Suggestions
Andreas Koch
Tech Industry Executive I Strategy, Corporate Development, Alliances & Corporate Venture I AI, Security & Networking I VMware
Andreas Koch is an AI-savvy technology industry strategist with a strong track record in leading product and GTM strategy, as well as M&A at prominent tech companies ranging from startups to PE portfolio firms.
He excels in corporate and BU portfolio optimization, strategic alliances, corporate development, and venture capital.
Andreas possesses expertise in AI/ML, cyber security, software-defined data center, edge computing, and networking.
His educational background includes an MBA in Strategy and Corporate Finance from the University of California, Berkeley, Haas School of Business, and completed Ph.D. coursework in Philosophy at UCLA.
With past roles at companies like VMware, Tempered, Pulse Secure, Citrix, and Juniper Networks, Andreas Koch's professional journey has seen him make significant contributions in strategic advisory, business development, and corporate strategy.
He has a deep network within the tech industry and is recognized for his strategic insights and ability to drive growth and innovation.
Highlights
A long time ago I studied philosophy. That's why I had to get the following off my chest:
Pure play neural networks (Transformers, CNNs, RNNs or even SSMs) are unlikely to ever emulate what we associate with fully fledged intelligence - such as our own human intelligence. The reason is that
nothing can be a genuinely intelligent system unless it has the capability to discover truths. It needs to have a fairly reliable mechanism for (collectively) arriving at truths, accumulating them and structuring them into theories. Pure play Neural Networks cannot succeed at this task. Here is why.
A system's ability to systematically accumulate knowledge requires that it follow essentially the same collective workflows that are familar to us as the "best practices" of modern scientific research. Scientists start by developing hunches based on data patterns. That's where admittedly neural networks can assist. But then comes the essential next step: If we aim at reliable discovery of truths, we formalize our hunches into precise testable hypotheses which ca then be shown to have deductive links to other propositions - downstream to observations, upstream or vertically to already established general truths. Hence the scientific process of discovery requires a systematic interplay of deduction and induction, and simpler versions of this process are required to gain knowledge of any kind.
True believers think we can build AGI on neural network "learning" alone. They are betting that the process of scientific research could be replaced by brute force neural network extrapolation from training data. The problem with this view is that there can no hypothesis testing and no organizing them into theories, without rigid adherence to deductive rules of inference. And these rules themselves cannot be extrapolated from actual human practice as reflected by online content - no matter how large this training set, how large a context of tokens the model can inspect at a time, or how many parameters it uses (see also the Forbes article below).
Another issue with reducing intelligence to neural network learning is the problem of recursive AI content generation: As the share of AI generated content in the incremental GenAI training sets grows, the quality of the model output declines, until a point of "model collapse" is finally reached.
I suspect that this now well known problem can be avoided by introducing side constraints requiring logical/deductive internal consistency and consistency with "observational" data, i.e. with content that is demonstrably simply a description of (actual/non-synthetic) data that comes directly from physical systems. Something to be investigated...😊
VMware is taking the next step in xDR. Unified threat analytics and unified policy across endpoint, network infrastructure and applications and data
