Welcome to this week's AI news roundup. Let's dive in!
Italy's ChatGPT Ban Disrupts Output
A recent study from Monash University reveals the consequences of Italy's ChatGPT ban on coding output. By comparing Italy with other European countries, researchers estimate a 50% decrease in GitHub releases in the two days following the ban. Additionally, VPN searches increased by around 50%, and Tor bridge usage increased by 9.4%. The authors note that while these measures may be well-intended, they often lead to short-term disruptions in output.
China Drafts Generative AI Rules
China has released draft rules governing the use of generative AI. As noted by Helen Toner, companies providing access to generative AI via APIs like those from OpenAI and Google are to be responsible for all content produced. This stance seems “appealingly simple but is likely to face practical hurdles”.
Stability AI Launches StableVicuna and DeepFloyd IF
Stability AI has released StableVicuna, the AI world's first open-source RLHF LLM Chatbot, and DeepFloyd IF, a text-to-image model that can integrate text into images.
GPT-3.5 and Psychiatry
Researchers from the Max Planck Institute and the University of Tubingen subjected GPT-3.5 to psychiatric testing. They found that GPT-3.5 responds robustly to a common anxiety questionnaire, producing higher anxiety scores than human subjects, and shows a strong increase in biases when prompted for anxiety-inducing text.
Scaling Transformers with Recurrent Memory
Researchers have scaled transformers to 1M tokens and beyond with Recurrent Memory Transformer. The recurrent memory mechanism allows the transformer to retain information across 2 million tokens on various tasks.
Hyperbolic Image-Text Representations
Research proposing the use of hyperbolic embedding space for image-text representations indicates that it may be better suited for capturing hierarchy than the Euclidean space used by CLIP embeddings.
Chameleon: Compositional Reasoning with Large Language Models
Chameleon, a method for plug-and-play compositional reasoning with large language models, shows strong results on science question answering and tabular mathematical reasoning benchmarks.
Visual Instruction Tuning and LLava
Visual Instruction Tuning introduces LLava, a large language and vision assistant. LLaVA is capable of explaining complex queries about images.
Physician vs. ChatGPT Responses
An article in JAMA Internal Medicine compares physician and ChatGPT responses on medical questions posted on Reddit. Professional evaluators preferred both the quality and empathy of ChatGPT responses.
Quick News Round-Up
SperBank announces GigaChat, a Russian-focused chatbot alternative to ChatGPT.
HuggingFace releases HuggingChat, an open-assistant chat model.
Microsoft reportedly working on its own AI Chip, codenamed Athena, since 2019.
StackOverflow plans to charge companies to train on its data, following Reddit's lead.
Palantir announces Artificial Intelligence Platform or AIP for defence, a multimodal chatbot that communicates with commanders, suggesting courses of action and then executing those commands if the operator approves them.
Samuel-API is released - a biological competitor to GPT-4 for text-completion.
RedPajama 2.8 B, an open-source replication of the LLaMA training dataset, continues to make progress.
AI legend Geoffrey Hinton leaves Google, voicing concerns about AI risks in a NYT article.
Tools Round-Up
NVIDIA open-sources NeMo Guardrails library, aiming to keep chatbots on topic and improve their security.
WhisperX is an open-source library for automatic speech recognition with word-level timestamps and diarization.
SEEM, an interactive segmentation tool supporting various prompts, releases its code.
If you prefer video summaries, you can find a video version of the newsletter here:
Acknowledgement: GPT-4 helped with editing.