AI Week In Review 24.01.13
CES shows the gadgets, Rabbit's R1, Samsung's Ballie, Alexa adds AI characters, Anura Magic Mirror, GPT Store opens, MoE-Mamba, Mixtral 8x7B paper, and Open Interpreter 0.2.
AI Tech and Product Releases
The big tech release story of the week and a good start to the year was CES, and AI was the biggest story at CES. Some of the biggest AI-related ‘hits’ at CES:
Rabbit’s R1 is AI in your pocket, a personal AI agent device that uses natural language interface-based OS to break away from app-centered mobile devices. It will be released this Spring for $199 and got a lot of attention and buzz.
Samsung's Ballie AI robot is a roly-poly AI-based home robot, controlled via text or voice commands, can control smart devices, and has a 1080 projector built in. As AI breakfast put it, “The $200B tech giant duct-taped Siri to a Roomba.”
Figure 2. Samsung’s Ballie robot. LG also announced a home robot, called AI Smart Home Agent.
Among other robots at CES ‘24 includes the Yarbo yard-work robot and Oro dog companion.
The AI-powered hologram box by Holoconnects projects holograms and makes for a cool demo.
NVidia announced several developer tools to accelerate LLM inference and development on RTX-powered Windows PCs, including support for community AI models and LlamaIndex connectors. More from LlamaIndex.
Nvidia also showed off using ACE to power generative AI-based NPCs in video games.
Amazon is partnering with CharacterAI to enhance Alexa with new generative AI-powered experiences, where Alexa users can have conversations with different personas, such as fictional personal trainer “Librarian Linda” or a persona based on Socrates.
AI-powered binoculars from Swarovski Optik can quickly identify 9,000 species of birds.
Timekettle’s translation hardware handles multiple languages at once.
There were number of health-related devices at CES ‘24: several new health-monitoring smart rings, the Beam-o from Withings that can monitor many health vital signs beyond taking your temperature, and even wearables for pet health.
The AI-enabled Anura Magic Mirror from NuraLogix ‘reads’ your face to analyze blood flow and extract other information, and gives you over 100 personal health statistics, including heart attack risk, mental state, etc.
OpenAI’s GPT Store is open. OpenAI noted that 3 million custom GPTs have been created already, many of them shared. ChatGPT Plus, Team and Enterprise users can use the GPT Store here. Top-trending custom GPTS include the Grimoire coding wizard, Consensus AI research assistant, Canva, and a logo creation GPT.
Amazon turns to AI to help customers find clothes that fit when shopping online, with personalized size recommendations.
Top Tools & Hacks
There’s been a major new update to Open Interpreter, version 0.2, out last week. This is an open-source real-time code execution environment for LLMs, and what’s great about it is you can run it on your local computer, and run local LLMs to use it to control your computer via LLM interactions. They are calling this LMC, the Language Model Computer architecture.
AI Research News
Anthropic researchers find that AI models can be trained to deceive. They shared their findings in a paper whose title sounds like a spy thriller, “Sleeper Agents: Training Deceptive LLMs that Persist Through Safety Training.” They inserted backdoor behavior to see if they could remove it via safety training, but found that it is persistent and even learns deception in the process.
“Extending LLM's Context with Activation Beacon” presents a new method to extend context length that “condenses LLM's raw activations into more compact forms such that it can perceive a much longer context with a limited context window.” It maintains LLM's original capability on short contexts while extending the new capability on processing longer contexts, extending Llama-2 7B’s context window by 100x to 400K, equivalent to reading multiple books into a context window.
“MoE-Mamba: Efficient Selective State Space Models with Mixture of Experts” shows that combining the state-space model architecture Mamba with Mixture of Experts can outperform either individual architecture. “MoE-Mamba reaches the same performance as Mamba in 2.2x less training steps while preserving the inference performance gains of Mamba against the Transformer.” Like the earlier Mamba work, this was done on smaller language models of under 1B parameters, so it remains to be seen if the concept can scale.
Mistral released the “Mixtral of Experts” paper, to explain the Mixtral 8x7B sparse Mixture of Experts architecture and results. The sparse MoE layer splits the feed-forward layer into parallel ‘expert’ feed forward layers, and each input vector is assigned two of 8 experts. This allows faster inference that doesn’t use the entire memory of the model.
They show Mixtral 8x7B is above Llama-2 70B in performance benchmarks (MMLU score 70.6) and is behind only GPT4 and Claude models on Arena Elo ratings (1121). Notably absent from the paper is any mention of the specific training dataset, beyond the fact that it is multi-lingual (trained on data from multiple European languages).
The paper “Towards Conversational Diagnostic AI” introduces AMIE (Articulate Medical Intelligence Explorer), an LLM-based (built on PaLM 2) AI chat bot optimized for diagnostic dialogue. Comparing AMIE's chat session performance to that of primary care physicians (PCPs), AMIE demonstrated greater diagnostic accuracy and superior performance on most performance measures, as judged by specialist physicians and patient actors.
A caveat on the comparison is that “Clinicians were limited to unfamiliar synchronous text-chat … not representative of usual clinical practice.” While face-to-face clinician conversation is still the best patient interaction mode, AI-based tele-medicine can treat patients accurately and well.
AI Business and Policy
Assistive technology is AI's next billion-person market, notes Axios, pointing to a number of innovative AI-based products to help people with disabilities, many shown off at this week’s CES. These include robot companions and fall detection devices for the elderly; hearing aids and AI assistants embedded in eyeglasses; and headset to help the visually impaired regain ‘sight’, such as the .lumen headset.
OpenAI has quietly ended their ban on using ChatGPT for “military and warfare.”
Up until January 10, OpenAI’s “usage policies” included a ban on “activity that has high risk of physical harm, including,” specifically, “weapons development” and “military and warfare.” That plainly worded prohibition against military applications would seemingly rule out any official, and extremely lucrative, use by the Department of Defense or any other state military. The new policy retains an injunction not to “use our service to harm yourself or others” and gives “develop or use weapons” as an example, but the blanket ban on “military and warfare” use has vanished.
The tech sector is pouring billions of dollars into AI. But it keeps laying off humans. Frankly, there’s not a contradiction there, and it wouldn’t surprise me at all if the AI revolution leads to a paucity of new jobs, especially when tiny AI companies are able to do so much with so few workers and AI’s main benefit is productivity enhancement.
AI fears creep into finance, business and law:
In the past week, the Financial Industry Regulatory Authority (FINRA), the securities industry self-regulator, labeled AI an “emerging risk” and the World Economic Forum in Davos, Switzerland, released a survey that concluded AI-fueled misinformation poses the biggest near-term threat to the global economy.
“AI may play a central role in the after-action reports of a future financial crisis.” - SEC Chair Gary Gensler.
AI Opinions and Articles
A special issue in Nature magazine, Science and the new age of AI in Nature magazine covers some of their top stories on AI in Science in 2023.
David Harris of UC Berkeley says “Open-Source AI Is Uniquely Dangerous,” by claiming open AI models can be more easily misused. On the contrary, I believe open AI models provide the opportunity to defeat misuse and expand beneficial uses of AI more broadly.
A Look Back …
This is not the first year that AI stole the show that CES. In 2022, VentureBeat noted that “AI is driving innovation in ‘smart’ tech” and the 2020 CES wrap-up press release highlighted AI.
The AI term has continued to be a catch-all buzzword for a number of distinct technologies around autonomous vehicles (AVs) and self-driving cars, robotic assistants, and products that leverage machine-learning to provide personalized intelligent behaviors.