27 May 24 - AI with Whiskey

Author

Shobhit Varshney
May 27, 2024

Recapping last week’s massive #GenAI announcements at IBM THINK, Microsoft Build, and Dell Tech World as of May 27th.

But first, cheers for Memorial Day.. 🥃 pairing suggestion.. Henry McKenna by Heaven Hill. Recommended by a good friend Madison Gooch from Kentucky. One of the longest aged, award-winning American whiskeys. Smooth oak, sharp spices and slight sweetness. Bottled-in-bond is a designation that sets a rigorous standard for quality ingredients, strict Govt regulated process, and distillery’s stamp of authenticity. Much like IBM’s open trusted transparent granite models 😉 (shameless plug)

IBM THINK

The Aha!

  • In my 12 years at IBM, this was our most impactful THINK flagship event. Real world client examples showcasing Open Trusted AI on Hybrid Cloud in action. The pace of innovation has never been faster, with a great ‘winning‘ attitude in the team.

  • Packed with 6k+ attendees (majority clients). Our strongest-ever showcase of clients sharing their AI successes on stage eg. Citi, Verizon, Elevance Health, PepsiCo, Merck, Toyota, Dunn & Bradstreet, Meta, Palo Alto, Takeda, Seville FC, Bloomberg, SAP, Adobe, Honda, Pfizer, Fiserv, NTT, BNP Parnibas, State Street, HSBC, Casper Labs, Lockheed Martin, SLB, Bolt, Broadridge etc.. the list goes on!

  • Love the quote from Arvind’s keynote “Open means choice. Open means more eyes on the code, more minds on the problems, and more hands on the solutions. For any technology to gain velocity and become ubiquitous, you’ve got to balance three things: competition, innovation, and safety. Open source is a great way to achieve all three.”

  • Rob shared a very comprehensive approach to enterprise automation, pulling together a singular narrative across IBM offerings, and introduced IBM Concert for AI-powered insights for all your IT apps

  • As always, the master class from Dario is a must-watch. He covers 3 critical steps to Enterprise’s AI mission and why each matter:

    1. Start with a trusted base model (ibm granite)

    2. Create a representation of your enterprise data (InstructLab)

    3. Deploy, scale and create value with your AI (redhat, watsonx)

  • Stanford released it’s Transparency Index for foundation models ranking IBM’s granite models as one of the most transparent in the industry, way ahead of AWS, Google and OpenAI

  • IBM open-sourced 18 granite models under Apache 2.0, and with training via InstructLab, our code models beat out the competition

  • IBM deepened its partnership with Mistral AI to bring Mistral-Large, their best-in-class proprietary model to watsonx.ai to be deployed anywhere - even on-prem : )

  • IBM showed meaningful progress with scaled client deployments of its AI Assistants eg. watsonx assistant, watsonx orchestrate (with 1,000+ pre-built skills) etc. Also released new watsonx Code Assistant for Enterprise Java, watsonx Assistant for Z, and AI Assistant Builder to build your own assistant.

  • IBM doubled down on ecosystem partnerships with AWS, Adobe, Meta, Microsoft, Mistral AI, Palo Alto Networks, SAP, Salesforce, and SDAIA to expand their AI capabilities and offer model choice, flexibility and governance through watsonx.

  • IBM Fusion 2.8 with OpenShift + watsonx + Nvidia L40s GPUs + Intel Sapphire Rapid CPUs delivers secure GenAI in a box

  • Overall, love IBM’s razor sharp focus on enterprise workflows, trusted models and innovating with the open-source community.

Uh-Oh!

  • With its outstanding InstructLab, IBM is taking on the huge challenge of codifying all the complex proprietary enterprise data for efficient and collaborative training of LLMs. IBM will need to ensure it runs effectively in all environments, and invest in domain taxonomies.

  • IBM understands the complexity of complex regulated processes in enterprises. As it brings more LLM Agent based agentic workflows, it would need to focus more on intelligence, transparency and control.

  • IBM needs a solid multi-modal LLM in its arsenal to help with images, charts, complex tables, video. Bring-your-own model somewhat addresses that today, and hopefully open models like Llama3 400B will be out soon and be offered on .ai

MICROSOFT BUILD

The Aha !

  • Microsoft made significant announcements at their BUILD event. 90 sec recap below, but honestly you should watch the full keynote.

  • Co-pilot is evolving into a platform, with co-pilot studio getting a lot of enhancements. Honestly, so far co-pilot was missing key enterprise features, and our team had to default to leveraging Azure AI services directly and build the complex end2end workflows for clients. With extensions, grounding with pre-built connectors etc, it’s getting closer to a mature platform to build your own copilots. Great demo.

