24 Aug 2023 - Emerging Technologies Meeting Agenda about zenith HOT 25 CLOSED

How embarrassing. I didn't record my own attendance!
Keith O'Donnell | Feynic Technology

(Don't beat me up, Rimma!)

from zenith.

I can't believe I missed this meeting! GREAT PPT BACKGROUND EVER (speaking as a D&D player)

from zenith.

I can't believe I missed this meeting! GREAT PPT BACKGROUND EVER (speaking as a D&D player)

Check out https://github.com/finos/zenith/blob/main/Governance/SIG%20meetings/2023-08-24/Zenith%20-%20Aug%2024%202023.pdf for the (non animated version of the) slides.

from zenith.

Patrick Downing | Morgan Stanley

from zenith.

Carly Richmond | Elastic

from zenith.

Kendall Perez | FINOS

from zenith.

Tadhg Wilson | ZNGLY

from zenith.

Conor Cassidy | Zngly | @cassidyconor1909

from zenith.

Charlie Greener | Theria

from zenith.

John Davies | Independent

from zenith.

Richard - Mobica

from zenith.

James McLeod / FINOS

from zenith.

Ricardo Sueiras / AWS

from zenith.

András Velvárt - Response

from zenith.

Nick Williams / Morgan Stanley

from zenith.

Junji Katto / Itau Unibanco

from zenith.

Joe Zang / FNMA

from zenith.

Alvin Shih / Morgan Stanley

from zenith.

Polina Levyant | GSR Markets

from zenith.

Marco / Gedox

from zenith.

Win Morgan | FINOS

from zenith.

Mike / ZNGLY

from zenith.

** Three drivers we are working on now at ZNGLY on the further integration and adoption of GenAI into the workflow of content publishing is 1. Entitlement. 2. Taxonomy and 3. Readiness - delighted to chat through any time - check out our integrations here https://www.synthesia.io/integration/zngly

from zenith.

Meeting minutes:

