AI Week In Review 23.11.11
OpenAI's DevDay, GPT4 Turbo, GPTs, Human AI Pin, Amazon Olympus, Samsung Gauss, the Matic, Levels of AGI
TL;DR - OpenAI DevDay made the biggest splash of the week, and other AI news hit: Humane AI Pin was announced; X’s Grok AI model is in the wild; both Samsung and Amazon are in the AI model race. CNN called it “the most momentous week for artificial intelligence since the launch of ChatGPT last year.”
AI Tech and Product Releases
The biggest news and release of the week is OpenAI’s DevDay We covered it in detail already. The top-line announcements were:
GPT-4 turbo, with longer (128K) context, better control, and more modalities
GPTs, customized versions of ChatGPT
Whisper V3, even better speech recognition
The Assistant API, putting all the power of GPT-4 turbo in OpenAI API plus other features to enable AI agents and apps
The Humane AI pin has launched. It’s calling itself “The first wearable device and software platform built to harness the full power of artificial intelligence (AI).” It offers a camera, projection-onto-hand screen, voice interface to search and AI, and build on Cosmos, Humane’s operating system. It will be available this week, starting at $699. Will it sell? Wired has more.
Google Generative AI in Search has been rolled out to 120 more countries. An incremental update in Search Generative Experience, but it means a lot globally. Access to AI is broadening.
In another incremental update, YouTube is testing two new AI features:
YouTube is testing two new features integrating generative AI into its viewing experience: The comment topics tool and the conversational AI tool. In a blog post, YouTube says these new features should help viewers better understand the content they're consuming and help creators connect with their audiences.
X has added an AI-powered “similar posts” feature. Meanwhile, X’s Grok AI model is rolling out to Premium+ users. It’s snarky persona seems to be the most prominent and notable feature. Elon Musk is reported saying Grok will be put into Teslas in addition to being hosted on X.
The AI Machine is shared by Runway CEO is a whimsical and retro take on creating a hardware AI video generation mixer. We posted a picture of this at the top. Real or AI? Or both?
A new AI-based robot vacuum cleaner, The Matic, is now on sale: “The Matic is a fully autonomous robot vacuum that its founders claim will clean your floors without getting stuck on cables or toys and without sending a map of your home to the cloud. And it’ll only cost you $1,800.”
Top Tools & Hacks
Via cocktailpeanut on X: “Mirror: A hackable AI-powered Mirror on Your Laptop.”
“Mirror is a simple yet powerful web app that runs 100% locally, where the AI @ggerganov's llama.cpp + @skunkworks_ai's Bakllava) constantly watches your webcam feed and sends you messages. And you can try it too, right now.”
The use cases are endless, but for one, will we disrupt the security industry?
AI Research News
So, even GPT-4 vision has powerful new capabilities, but it suffers from the same hallucination issues of ChatGPT and GPT-4 itself. Holistic Analysis of Hallucination in GPT-4V(ision): Bias and Interference Challenges puts GPT-4V through a new benchmark, namely, the Bias and Interference Challenges in Visual Language Models (Bingo).
The term Bias here “refers to the model's tendency to hallucinate certain types of responses, possibly due to imbalance in its training data.” They identified a notable bias in GPT-4V toward Western images and English writing, vulnerability to leading questions, and confusion when presented with multiple images. They found similar in LLaVA and Bard. Further, self-correction and chain-of-thought reasoning do not resolve these challenges.
CogVLM: Visual Expert for Pretrained Language Models presents a powerful open-source visual language foundation model:
CogVLM bridges the gap between the frozen pretrained language model and image encoder by a trainable visual expert module in the attention and FFN layers. As a result, CogVLM enables deep fusion of vision language features without sacrificing any performance on NLP tasks. CogVLM-17B achieves state-of-the-art performance on 10 classic cross-modal benchmarks.
The paper Efficient LLM Inference on CPUs comes from Intel, and proposes an efficient way to deploy LLMs on CPUs, by using automatic INT-4 weight-only quantization and an LLM runtime that has highly optimized kernels to accelerate the inference on CPUs.
From Google Deep Mind comes Levels of AGI: Operationalizing Progress on the Path to AGI. It classifies the capabilities of Artificial General Intelligence (AGI) models and their precursors in a framework that introduces levels of AGI performance, generality, and autonomy.
To develop our framework, we analyze existing definitions of AGI, and distill six principles that a useful ontology for AGI should satisfy. These principles include focusing on capabilities rather than mechanisms; separately evaluating generality and performance; and defining stages along the path toward AGI, rather than focusing on the endpoint. With these principles in mind, we propose 'Levels of AGI' based on depth (performance) and breadth (generality) of capabilities, and reflect on how current systems fit into this ontology.
With this structure in mind, they have placed existing AI models in their framework, see below. It seems to be a helpful roadmap to help answer the question: When AGI?
AI Business and Policy
Amazon is reportedly racing to build an AI model called Olympus to take on ChatGPT and Bard. The reports says it may have as many as two trillion parameters, and would be used as a superior AI to back Alexa.
Samsung unveils Gauss AI models that can generate text, code and images. Everyone is getting in on the AI act! The Samsung effort is 3 AI Models: Gauss Language, which is similar to ChatGPT; Samsung plans to incorporate it into its phone, laptop and tablet devices. Samsung Gauss Code for coding assistance. And Samsung Gauss Image, an AI image generation tool.
The AI talent war is really heating up: OpenAI’s New Weapon in Talent War With Google: $10 Million Pay Packages for Researchers.
Google in talks to invest in AI startup Character.AI. This is another investment in the hundreds of millions at a valuation of possibly $5B. Character.AI seeks capital to “train models and keep up with user demand.”
A study on 100s of consultants at BCG using ChatGPT with an intuitive finding that LLMs boost performance of the less skilled.
AI Opinions and Articles
Ben Goertzel Says the Singularity Will Happen by 2031:
Ben Goertzel, CEO of SingularityNET—who holds a Ph.D. from Temple University and has worked as a leader of Humanity+ and the Artificial General Intelligence Society—told Decrypt that he believes artificial general intelligence (AGI) is three to eight years away. AGI is the term for AI that can truly perform tasks just as well has humans, and it’s a prerequisite for the singularity soon following.
His argument, that AI progress is happening fast due to the heavy attention and investment in it, seems borne out by recent events.
In a FoxNews interview, Dr. Michio Kaku says: “I don't believe AI will be the death of civilization anytime soon.” Phew, but he had the caveat ‘anytime soon.’
A Look Back …
While the Humane AI Pin is an innovative and attractive device in many ways, it reminds me of another innovative wearable smart device, Google Glass.
Google Glass was first launched in 2013 with much hype, but it failed to take off. You could blame a lack of product-market fit, about 10 years ago. it faced questions about use-cases as well as “privacy concerns and social acceptance challenges due to its built-in camera and ability to record video without others’ consent.”
After two years, Google pivoted to niche enterprise applications for it, and continued to develop it, but it never caught on seriously. It was discontinued earlier in 2023.
Whether other wearable technologies will find the user base and use models that are enduring is an open question. Google Glass was a failure, but it’s possible it was just too ‘early’ and premature as a technology. The state of AI is much more advanced than 10 years ago, so the power and utility of such devices are potentially greater as well.