  • Extending co-pilot to Teams, Project Mgmt etc. drives productivity, but the real unlock will be with Agents. This was a real good showcase of building agentic workflows with natural language like Google Vertex AI

  • Phi-3 series had great updates to their small open LLMs. Phi-3-mini with 4k and 128k context window is punching above its weight class with great MMLU scores. Cost conscious high volume clients like Khan Academy are already leveraging Phi models.

  • I was most excited about Phi-3-vision 4.2B model which runs on a phone, and is showing great potential for AI on the edge on IoT devices. Great with images, tables, charts etc. Competing with Google’s PaliGemma open vision LLM.

  • Co-pilot + PC - Huge jump in windows PC hardware performance with new NPUs to speed up on-device AI, locally run Phi-silica, RAG etc. This is a make or break moment for PC sales. Apple should be extremely worried about the huge lead Microsoft has with AI infused across Windows. Lots of WWDC rumors with OpenAI.

  • As expected, after OpenAI’s GPT4o demo and mac-only app release, it was inevitable that co-pilot would bring GPT4o capabilities natively to Windows. Like GPT4o, it can watch your screen/webcam and interact with you in real-time with voice, but goes a step further by even allowing co-pilot to control the screen/app eg. add items to cart. Watch 2 great demos here.

  • One of the least talked about updates was AI flows in Power Automate (in early preview). It lets you share your screen and voice over what process you are doing (eg. I go to outlook, and look for emails from vendors with attachments, grab and summarize the invoice dates, $ etc.) and it auto-creates the power automate flow, and manages updates etc. Watch the 2nd video here.

  • Introduced Real-Time intelligence in MS Fabric to discover, consume and manage real-time streaming data from hybrid clouds, with co-pilot baked in.

The Uh-Oh!

  • Microsoft released Phi-3 under MIT license v/s IBM released granite models under Apache 2.0 license. With the Apache 2.0 license, the enterprise can:

    • Freely use, modify, and distribute the library, including in commercial products .

    • Claim patents on any modifications or derivative works they create based on the library, without fear of infringing on the original contributors' patents .

    • Distribute their modified version without revealing the actual source code changes, as long as they provide a notice describing the modifications .

    However, if the same library was licensed under MIT, the enterprise would not have the explicit patent grant from contributors. This could potentially expose them to patent infringement claims if any of the contributors later obtained relevant patents. Additionally, there would be no requirement to disclose modifications made to the source code.

  • Windows Recall feature that remembers everything you do on your windows PC. Practically takes a screenshot of every thing, tracks every activity, and makes it instantly searchable. Brilliant! but news is blowing up on how huge a privacy risk this could be.

  • Microsoft data offerings have been constantly changing, but enterprises can’t keep jumping from one offering to another. Need some stability on core offerings and more focus on total cost.

Misc AI news

  • Google’s new AI Overview feature ran into some embarrassing fails eg. recommending glue to make cheese stick to pizza, eating 1 rock a day. Unfortunately some folks even posted mocked up fake responses. The core issue was, just like humans, Google was gullible to content from satire/comic sites like the Onion’s one rock article.

  • NVIDIA stock crosses $1,065 with another crazy quarter, with profits up 628% vs. 1Q2023. $2.62T market cap. 10-to-1 stock split in June.

  • OpenAI got into a controversy with their ChatGPT ‘sky’ AI voice that closely resembled Scarlett Johansson’s voice from HER. OpenAI had approached her for her voice, but showed evidence that the voice was from another hired actor, and still removed Sky and issued explanation.

  • At Tech World, Dell announced new AI Factories and Project Helix with Nvidia for secure on-prem GenAI. New Co-pilot+PC laptops.

  • Slack got into trouble as users realized their data was being used to train SlackAI by default unless they specifically opt-out.

  • Elon Musk’s x.ai raised $6B ($24B valuation in 1 yr) from US and Middle East to accelerate building open-source AGI. $1B will go towards just the Nvidia GPUs.

  • ScaleAI raises $1B from Accel/Amazon/Meta/Nvidia etc., to boost its valuation to $13.8B. They have done a phenomenal job at addressing the data labelling challenge to accelerate AI workloads particularly with industry specific models eg. Public Sector, Autonomous vehicles.

  • Meta gets called out for establishing an AI Advisory Council with only men on it.

--------
🔔 If you like such content, I encourage you to connect on my LinkedIn ♻️ Recommend your friends to subscribe to this free ‘AI with Whiskey’ newsletter.

Disclaimer - these are my thoughts, made in a personal capacity on an independent platform, and not on behalf of my employer
--------