still true?

But the development of those tools is amazingly fast. If you would retry today with #GPT-4 (for which you will need a paid "plus" account, and perhaps you can't activate a new one at the moment), many things changed since then. It broadly enhanced the linguistic capabilities, and added some majors features:

• Complex Problem-Solving Abilities: GPT-4 can synthesize information from multiple sources, providing comprehensive and nuanced answers to complex questions and properly citing sources when generating text. It can solve mathematical and scientific problems, extending to advanced calculus and scientific concepts analysis.
• Image and Graphics Understanding: GPT-4 can analyze and comment on images and graphics, which is a step beyond the text-focused capabilities of its predecessor. It can interact with and generate content that includes pictures and other media types.
• Use of Plugins: GPT-4 can integrate with external tools and services through plugins, enhancing its functionality and allowing it to perform a wider range of tasks.
• Web Browsing Capability: It can search the web to pull in information from the internet, offering more accurate, up-to-date information and providing a broader range of data sources.

Especially the last one - including a web search into a query - is game-changing for the private use. It combines the world of web search with the power of a fully natural language-interface to your digital device.

The use of plug-ins, instead, is a game-changer for industry business. It links generative AI to the world of business processes, document analysis, systems engineering and so on.

Don't forget: LLM's are underlying fundamental constraints due to statistics and training, and they are no "AGI" (generic AI) yet.

Bing with AI - the worst User Experience since Clippy

The integration of all these new capabilities into something productive, adding real value to your private or business life, is however tricky. If you are using the bing.com website for search, you can experience why.

Microsoft tries to offer you an "AI co-pilot for the web", but the User Interface is so incredibly bad that it reminds me of Clippy, the most annoying bot ever ("Karl Klammer" in Germany).

The integration of AI-powered assistance, with the right level of pro-activeness (offering more), factfulness (showing references) and user control (the user controlling the tool, not vice-versa) is not easy. But 微软 is already improving, and hopefully they will get the integration into the complete Office365 suite right.

Will be AI bots your new colleagues?

The example of Clippy shows that it's not the great idea or potential of an innovation - it's also about implementation, understanding the use cases / users and the day-to-day challenges, or getting the algorithm so mature that it really can take over a business process on its own.

Implementing AI in your business is a marathon of craftsmanship: many small and continuous steps of adaption and improvement.

We at 埃森哲 are investing heavily in (Generative) AI, with a massive Center of Competence on the training and deployment of AI algorithms in industry environments. The first, most obvious use cases to take AI bots to run the customer interface are already state-of-the-art and we are widely in implementation projects. This is especially interesting when the "device" is not just a website, but for example a vehicle - how could we bring an infotainment screen inside a car on steroids by ChatGPT & Co?

But this is only the obvious use case. If we look on the complete value chain of development, systems engineering, manufacturing, supply chains and so on, you will see a lot of experts taking pieces of information from one process and putting into another, following process. Or taking decisions in between, which are based on a (rich) set of data. Many of those decisions and processes are great candidates for AI support, or even AI automation. We are running a huge experimental framework to identify and implement such use cases, and scale them to all of our customers in a fast way.

Reach out to me and my colleagues to understand how our solutions can speed you up on AI: Kathrin Schwan Max Haberstroh Vlad Larichev Benno Stützel Jens Frühling and many more.

Who's in the race?

As you know, OpenAI (with Microsoft as biggest promotor) set the pace - even if the transformer model was introduced with the help of Google in 2015. And probably OpenAI still has the most advanced model, but that's not all that counts.

Interestingly, one big player is missing: 苹果 . #Siri hasn't made any major step ahead in the past years, but it is expected that iOS 18 (next year's release) will (or better: has to) deliver a significant step. And Apple will focus for sure on a fundamental integration into the operation of its devices. Wait and see...

Mostly based on Meta's Llama model, another significant competitor is creating stress for the big ones: the open-source community. Looking at the massiv invest into data gathering, training and fine-tuning of the models, a convergence of cost vs. benefits is foreseeable. Open-source models could take away the much-needed repayment of the first movers which are losing currently hundreds of millions each year.

And Europe?

Prof. Bernhard Sch?lkopf is among the most important AI researchers in Europe and heading the Cyber Valley at #Tübingen. He formulated an action plan which was published a few weeks ago in the Frankfurter Allgemeine Zeitung (behind a paywall). Here are his top recommendations:

• Adapt to the AI revolution by embracing it as a tool for competitive advantage and economic success.
• Shape AI development actively within Europe to ensure economic success and alignment with European values.
• Regulate AI to address potential negative impacts without stifling innovation and competitiveness.
• Foster a European AI ecosystem that is compatible with the values of a liberal European democracy.
• Prioritize the creation of substantial, sustainable computing infrastructure and curated training datasets to ensure technological sovereignty.

With the ELLIS initiative, a European platform has been created to promote young AI talents. More than 1000 entities are part of the network, and a first dedicated institute has been created at the Cyber Valley Tübingen.

The website AI Watch of the European Commission gives a structured dashboard on all entities, companies and investments into AI within the European Union. There is more than expected - but still a long way to go.

What are your experiences with AI so far?

I'd like to hear from you - where are you? #Honeymoon, #divorce or #lifelong partnership with #AI? Happy to read your comments!

Dennis R?hr?Marcel Friebel?Fran?ois BARTHET?Jakub Borowczyk?Timo Dieringer?Deepti Bhutani?Celil Ayd?n?Christian Neidel?Dr. Marco Langerwisch?Juergen Reers?Hans Loes?Philip Dorn?Daniel Tatomir?Modar Horani?Amarnath Bharadwaj?Cristian Cuna?Martin Thiede?Wiktor Walczak?Joris Bauer-Ludwigs?Jan Evers?Verena Aulenbacher?Raffaele?Menolascino?Wolfgang K?cher?Gerd Schaefer?Liam Friel?Dominic Craciunescu?Dule Petkovic?Milica Ruvidic?Hendrik Dettmering?Benno Stützel?Bettina Blum?Dharanisha Siddalingaiah?Marcus Hammes?Jonas Kampik?Darani Yogalingam