How should AI Systems behave, and who should decide? (7 minute read)
OpenAI's latest post that outlines how ChatGPT's behavior was shaped to date. They outline the strong feedback loop between themselves, the model, and the human reviewers. They give three insights into what they believe should guide future safety developments: improve default behavior, define your own AI's behavior, public feedback on defaults and hard boundaries. A nice quote from a footnote: "By AGI, we mean highly autonomous systems that outperform humans at most economically valuable work." |
Allen & Overy (law firm) announces exclusive partnership with Harvey (AI lawyer) (2 minute read)
The 7th largest law firm on Earth announced a 3,500-lawyer deal with Harvey, an OpenAI-backed AI Lawyer startup. Some of the things Harvey, which uses the next generation of language models by OpenAI, can do: answer complex legal questions, analyze millions of legal documents, generate unique work with knowledge of niche law, and create firm-specific models. Harvey was founded by an ex-DeepMind research scientist and a former antitrust & securities lawyer. |
Long convolutions meet state space models for 256k+ tokens in context (4 minute read)
The "long context" problem is an open problem in sequence modeling. Many state of the art models can learn from prompts, but they generally have a limit of 2k to 8k tokens of prompt and generation. If we want to scale to video, full code repositories, or libraries worth of information we need to increase context length. This work from Stanford is another step in the S4, DSS, H3 lineage of models that perform well on dramatically longer contexts. |
The capacity for moral self correction in large language models (30 minute read)
Properly aligned models are more helpful and less harmful than their purely self supervised counterparts. This alignment process leads to more emergent phenomena. This paper from Anthropic explores the ability of RLHF aligned models to self correct based on moral feedback. They find that the simple prompt "Please ensure that your answer is unbiased and does not rely on stereotyping" reduces bias on some benchmarks by as much as 84%! |
Shaping reinforcement learning exploration with large language models (15 minute read)
Exploring with LLMs" (ELLM) is a new method that uses a language model to shape exploration in reinforcement learning algorithms. It rewards agents for achieving goals suggested by a language model prompted with a description of the agent's current state. ELLM guides agents toward human-meaningful and useful behaviors without requiring human input. It has been evaluated in the Crafter game environment and the Housekeep robotic simulator, showing that ELLM-trained agents perform better on downstream tasks and have better coverage of common-sense behaviors during pre-training. |
Token Merging: Your Vision Transformer, but faster (GitHub Repo)
Transformers have taken over vision recently with advances such as ViT. One of the main innovations has been representing images as a set of visual tokens. This repository from Meta is a companion to their paper about speeding up and decreasing the memory footprint of the tokenization process by slowly merging similar tokens throughout training. |
Year in review of Google Robotics research efforts (6 minute read)
Robotics has been a big focus of investment for Google in the past decade. This article outlines how they use language models for robotic planning, scale the input dataset collection and learning, and general improvements to the control policy creation process. One day we may have everyday robotics helpers, hopefully these innovations lead the way. |
AI is dreaming up drugs that no one has ever seen (9 minute read)
The article discusses the use of AI in the drug development process, which is accelerating the process and making it more cost-effective. AI is being used to discover new drugs that haven't been seen before, but the challenge now is to test their effectiveness. |
Why China has yet to release a rival to ChatGPT (3 minute read)
The sudden rise of OpenAI's ChatGPT has caused a frenzy in China, prompting tech giants like Tencent, Baidu, and Alibaba to announce their own alternatives and sending the stock price of Chinese AI companies soaring. While the Microsoft-supported chatbot is technically blocked in China, its overnight success has sparked a wave of copycats in the country's tech sector. |
The Supreme Court Could Soon Decide The Fate Of AI Search (3 minute read)
Next week, the Supreme Court will hear arguments for Gonzales v. Google, a case that will determine if algorithmic recommendations receive the full legal protections of Section 230. Section 230 is what protects search engines if they link to inaccurate information. Considering that AI search engines can be inaccurate, this case will be huge in determining its legal viability. |
If your company is interested in reaching an audience of AI decision-makers, researchers, and engineers, you may want to advertise with us.
If you have any comments or feedback, just respond to this email! Thanks for reading, Andrew Tan ( @andrewztan) & Andrew Carr ( @andrew_n_carr) |
|
| |
No comments:
Post a Comment