Nils Durner's Blog Ahas, Breadcrumbs, Coding Epiphanies

AI Security Concerns

Following up on the publication of “AI security concerns in a nutshell - Practical AI-Security guide” by the German BSI (Federal Office of Information Security), I highlighted that another, more specific security concern to generative AI is prompt injection, a throwback to the SQL injection attacks of the… 2000s? 2010s? Anyway, OpenAI’s proposed... Read more

Aleph Alpha at the Federal Computing Center

Use Cases for AI presented by Matthias Lichtenthaler of the Federal Computing Center 🇦🇹 (“Bundesrechenzentrum”), at Future-Law “Digitaler E-Signatur Tag 2023”: Virtual Assistant: Case Categorization(?) Consultation/Assessment Semi-automated intelligent processing/answering Recognizing intentions Providing forms A... Read more

Visual GPT

Visual GPT, the multi-modal interface that a German Microsoft manager likely referred to in their earlier hints, has been released as source code: Github. And it’s a bit disappointing… at least when running on CPUs only, rather than GPUs. Details: Early reports are light on details, so here’s what to expect with the “Advice for CPU users”: t... Read more

Open Source LLMs

Entire language models are being pirated and touted as “Open Source”. This includes Vicuna, which has been making headlines as “90% quality of OpenAI ChatGPT” (which is B.S.). This illustration by Sahar Mor makes a distinction between “Research” and “Commercial”, and “Research” partially implies “legally tainted” (may also be true for ⚡ Lit-LlaM... Read more

Training LLMs

From the Amazon (pre-)announcement of Bedrock: One of the most important capabilities of Bedrock is how easy it is to customize a model. Customers simply point Bedrock at a few labeled examples in Amazon S3, and the service can fine-tune the model for a particular task without having to annotate large volumes of data (as few as 20 examples is... Read more