This has been another incredible month. I won’t get into the speculation regarding the OpenAI affair. A lot has been written, let’s just wait and see if it is just a compute war, a next generation training technique or both or maybe just a matter of money. In the meantime the landscape of artificial intelligence is evolving at a breakneck pace, with industry giants like OpenAI, Amazon, Microsoft, and Google leading the charge.
OpenAI first Developer Day did not disappoint. Here are the critical takeaways:
The newly introduced GPT-4 Turbo has set a new benchmark for large language models (LLMs) with its expanded 128K context window and training data updated until April 2023. This advancement means it can understand and generate much longer texts with greater coherence than ever before and with costs basically cut in half.
OpenAI’s Assistants API and GPT Store are game-changers for developers and entrepreneurs alike. These platforms are not only allowing for the creation of custom AI agents but also offering a marketplace to publish and potentially monetize these innovations. Now the exact potential has to be seen, given the fact that OpenAI screwed many plugins creators, when plugins seemed to be next “apps”.
The unveiling of DALL-E 3 and new text-to-speech APIs with seven AI voices extends OpenAI’s portfolio into the realms of visual and auditory creation, marking a significant expansion of AI’s creative capabilities.
In an important move to make their models more viable for business, OpenAI’s Copyright Shield aims to protect clients from intellectual property infringement claims, showcasing their commitment to supporting businesses in the integration of AI technologies.
Amazon is not far behind, with its announcement of Olympus, a new LLM poised to be smarter, faster, and more cost-effective than its predecessors. With an impressive 2 trillion parameters, Olympus promises updated knowledge, a longer context window, and versatility in web browsing, coding, and data analysis. Also form Amazon the first timid try to par Microsoft Copilot, with Amazon Q. Amazon Q emerges as an AI-powered assistant designed to understand business intricacies and provide generative content. While it primarily focuses on chat and knowledge mining, its extensive connectors offer significant implementation advantages, still it lacks actions. i.e. it gives instructions on how tos but it’s not able to complete the actions by itself, like Copilot does.
Speaking of Microsoft, this month has been the Ignite 2023 month. The first thing the I want to highlight is the struggle for compute power and independence. This is so true when Nadella in the opening keynote presents the first MS chip, Azure Maia 100 is a chip specifically targeted at AI marking a significant move towards hardware independence. Almost everything in Ignite was AI permeated. New Azure AI Studio, new models and a new models as a service platform that allows access to large models without extensive infrastructure and obviously Copilot for Microsoft 365 generally available from November 1st. A news with less mediatic impact but that will make life a lot easier for model teesting and training is the integration of OneLake datastore in Azure Machine Learning, facilitating a seamless transition between Microsoft Fabric and Azure Machine Learning. This integration allows data engineers to share machine learning-ready data assets developed in Fabric, enabling machine learning professionals to directly utilize them for model training in Azure Machine Learning. Additionally, machine learning professionals can write model predictions back to OneLake for further processing in Fabric or to surface insights through Power BI.
A lot more has happened during November. Azure OpenAI has broadened its horizons by supporting new data sources. Now, you can leverage Azure Cosmos DB for MongoDB vCore and URLs as data sources. This update simplifies the process of ingesting data for conversation with Azure OpenAI models, making it more versatile and efficient. GPT-4 Turbo Preview has been added to the supported models. OpenAI’s GPT-4 Turbo is a leap forward with a staggering 128,000 token context window, capable of generating 4,096 output tokens. Trained with the latest data up to April 2023, it’s designed for those who need cutting-edge AI capabilities. However, it’s worth noting that GPT-4 Turbo is still in preview and not yet suitable for production environments. Also GPT-3.5-Turbo-1106 has been added, it offers a 16,385 token context window and the same output token capabilities. It’s the current go-to for production applications, with regional availability and unique quota allocations. Finally DALL-E 3 landed on Azure OpenAI. Itrepresents the pinnacle of image generation models from OpenAI, delivering enhanced image quality, the ability to render more complex scenes, and improved text rendering in images. DALL-E 3 is accessible via OpenAI Studio and the REST API but requires your OpenAI resource to be in specific regions. What sets DALL-E 3 apart is its built-in prompt rewriting capability, aimed at enhancing images, reducing bias, and promoting natural diversity in the generated content.
Responsible AI support has also been improved. Customers now have the power to set severity levels for content filtering across a range of sensitive categories. This ensures that the AI aligns with the ethical standards and regulatory requirements of diverse global users. In an age where distinguishing between real and AI-generated content is increasingly challenging, Azure OpenAI now includes digital credentials in all DALL-E generated images. New layers of security and compkliance have been added:
- Jailbreak Risk Detection. This optional feature scrutinizes user prompts, preventing the AI model from engaging in undesirable behaviors or bypassing predefined rules.
- Protected Material Text and Code Models. These models ensure that outputs from the large language models are not inadvertently infringing on intellectual property, such as song lyrics or source code, by identifying and filtering protected material.
- Blocklists: Customizing AI Interactions. Azure OpenAI now enables customers to create custom blocklists that prevent specific terms or patterns from being processed or generated by the AI. This is in addition to the provided Microsoft profanity blocklist, further tailoring AI interactions to match user preferences and compliance needs.
November has also been the month of GitHub Universe. here again AI everywhere with the release of the GitHib copilot chat and much nmore:
- Copilot Chat, a new feature that enables natural language coding assistance, code-aware guidance, inline editing, and slash commands for developers on GitHub and in JetBrains IDEs.
- Copilot Enterprise, a new offering that provides enterprise-grade security, safety, and privacy for teams of developers, with access to internal and private code, pull request summaries, and smart actions.
- GitHub Advanced Security, a new feature that integrates AI-powered application security testing to detect and remediate vulnerabilities and secrets in code, with code scanning autofix, secret scanning, and regular expression generator.
- Copilot Workspace, a new feature that helps developers scale the barrier of putting an idea into code, by proposing an automatically editable plan, running and testing code, and fixing errors.
- Copilot Partner Program, a new program that creates an ecosystem for third-party developers to build plugins for GitHub Copilot, with over 25 debut partners.
- Future plans for expanding GitHub Copilot to more use cases, languages, and platforms, and to enable the intersection of human and artificial intelligence for software development.
Waiting for Gemini debut Google empowered Performance Max to use AI to create marketing content, showcasing AI’s growing influence in the advertising sector.
A Stanford-led initiative introduced the Foundation Model Transparency Index, assessing AI models’ openness about their training, architecture, and usage. This effort reflects the AI community’s push towards greater transparency and accountability. In Europe is moving also, worth of notice the DEXAI startup https://www.dexai.eu/ that tries to build a model to ADDRESS THE RELATIONSHIP BETWEEN ARTIFICIAL INTELLIGENCE, ETHICS & SUSTAINABILITY.
Generative AI costs are a debatable subject given the different pricing models and the different capabilities. What I get is that while GPT4-turbo will significantly cut costs, it will remain the most capable and expensive model, while for example GPT-3.5 is somewhat on par with LLAMA 2 running on dedicated hardware. Very interesting study here https://cursor.sh/blog/llama-inference
With rumors of OpenAI releasing GPT-5 in Q1 or Q2, Google Gemini set for Q1 and NVIDIA showcasing their new H200 chip, the boundaries of AI are continually being pushed. Additionally, GitHub’s Copilot for Enterprise and DreamWorks’ AI-driven cost-cutting strategies for animated films point towards AI’s expanding influence in various industries. We’ll see what December is crafting for AI.
Quae Nocent Docent 