My #GenAI workbenches for OpenAI, Anthropic and Amazon #Bedrock are a pastime that has kept me interested for quite some time now, so I'm confident to share these more broadly: these tools serve as my personal Swiss Army knife and intelligent utility for diverse applications across work, play, and vacation.

🛠️ Workbench variants
on OpenAI: https://huggingface.co/spaces/ndurner/oai_chat
on Anthropic: https://huggingface.co/spaces/ndurner/claude_chat
on Amazon Bedrock: https://huggingface.co/spaces/ndurner/amz_bedrock_chat

Distinctive features:
💫 ahead of the curve: custom prompting and sometimes pre-release model access, with reliable GPT-4 "classic" available for fallback, elevate performance beyond what's possible with ChatGPT etc.

🌞 accessible: allows access to Claude 3 in the EU, and unlocks higher usage limits (subject to #AWS, OpenAI or Anthropic agreements). Better user experience than handling Google Collab sheets.

🌟 cost effective: the pay-per-use model using your own API key allows for cost distribution across team members, which can be more economical than individual flat-fee subscriptions.

🌄 vision capabilities to discuss images and photos, a feature not commonly available in typical #LLM Playgrounds.

💡 educate about & experiment with #GenerativeAI, and experience its job-transformative potential beyond #ChatGPT

✨ bonus features: file upload including basic Word file reading, history export/import for reuse or sophisticated prompting techniques, file download, reproducibility, support for Mobile, custom system prompt for AI personas, … and perhaps more to come.

🔒 mitigation against the AI Assistant snooping attack by Roy Weiss et al.

🚀 self-hosted deployment option or ready-to-run hosted variant.

(🌖 frozen models with static world-knowledge, not internet-enabled like #PerplexityAI)

These tools are shaped around my personal and local community's use. I am open to suggestions however, and a very modest write-up to get started with some of the more advanced features is here: https://lnkd.in/eS7xvEGk.

As a Swabian from #TheLänd, I really only pay if I am really, really convinced. 👛🔒,💸🙅🏻‍♂️💯. The professional-grade services underlying these tools justify the effort and time for any dedicated professional - as opposed to the consumer-grade offerings that may just refuse to work when demand is high and generally only give access to technology that is so very behind by today's standards. "You get what you pay for", as the saying goes.
Through the underlying frontier language models, these AI tools can almost be likened to a young apprentice: capable of conceiving fresh & brilliant ideas and eager to tackle the tedious tasks. Yet, it remains crucial for the Maestra or Maestro to diligently check and co-iterate on results.

Nota bene, thus: This tool is an ongoing experiment, provided with no warranties. You are solely responsible for its use. Follow the science and share your experiences.

If you are aware of similar projects or have insights to share, I would appreciate hearing about them.

🔔 Keeping up
on Hugging Face: https://huggingface.co/ndurner
on GitHub: https://github.com/ndurner/

🛫 Recommended high-level background
"Co-Intelligence" by Ethan Mollick: https://www.linkedin.com/posts/emollick_co-intelligence-by-ethan-mollick-9780593716717-activity-7183949270348124160-cR1J

Healthcare AI Build vs. Buy: Lessons on building genAI solutions in house: https://elion.health/resources/webinar-ai-build-vs-buy

AI Index report 2024 by Stanford HAI: https://aiindex.stanford.edu/report/?sf187708151=1

🤿 Recommended deep dive
"Writing Principles for Task-Tuned Prompt Engineering" by Karina Nguyen: https://www.youtube.com/watch?v=6d60zVdcCV4

Anthropic Prompt library: https://docs.anthropic.com/claude/prompt-library

OpenAI Cookbook: https://cookbook.openai.com/

🔬 Recommended research preprints
Sparks of Artificial General Intelligence: Early experiments with GPT-4: https://arxiv.org/abs/2303.12712

A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models: https://arxiv.org/pdf/2401.01313.pdf

A Systematic Survey of Prompt Engineering in Large Language Models: Techniques and Applications: https://arxiv.org/pdf/2402.07927.pdf

An Empirical Categorization of Prompting Techniques for Large Language Models: A Practitioner's Guide: https://arxiv.org/pdf/2402.14837.pdf

Lost in the Middle: How Language Models Use Long Contexts: https://arxiv.org/abs/2307.03172

Ada-LEval: Evaluating long-context LLMs with length-adaptable benchmarks: https://arxiv.org/pdf/2404.06480.pdf

OpenEQA: Embodied Question Answering in the Era of Foundation Models: https://open-eqa.github.io/assets/pdfs/paper.pdf

AI Attribution in Art offers perspectives for tech workers pondering #GenAI disclosure: from subtle references to detailed disclosures, artists are using various approaches to acknowledging AI contributions in their work.

https://www.linkedin.com/pulse/ai-attribution-art-perspectives-software-engineers-nils-durner-mx68f

My article examines:
🔍 The "AI Assisted" label used by curator Prof. Janet Bellotto

🔍 Segregated disclosure at the Singapore Art Museum

🔍 Integrated disclosure by artist Daito Manabe

🔍 Comprehensive disclosure practices by:

🌠 the Italian art studio fuse* and

👨‍🎨 artist and writer Prof. Lev Manovich

As AI becomes increasingly interwoven with our development lifecycle, tech stacks and products, the #ModernArt community's thoughtful exploration of authorship, originality, and ethical AI use may offer valuable insights for our own practices. While it's unlikely that there's a one-size-fits-all solution, these examples provide food-for-thought as we consider the nuances of transparent and responsible AI attribution in our own fast-paced industry.

Whether you're a software engineer or in any other role shaping tech solutions, I invite you to explore these perspectives and consider how they might inform our work.

#GenerativeAI #SoftwareEngineering #SolutionConsulting # DigitalArt

The perceived RSA-3000 crypto mandate by the German Federal Office for Information Security (BSI) has been reported by @heiseonline highlighting that:
💡 a BSI speaker confirmed that this is a recommendation, not a mandate
💡 the TLS certificate of the #BSI website still uses RSA-2048 as well
💡 the wording, especially across BSI publications, is confusing and could be misleading

This reporting¹ is in the context of TLS (publications TR-02102-2, TR-03116-4), but the same issues are present with the general "Technische Richtlinien" document on cryptographic algorithms and key lengths (TR-02102 part 1), which is cited by sources like keylength.org, often without the nuance from the preamble, such as:
👩🏻‍⚖️ the recommendations do not preempt regulatory approval processes
🧑🏻‍💻 they target developers planning new systems
💫 they may exceed the stated goal of achieving 120 bits of security

The Heise article¹ concludes that "A #signature algorithm for TLS needs to be secure for only as long as the certificate is valid, which is typically one year." It also notes that the US National Institute of Standards and Technology (NIST) "considers #RSA with a key length of 2048 bits to be sufficiently secure for signatures until the year 2030".

¹ Heise article (German): https://www.heise.de/news/BSI-Verwirrung-um-Anforderungen-an-Schluessellaengen-fuer-TLS-Verbindungen-9596072.html