00:00:35 - The meeting begins, Keith welcomes participants, and James enters the room.
00:00:48 - Patrick Downing mentions recording the meeting.
00:01:21 - Keith discusses the music and mentions Charlie's presence.
00:02:04 - James talks about building a music playlist for calls.
00:02:18 - Laughter and random sounds.
00:03:18 - James welcomes attendees and introduces the theme.
00:04:38 - Discussion about starting calls with music.
00:05:40 - Kendall provides legal and conduct reminders.
00:06:46 - Keith gives information about upcoming events.
00:08:11 - A fantasy-themed introduction to the meeting's agenda.
00:08:29 - Introduction to the agenda and the flexible nature of the meeting.
00:10:03 - Keith announces the upcoming showcase by Matt Bain.
00:10:30 - Keith announces the new schedule for SIG meetings.
00:11:45 - Keith introduces the deep dive segment by Matt Bain.
00:15:14 - Matt demonstrates AI-powered analysis of application models.
00:20:39 - Matt discusses potential use cases of AI language models.
00:22:05 - Keith presents administrative details, Github repo, mailing list.
00:22:21 - Keith invites participants to take action after the meeting.
00:23:20 - Keith concludes the presentation and opens the floor for discussion.
00:24:10 - James McLeod references a question in the chat about the stability of Chat GPT and its non-deterministic nature.
00:24:21 - James raises the question of whether Chat GPT is stable and discusses the variance in its quality when run multiple times.
00:24:34 - Keith talks about his experience using Chat GPT for code reviews and vulnerability testing, leveraging its non-deterministic nature to explore various perspectives.
00:24:46 - Keith continues discussing the potential applications of Chat GPT for penetration testing and addressing challenges related to determinism.
00:25:04 - Keith refers to a presentation by Matt using Chat GPT for architecture discussions and mentions the need to enhance open source tools like LLAMA to be more robust.
00:25:22 - Keith opens the floor for others to share their views on the topic of Chat GPT's determinism and stability.
00:25:38 - Keith acknowledges a correction from Ricardo about LLAMA not being open source.
00:25:59 - Ricardo Sueiras expresses interest in the topic of AI and its simultaneous appeal and frustration due to its non-reliable nature.
00:26:17 - Ricardo discusses the challenges of using AI, especially in regulated industries that require repeatability and consistency.
00:26:35 - Keith shares his perspective on ChatGPT, seeing it as an assistant that offers alternative viewpoints rather than a fully reliable tool.
00:26:55 - Keith emphasizes the potential use of ChatGPT as an assistant and points out its limitations for full-fledged production.
00:27:24 - James McLeod asks a question to Ricardo and Keith about the differences between LLAMA and ChatGPT, as well as the challenges of open source AI.
00:27:29 - James requests clarification on LLAMA and open source AI models and their differences from proprietary models.
00:27:48 - James reflects on the importance of understanding terminology for those new to the topic.
00:27:55 - Keith asks Ricardo to elaborate on LLAMA and open source AI models.
00:28:01 - Ricardo explains the challenges of defining open source AI, including data and infrastructure considerations, and mentions efforts to define open source AI standards.
00:28:15 - Ricardo discusses the complexities of open source AI models and the differences between open source and open innovation.
00:28:49 - Keith, Ricardo, and Alvin share insights on open source AI, the challenges of sourcing information, and the importance of trust and transparency.
00:29:59 - Ricardo poses a question about the group's comfort level with using proprietary AI models as a foundation for open source tools.
00:30:50 - Keith reflects on the potential challenges of using proprietary AI models as an abstraction layer for open source tools.
00:31:06 - Alvin discusses the idea of using AI models as a starting point and mentions a project called Supercharger that aims to generate code with nonzero temperature.
00:32:04 - Alvin and Keith share insights on AI models and their potential applications, along with the challenges of trust, sourcing, and innovation.
00:33:14 - James McLeod suggests introducing participants and their roles to provide context for the discussion.
00:33:27 - Alvin Shih introduces himself as a developer from Morgan Stanley, specializing in open source and machine learning research.
00:34:17 - Ricardo Sueiras introduces himself as a developer at AWS with expertise in open source and innovation, focusing on open source strategy and emerging technologies.
00:35:10 - Keith asks for further contributions from participants on the topic of open source AI and proprietary models.
00:36:11 - Ricardo discusses the challenges of vendor lock-in, trust, and standardization in the context of open source AI and proprietary models.
00:37:04 - James McLeod shares his perspective on the benefits of open source, the importance of transparency, and the challenges of building trust in AI models.
00:40:52 - Carly Richmond emphasizes the importance of sourcing information, references, and documentation in AI models, along with the challenges of determining accuracy and building trust.
00:42:23 - James McLeod introduces the topic, reflecting on the recent London user group meeting where concerns were raised about the transparency of AI-generated responses.
00:43:01 - Ricardo Sueiras builds on the concern of transparency and control in AI systems, especially in the context of financial services. He highlights the challenges of understanding and influencing AI-generated responses.
00:43:37 - Ricardo Sueiras discusses the difficulties architects face in knowing who has control over AI systems and how to ensure reliable responses, given the growing complexity of these tools.
00:44:18 - The conversation shifts to the Brain Trust, an amalgamation of financial services members focusing on open source AI. Michael Wilson shares the progress his team has made in integrating generative AI with platforms and discusses the importance of taxonomy and entitlement in handling large-scale AI content.
00:45:40 - Michael Wilson explains their work in API integration and publishing AI content, emphasizing the significance of proper entitlement and organizational taxonomy to manage the increasing volume of AI-generated content.
00:47:08 - Michael Wilson provides information about how interested parties can join the Brain Trust and engage with the community. Keith J. O'Donnell elaborates on the structure of the Brain Trust and its involvement in various financial services discussions.
00:49:26 - Keith J. O'Donnell encourages interested individuals to reach out for membership details and highlights the benefits of being part of the Brain Trust and broader Finos community.
00:49:31 - Keith J. O'Donnell acknowledges the progress and playfulness in the Zenith meetings and announces upcoming themes.
00:49:40 - Participants discuss their experiences with generative AI tools. Ricardo Sueiras shares his observations about the effectiveness of such tools, particularly in working with established technologies versus emerging ones. The limitations and challenges of using these tools, as well as the importance of developer experience, are discussed.
00:52:04 - Michael Wilson commends the progress made in integrating AI with platforms and acknowledges the fun and innovation that participants bring to the discussions.
00:53:00 - James McLeod invites participants to share updates about their work, experiences, and experiments with AI tools.
00:53:19 - András Velvárt presents a different perspective, suggesting that generative AI tools can be helpful for experienced developers to learn new technologies and overcome hurdles. Ricardo Sueiras and others express their views on the potential benefits and limitations of these tools in different learning scenarios.
01:00:01 - Ricardo discusses copyright concerns related to code generated by AI tools. He highlights the recent legal developments, such as the inability to copyright content created solely by AI.
01:00:58 - Keith wraps up the meeting, summarizing the discussions and encouraging participants to stay engaged through the website and related activities.

from zenith.

Shared links throughout discussion:

• supercharger wraps the LLM so that it writes variations of code and unit tests, runs them in a container, and then has a "reflection" stage to pick the best one

https://github.com/catid/supercharger/

• Falcon 40b might be untainted, but there is a concern that Falcon 40b Instruct is not.

https://huggingface.co/tiiuae/falcon-40b-instruct

• "Baize is an open-source chat model trained with LoRA. It uses 100k dialogs generated by letting ChatGPT chat with itself."

https://github.com/project-baize/baize-chatbot

• An LLM can still hallucinate if the quality of the vector search results are poor. Context-obedient LLMs are one direction being explored.

https://www.reddit.com/r/LocalLLaMA/comments/1408ued/airoboros_gpt4_instructed_contextobedient/?rdt=65232

• Ontology vs Tags:

https://gwern.net/doc/technology/2005-04-shirky-ontologyisoverratedcategorieslinksandtags.html

from zenith.