AI News in October 2023: Disney Robotics, Creative Tools and Synthetic Driving Data

This October 2023 roundup covers an expressive Disney robot, Adobe’s AI products, voice dubbing, developer costs, image provenance, satellite internet and synthetic driving data. Some items are reports or future plans as of the video, not completed launches or current product guidance.

Disney Research in Zurich created a small robot designed to move with personality. I like how its expressive motion recalls WALL-E. The team used mostly 3D-printed modular hardware and actuators, with engineers, designers and animators contributing to its development.

Disney used a reinforcement-learning-based pipeline to balance an animator’s intended expression with robust movement. I can imagine such robots entertaining visitors at Disney parks and perhaps becoming smaller toys someday. Those are my possibilities, not announced product plans.

Adobe announced updates to Firefly, including additional training data and controls such as referencing an image’s style. It also showed text-to-vector generation in Illustrator and AI features for Express. These are the capabilities described in 2023; availability and features can change.

Adobe previewed Project Stardust, showing objects and shadows selected, moved or removed with the background filled in. I welcome tools that make photo editing easier. I describe Firefly content as copyright-approved, but neither a training-data policy nor a product claim guarantees that every output is free of rights issues.

I say Adobe acquired Figma, but that was incorrect: at the time the roughly $20 billion transaction was proposed and had not closed. Adobe and Figma mutually terminated the agreement in December 2023. Figma is a collaborative design platform, and my speculation about future integration was not an announcement.

ElevenLabs introduced AI dubbing that it said could translate content into 29 languages while retaining voice characteristics and handling multiple speakers. I think this could help creators reach audiences beyond their original language. A generated voice is a synthetic imitation, not the original person actually speaking the translated words.

I compare the approach with human dubbing used by creators such as MrBeast and speculate about cost and the effect on voice-acting jobs. Those outcomes are predictions, and the actual quality and suitability of a dub would need to be assessed for each use.

Amazon launched its first Project Kuiper internet-satellite prototypes, entering a field where SpaceX’s Starlink already had a large constellation. At the time Amazon planned a network of 3,236 low-Earth-orbit satellites, production launches in the first half of 2024 and customer beta testing by the end of that year. Those were 2023 plans, not proof of coverage or deployment.

Reports said OpenAI planned developer updates intended to reduce costs and add vision capabilities. As an AutoGen user, I had found GPT-4 agent experiments expensive and encountered rate limits; I say a few iterations could reach $10. That is my experience with a particular setup, not a general price for all agent workflows.

Reuters also reported that OpenAI was exploring its own AI chips and evaluating a potential acquisition. I discuss GPU scarcity, possible benefits of custom silicon and the time needed to develop chips. These were reported possibilities, not a completed acquisition or a guaranteed price reduction.

I discuss then-rumored NVIDIA RTX 4080 variants and possible pricing, as well as Samsung’s announced next-generation mobile processor. Samsung claimed 1.7 times faster data processing and 14.7 times the AI performance of its predecessor. These figures and the phone integration were claims or expectations from 2023, not independent measurements here.

Meta researchers introduced Stable Signature to embed an invisible watermark into a generative model’s decoder so that outputs can be traced to a model version. I support provenance tools because synthetic images can mislead people. The researchers evaluated robustness against particular modifications; this does not guarantee the watermark survives every edit or attack, or that detection is infallible.

Reports also described Meta’s AI sticker generator producing objectionable images despite keyword filters. I compare this with moderation in other image-generation systems and worry about exposure in messaging products used by children. No filtering method demonstrated in this segment should be assumed to prevent every problematic output.

Researchers from the National University of Singapore presented Show-1, a text-to-video system combining pixel and latent diffusion. The team reported competitive results and a 15 GB versus 72 GB inference-memory comparison with a pixel-based model. That comparison is specific to the tested models and conditions, not a general claim about every commercial generator.

Reporting on Books3 raised authors’ concerns that books included in a dataset were used to train AI without authorization. Lawsuits included allegations about training data; a court order requiring models to be destroyed was my speculation, not a stated outcome. I also worry that valuable datasets may become less accessible as their owners restrict access.

Wayve introduced GAIA-1, a generative model for driving scenes that could help create synthetic training data. I compare approaches to real-world driving data and mapped robotaxi service areas, and express my preference for Tesla’s approach. The video does not establish that any system can safely handle every unfamiliar road, or that other automakers cannot compete.

Synthetic driving scenes could expand the range of situations available for training, but I am unsure how well generated data will transfer to real driving. I worry that repeatedly generating derivative data may have limits, while recognizing the potential to augment real-world data. Its safety and effectiveness require testing.