This Week in AI 23.03.18
The AI model Horserace heats up
AI Tech and Product Releases
A big week for product announcements:
AI image synthesis service Midjourney version 5 has been released, and it’s stunning: AI-imager Midjourney v5 stuns with photorealistic images—and 5-fingered hands. This figure shows their comparison of version 5 with previous versions. It’s getting very photo-realistic.
Microsoft announces Microsoft 365 Copilot, an LLM-powered chat tool “embedded in the Microsoft 365 apps you use every day — Word, Excel, PowerPoint, Outlook, Teams and more.” Copilot is more than just a chat bot, but an integrated system that ties in with the Office tools and grounds itself in emails, docs and other business data and feed that to the LLM. Their interface for this is called business chat:
Business Chat works across the LLM, the Microsoft 365 apps, and your data — your calendar, emails, chats, documents, meetings and contacts — to do things you’ve never been able to do before. You can give it natural language prompts like “Tell my team how we updated the product strategy,” and it will generate a status update based on the morning’s meetings, emails and chat threads.
What Github Copilot did for code, this will do for business documents. There was a time when Microsoft was the more stodgy of the Tech giants, but Microsoft 365 Copilot seems just so darned useful - just as Github Copilot is proving - that it surely will become a hit.
Anthropic anounces Claude, “an AI assistant by Anthropic and ChatGPT rival promising a better AI experience with less toxicity, bias, and hallucinations.” They mention use in partners such as Quora, DuckDuckGo, and Notion. In particular:
Quora has incorporated OpenAI GPT-4 and Claude-v1.2 into their new app, Poe.
Google announces PaLM API and MakerSuite - "PaLM API & MakerSuite: an approachable way to start prototyping and building generative AI applications" These announcement are limited releases and access. Here’s more on the Google ecosystem - there are many third party companies in the AI space in Google’s orbit.
And not to be out-done by Microsoft’s Copilot, Google announces AI features in Gmail, Docs, and more to rival Microsoft. Integrating generative AI and LLM features in Gmail and google docs, it includes the “option to generate full emails in Gmail based on users’ brief bullet points, and the ability to produce AI imagery, audio, and video to illustrate presentations in Slides.”
Edit: Wait, one more! On Thursday, Baidu announced Ernie bot, a chatGPT-like chatbot for the Chinese market. We should not forget that China is pursuing AI as vigorously as the US Big Tech companies.
There’s so much going on in the AI space! With these and the GPT-4 announcement a whole 5 days ago, you have six seven big announcements for this week, not to mention others we haven’t covered.
AI Research News
I shared Alpaca instruction-following model news in a prior post. This is a big deal in a positive way for AI research.
Because of so many product announcements, I’ll go light on the research this week, but promise to get more papers reviewed and discussed next week. The pace of AI research and AI product development is not slacking off.
AI Business News
The above product announcements show the AI land rush is upon us, and one clear winner is Nvidia, selling the GPUs that are used to train and implement these powerful AI models. Stay tuned for the upcoming NVidia GTC on March 20-23. I will post some summaries on it.
Snippet from a Job Posting for ML Engineer at RobinAI, which gives a glimpse of a typical tech stack for AI startups in this era of LLMs:
"We incorporate Large Language Models (LLMs) from third-party vendors ... For defining in-house models, we use two ecosystems - HuggingFace Transformers and scikit-learn. We use AWS services such as Sagemaker, ECS and Lambda to make these models available for use in our products (which are also hosted on AWS). We use CircleCI to build and deploy, and all of our infrastructure is managed using Terraform"
AI Opinions and Articles
The not-so-good trend towards closing AI research was discussed in a twitter summary on OpenAI's Decision to Limit Disclosure of GPT-4 Sparks Debate in AI Community. More on this topic from the Verge. OpenAI not disclosing technical details on GPT-4 portends a bad trend in AI research that will slow progress and make foundation models opaque. In my view, AI model makers need to be transparent if not fully open about their models, and more open AI models will be better and safer for us than closed AI models. Most of the AI community agrees with this perspective, but most of us aren’t the decision-makers in the BigTech organizations building these models, and those people, like Ilya Sutskever, are inclined to close the door.
With so many useful research results and amazing AI products coming out, this is not the time to opine about the future of AI or to quibble over the pessimistic think-piece at The Atlantic. Instead, I will try out the present of AI and figure out what it means for how we can be productive with these incredible new tools.
AI Useful Items
I hope to share a lot more in future Substack posts about interesting use cases and applications for GPT-4 and other models, as well more in-depth on AI model prompt engineering. There is so much coming out that I think this Substack’s utility will be in helping readers like you fight through the clutter. Let me know in the comments what things you’d like me to focus on.
In the mean times, via a tweet from Charles Miller, suggestion on useful AI for copywriting: “1. Chat GPT - Research 2. QuillBot - Paraphrasing 3. StoryLab - Hooks and outlines 4. Grammarly - Grammar/spelling 5. Hemingway - Conciseness/clarity 6. Power Thesaurus - Thesaurus 7. Tweet Hunter - Content creation.”
A Look Back …
Ars Technica article on PLATO, an advanced educational computer developed from the 60s to the 80s. I recall playing with a PLATO terminal as an undergraduate. It’s a great example of the kind of ambition that was attempted on machines with far less memory and power than today.