. . AI: Wizards, Dreams, Fears, and Perplexity

Author

Marshall Kirkpatrick
January 04, 2024

Friends, today’s another day of big AI developments. I’ve surfaced the 6 news stories that the data and my editorial judgement deem most interesting. If you want to watch me walk through these stories in 6 minutes, I’ve posted a video on LinkedIn.

I hope these stories are interesting and valuable to you. And remember, only a third of 2000+ surveyed AI researchers (38%, according to our 2nd story today) put at least a 10% chance on extremely bad outcomes, like human extinction!

On the other hand, take a look at Dave Sifry’s amazing new startup Questy.ai. I asked it this morning if it might find me a good alternative to Simon Wardley’s super helpful “pioneer/settler/city planner” model of organizational requirements for innovation, instead in the field of ecological succession (thanks, Christopher Daradics). Questy summarized 7 research papers in 30 seconds to output a really clear answer to this question I’ve had for years! I then took its answer to ChatGPT to ask for a direct replacement of terms and visualization of the three types of functions you need in an innovation community, now using the thinking provided by Questy. Check it out below. Thanks AI and friends!

Ok, now here are today’s top AI stories.

First impacted: Investors, AI technology developers
Time to impact: Short

AI firm Perplexity says it has raised $73.6 million in Series B funding "led by IVP with participation from NVIDIA, NEA, Bessemer, Elad Gil, Jeff Bezos, Nat Friedman, Databricks, Tobi Lutke, Guillermo Rauch, Naval Ravikant, Balaji Srinivasan." I will confess that of the 4 queries I've sent to Perplexity in the past 24 hours, the results were: bad, great, and two so-so outputs. I’ll keep trying. [wsj.com] Explore more of our coverage of: AI Funding, Perplexity. Share this story by email

First impacted: AI researchers, Risk management professionals
Time to impact: Medium to long

A survey of 2,778 AI researchers, performed by researchers from AI Impacts, the Universities of Bonn and Oxford, forecasts that "If science continues undisrupted, the chance of unaided machines outperforming humans in every possible task was estimated at 10% by 2027, and 50% by 2047. The latter estimate is 13 years earlier than that reached in a similar survey we conducted only one year earlier." On the other hand, "over a third of participants (38%) put at least a 10% chance on extremely bad outcomes (e.g. human extinction)." [aiimpacts.org] Explore more of our coverage of: AI Milestones, Job Automation, AI Risk Research. Share this story by email

First impacted: Robotics Engineers, AI Researchers
Time to impact: Medium to long

Google DeepMind has launched three systems, AutoRT, SARA-RT, and RT-Trajectory, with the aim of enhancing robot data collection, speed, and generalization. Ultimately, they say, so you can have a personal robot assistant that will cook and clean for you. Over a seven-month period, the AutoRT system reportedly managed to safely control up to 20 robots at once in various office buildings and collected a wide-ranging dataset from 77,000 robotic trials across 6,650 tasks. [Shaping the future of advanced robotics] Explore more of our coverage of: Google DeepMind, Robotic Systems, Robot Training. Share this story by email

First impacted: AI developers, Data scientists
Time to impact: Short to medium

The creators of WizardLM's language model, WizardCoder-33B-V1.1, claim that it has outperformed competitors such as ChatGPT 3.5, Gemini Pro, and DeepSeek-Coder-33B-instruct in several coding tests. WizardML is quietly based out of Microsoft and aims to build AGI. [WizardLM/WizardCoder-33B-V1.1 · Hugging Face] Explore more of our coverage of: AI Language Models, Coding Competitions, Data Contamination Checks. Share this story by email

First impacted: Animation Artists, Video Game Developers
Time to impact: Medium

A new system called DreamTalk has been developed that its developers, researchers at two Chinese universities and Alibaba, claim can generate expressive talking heads with enhanced lip movemement precision. The system uses diffusion models and includes a denoising network, a style-aware lip expert, and a style predictor to create high-quality, audio-driven facial movements across various expressions. In case you want to make videos of someone saying something they never said. In the last paragraph of the 17 page research paper, the developers say the videos will be watermarked to prevent misinformation. [Paper page - DreamTalk: When Expressive Talking Head Generation Meets Diffusion Probabilistic Models] Explore more of our coverage of: AI Technology, Deepfake Prevention, Audio-Visual Synthesis. Share this story by email

First impacted: AI Researchers, AI Developers
Time to impact: Medium

A study has identified 32 techniques aimed at mitigating the issue of hallucination in LLMs, a challenge that hinders their safe deployment in critical areas such as medical record summarization and financial analysis. The research, which includes methods like Retrieval Augmented Generation and Knowledge Retrieval, also presents a detailed classification of these strategies and discusses the inherent challenges and limitations, providing a foundation for future research in this field. [A Comprehensive Survey of Hallucination Mitigation Techniques in Large Language Models] Explore more of our coverage of: LLM Hallucination, Mitigation Techniques, AI Safety Research. Share this story by email

That’s it! More AI news tomorrow.