Data Machina #188
GPT research, hacks & startups. Advanced Multimodal ML. SoTA Anomaly Detection. Google FLAN-T5 XXL on AWS Sagemaker. GPT in 60 lines of NumPy. SoTA Speech-T5 in Hugging Face.
More GPT Research, Hacks & Startups. Lots of interesting stuff happening in many AI/ ML areas. But it seems anything related to GPT is getting the main attention.
Every week, I come across some fascinating research around the Generative Pre-trained Transformer (GPT). A few examples:
Theory of Mind (ToM) is the ability to attribute mental states to ourselves and others; a foundation for social interaction, empathy and self-consciousness. In this paper, Michal @Standford_Uni claims that GPT-3 (davinci-003), solved 93% of ToM tasks, a performance comparable with that of nine-year-old children.
A team of researchers @MetaAI and @PompeuFabra_Uni explained that [GPT-J] Toolformer can teach Itself to use tools, and achieve zero-shot performance across a variety of downstream tasks, while competing with much larger LLMs.
A group of researchers @Massey_Uni developed Chat2Viz, a system that generates Data Visualisations via natural language using ChatGPT, Codex and GPT-3. Pretty cool.
A team from @MaxPlanck_Institute researched why GPT-3 fails miserably in causal reasoning tasks. Paper: Using Cognitive Psychology to Understand GPT-3.
Still it’s unclear why and how GPT models generate some weird output. Researchers @SETA_MARIS found a set of mysterious, anomalous tokens which result in a previously undocumented failure mode for GPT models. Which brings me to…
…Jailbreaking ChatGPT. Since Microsoft has now entered the chat room, ChatGPT and New Bing ChatGPT are becoming more AI-Fair, RLHF-controlled, and AI-Safe.
What happens next? People loving to prompt-inject and jailbreak anything involving Open AI GPT models and MS New Bing ChatGPT.
Marvin, a student @CDTM_Munich shared what seems to be be the whole original prompt and rules of Microsoft Bing ChatGPT which apparently is codenamed Sydney. I hear that he was later kicked out from MS New Bing ChatGPT.
Vaibhak tried a token-smuggling attack on New Bing ChatGPT, and interestingly it seems that New Bing ChatGPT -unlike Open AI ChatGPT- generated and adversarial prompt against the malicious jailbreak attempt. Some people speculate that New Bing ChatGPT is powered by GPT-4.
A cool fad now is a jailbreak prompt called DAN (Do Anything Now), which includes a token-based system that punishes ChatGPT for refusing to answer questions. It’s pretty wild and clever. And evolving fast. Five days ago someone presented DAN Version 6.0 in Reddit.
Along the lines, many people are sharing their wildest ChatGPT conversations at ShareGPT hoping that they can trigger new jailbreak prompts and hacks.
And there are some startups that are assembling red teams that pay people to jailbreak GPT and other LLMs.
Back in December, @acheong08 showed how he reversed engineered ChatGPT API.
Prakhar built an open source tool that enables users to work around the limitations imposed by ChatGPT.
GPT and money, money. Three important Qs when using GPT family models are: 1) Which specific GPT model should I use for what? 2) How much is this going to cost me? and 3) How long will it take to run this GPT model? Allistair @BuildT wrote a nice post on this. Use GPT-3 incorrectly: reduce costs 40x and increase speed by 5x.
GPT Startups. The number of startups building apps (features?) on the back of GPT models family is absolute madness. Checkout GPT-3 Demo: A showcase of +600 GPT use cases & apps. Some startups are also starting to build ancillary apps around the GPT models family:
Helicone: Meaningful log analytics for GPT
HumanLoop: Make GPT-3 faster, cheaper, more effective
LlamaHub: Custom data sources and data loaders for GPT Index
Fixie: Extend ChatGPT with endless capabilities and interfaces
It’s a wild zoo of AI apps out there! See for example AI Search Tool: a directory with literally 1,000s of AI startups, apps, and etc mostly built on top of GPT models. How crazy is that?
Are these startups building a “feature,” or “a product” or a “biz model” on top of GPT models? It seems to me that most of them are building “features” but not building “biz models” with repeat use cases, and true value generation. Will many of these startups go belly up? Ummm… This reformed VC says that these days AI Looks Like a Bubble
Have a nice week.
10 Link-o-Troned
the ML Pythonista
the ML codeR
Deep & Other Learning Bits
AI/ DL ResearchDocs
Expert Language Models vs. Instruction Tuning (paper + code)
The Self-Supervised Learning Interplay: Insights for Practitioners
El Robótico
data v-i-s-i-o-n-s
DataEng Wranglings
AI startups -> radar
ML Datasets & Stuff
Postscript, etc
Tips? Suggestions? Feedback? email Carlos
Curated by @ds_ldn in the middle of the night.