Nils Durner's Blog Ahas, Breadcrumbs, Coding Epiphanies

Model "Air"

~Friday fun~ Saturday Shenanigans: AI models don’t have bugs, but they may have “an air”. (screenshot feat. Megan Leszczynski) Read more

Google Bard's Vulnerability to Adversarial Images

Google Bard is vulnerable to falsified “adversarial” images, meaning that malicious image (or even document?) creators can mislead it to form wrong conclusions and circumvent its security mechanisms: By attacking whitebox surrogate vision encoders or MLLMs, the generated adversarial examples can mislead Bard to output wrong image description... Read more

Document Processing

Quick recap on recent advancements in Document-processing below. Following up on the huge context size of the Llama 2 code model (> 100K), contributors of non-profits EleutherAI and Nous-Research have expanded stock Llama 2 attention windows to 128K tokens Note that this size is only very comparable to OpenAI GPT, Claude etc. b... Read more

ChatGPT-on-GPT-4: Knowledge Cut-off and Performance

ChatGPT-on-GPT-4 now reports knowledge cut-off in January 2022 and will answer questions even beyond. However, anything I‘ve tried turned out hallucinous (or vague/anticipated, e.g. the French elections in February). GPT-4-via-API still reports cut-off as of September 2021. (via https://x.com/simonw/status/1704144941172117686?s=12) X (for... Read more

OpenAI's GPT 3.5 Instruct Model Announcement

OpenAI’s new GPT 3.5 Instruct model announced through E-Mail, pasted below. This is a drop-in replacement to text-davinci-003. It is widely used in code snippets around the web, so it’s good to finally have an optional available that combines the traditional TextCompletion interface and mode of operation with the benefits of GPT-3.5-turbo. Notew... Read more