Dirk Roeckmann

We should have noticed by now, that the scaling laws of #GenAI fail and we’ll enter a phase of diminishing returns. Equally problematic is the uncritical re-emergence of failed #symbolic paradigms which brought as #AI winter 1. #Neurosymbolic - yes but not based on failing monotonic reasoning

I have a hunch that #AI research will go beyond glorified matmul and #GenAI as we know it including its hardware will be toast. #LLM will be smaller, will run on turing-complete #CPU again and serve as what they were designed for - a great natural language interface and bridge to symbolic reasoning

“It doesn't make sense to hire smart people and then tell them what to do; we hire smart people so they can tell us what to do”
Steve Jobs

Think about that when you hire consultants 🧐
#SAP #ABAP

I can’t believe how much money is put into such hyper-specialized hardware like a #GPU which is not even turing-complete and can only do parallel matmul. This will all be immediate trash once #AI goes beyond tensor-based paradigms because it can’t be generally repurposed like a #CPU. #SAP #ABAP

In order for #neurosymbolic #AI to succeed we need a renaissance of #homoiconicity. Bring back #Lisp as 2nd lingua franca for #AI next to #Python and perfect their interoperability. IMHO the ideal scenario would be to host a performant Lisp on #Mojo! #Modular @Chris Lattner

Hi all - are #UI #annotations anachronistic in the age of #agentic #AI? Should only the final code be generated and not some intermediate code which has still to
go through the annotation engine? Let me know what you think #SAP #RAP #OData #CAP

In consumer web search and commerce the browser will become a module of #AI agents, #UI will be ephemeral i.e rendered on the fly. In enterprise this translates only partially. #GenAI is not intelligent but transformative. UI annotated code generation is an anchronism. #SAP #ABAP

Here is some common-sense and sad truth: Most customers fail to pen a consistent specification of what they need in natural language to human #developers - what makes deciders think that these very same people could instruct a #codingagent and then verify the results on top of that!? 🧐 #ABAP

There is no need for the #nocode, #locode and #procode silos anymore. No matter whether you #vibecode, use a lo-code interface or code manually in an IDE/editor: The resulting code should always be fully accessible for further optimization and not be hidden behind an opaque UI. #SAP #ABAP

While #SAP touts that it is advocating #OpenSource and being open in general, many recent actions suggest otherwise:
- OpenSAP courses : gone
- CAP 9 : only with BTP license
- OpenPartner Programme : discontinued
- UI5 community: underfunded
What’s going on??? #ABAP

🔥 #zed.dev is the best #IDE I saw in a long time. Fast, written in #Rust and SOTA #AI capable. Who wants to write a #SAP #ADT based #ABAP extension for it? 🚨

To all vibe coders their managers and CIO/CTO/CEOs out there: In case of failure, there is no one to blame but yourself: Not IT, not developers, not consultants, certainly not the #AI itself. To quote Jack Reacher: “Remember - YOU wanted this”. #SAP #ABAP

I like the fact that #SAP tries to be ‘#AI ready’ but at least as for coding it is my firm belief that instead of coming up with their own agent (#Joule) they should have distributed a curated #ABAP training dataset to the real players.Would have made a world of difference for customers. Am I right?

In the age of #AI agent driven flow, #NoUI and #EphemeralUI become more central and will push browser, websites and apps into the background. Questions will be answered directly w/o UI - if necessary, it will be generated on the fly and vanish after use. Exposing data to AI is king #SAP #ABAP

Why wait on #SAP #Joule? If you are on S/4 with #OData (doesn’t matter whether on-prem or cloud), expose your business data 2-fold: 1. webapi #RAP V4, consume with #CAP and 2. webapi V2, consume with #MCP OData bridge. You are ready for #AgenticAI using SOTA coding and #noui agents. Done! #ABAP

We need #neurosymbolic architectures which augment the #LLM prompt with the result of symbolic reasoning. It’s not enough to augment with symbolic facts from a #KnowledgeGraph or #Ontology alone. Only then will #LLM output be symbolically verifiable. #AI #SAP #ABAP #KnowledgeRepresentation

#SAP: Noone wants to deal with 4 IDES!!! That’s unprofessional. We don’t need tool silos like SAP GUI, BAS, Eclipse ADT and VSC with Fiori Tools. Stop overcomplicating things at every turn and streamline your dev strategy. It will not end well otherwise. Take it from a well-meaning friend. #ABAP

#SAP: Focus on #AI compatible exposure of business data and let #AI native solutions deal with it. Get rid of annotation-based generation and let customers use SOTA coding agents to generate UI directly. This eliminates the artificial distinction between low/pro/free-style. And build ONE IDE! 🤔

Why let #Joule generate code with annotations which then in turn generates the final code instead of generating the final code right away? #SAP #ABAP

This emancipates customers from proprietary tools, from slow-rollout and expensive #Joule and get #SOTA #AI running independently of whether they are on-prem or not.On top of that #MCP your OData to make the data queriable.This can be done in ABAP or in #node.js or #Python or or or - #SAP #ABAP
2/2

Why not get rid of annotation hell and and why distinguish between free-style and annotated list reports and object pages? Let’s expose the relevant data using web-api #RAP without ui annotations and let coding agents render the UI via #CAP according to fiori elements specs! #SAP #ABAP
1/2

I’m pretty sure that #SAP customers will learn fast how to expose their relevant business data via #MCP, use their #ABAP codebase to equip SOTA models and use more #CAP to leverage the best coding agents w/o the need to rely on a slow and expensive roll-out of a proprietary #Joule offer.
2/2

With Gemini 2.5 Pro, Gemini Jules, Gemini Code Assist and now Gemini CLI, Google is fast pacing to the top. It’s only a matter of time until customers understand that #SAP will never be able to ship things of that quality for that pricing - they are just not an #AI first company #ABAP
1/2

The out of distribution (#OOD) dilemma of #LLM is comparable to the closed world assumption (#CWA) dilemma of classical rule based systems. The incapability of #AI to reason out of the box was and is the core problem. People have to understand this to escape the hype. #LLM #AgenticAI #SAP #ABAP

In light of more and more statements like “30% of our code is meanwhile generated by #AI” (the catch here is ‘generated’).
The real question is: How much code is being generated by #AI and deployed to production without any humans in the loop after the prompt? It’s a big fat zero! #ABAP

Some nice lecture for the weekend:
‘ARTIFICIAL INTELLIGENCE MEETS NATURAL STUPIDITY’ Reminder: This was written in 1976. Similarities with 2025 players are purely coincidential in nature 😉
cs.fit.edu/~kgallagh...

This new paradigm would also avoid the current problem that niche and new languages will not have a chance to be supported by #AI coding agents due to lack of training material. #ABAP
2/2

#AI coding should be based on understanding higher order coding concepts being agnostic to any specific language. Rendering into a language is only the last step in that chain. I fear that current transformer architectures simply predicting code token by token are not well-suited for this. #ABAP
1/2

My personal view is that everything that’s not explicitly extending #Fiori floorplans should be done with #CAP leveraging webapi exposed #RAP services in a hybrid approach. This means a much better coding universe and better embedding in #AI coding made by the big 5. #ABAP
2/2

#ABAP as a niche language whith much less training data will be outperformed by its competitor coding agents. Languages which perform well will likely get much more footprint in the future in #SAP applications. Will that mean a shift from #RAP to #CAP where this is possible?
1/2