AI News in July 2023: Llama 2, Generated Video and Model Behavior

In a July 2023 news roundup, Matthew Berman discusses AI-generated entertainment, reports about Apple and OpenAI, company uses of language models and Meta’s Llama 2 release. His examples, estimates and predictions describe the news as he understood it at the time.

Berman opens with the writers’ strike and concerns about AI replacing creative work. He shows several AI-video examples, including work using WarpFusion and Runway Gen-2, and is impressed by their visual consistency compared with earlier short clips.

He predicts that content production could move closer to individual viewers, allowing highly personalized stories. At the same time, he says he has not yet seen AI reliably reproduce the emotion, suspense and empathy needed for compelling storytelling.

He highlights a Fable Studio paper and a demonstration that imitates South Park-style animation using AI for visuals, voices and story. He is impressed by the production but remains skeptical that the example achieves gripping storytelling. He plans a separate discussion of the paper.

Turning to Apple, Berman discusses reports of an internally tested AI effort code-named Ajax. He speculates that deep integration with Apple’s ecosystem could encourage adoption. His comments are a 2023 prediction, not confirmation of a product roadmap or user count.

Berman discusses a Stanford and UC Berkeley study comparing March and June 2023 versions of GPT-3.5 and GPT-4 on selected tasks. He says results changed across math, sensitive questions, code generation and visual reasoning, with some improvements and some declines.

He attributes perceived declines to a tradeoff between safety measures and model quality. That is his interpretation; the comparisons shown do not by themselves establish that safety changes caused the differences or that every capability declined.

He points to changes in the rate of answers to a sensitive question and to lower scores on the code-generation measure shown in the video. These are results for particular prompts and evaluation methods in a historical comparison, not a general measure of current model quality.

Berman recounts reporting that OpenAI delayed broader image-recognition features over privacy and security concerns, including face identification. He also describes Be My Eyes, where volunteers help blind people interpret their surroundings, and the potential for AI assistance alongside concerns about misinterpreting people’s faces.

On financing, he recounts a report that Hugging Face was raising $200 million at a $4 billion valuation. He explicitly says the round was still in progress when he recorded, so it should be read as a report at that time rather than a completed transaction.

Berman cites Stripe CEO Patrick Collison’s description of an internal language-model tool with prompt sharing, privacy controls and configurable models. Collison said roughly one-third of Stripe employees were using it weekly. Berman speculates about the underlying provider and sees a business opportunity in privacy-aware internal tools.

He then describes Petals, a project for distributed inference in which different devices contribute portions of a large model. He compares the distribution idea to torrents and notes support for Llama 2; his suggestion that this could form an exceptionally powerful network is a prediction, not a measured result.

Finally, he recaps Meta’s Llama 2 launch: base and chat variants in 7B, 13B and 70B sizes, and a reported increase in training data over the first Llama. He notes that a 34B model was not released at that time.

Berman highlights commercial use under the Llama 2 license and mentions its 700-million-user threshold. That threshold is not a substitute for the full license terms. He expects many fine-tuned versions to appear and points to his separate testing video.