DeSantis Uses Deepfake Images
In a new low for political advertising, an attack ad supporting Ron DeSantis uses AI-generated images of Trump embracing Fauci. As reported by AFP Fact Check, the deepfake gave away its non-human origins through noticeable text inaccuracies. It begs the question of how far the technology will move into political campaigns and the risks it poses.
Global Summit on AI Safety
The UK government announced it would be hosting the first major global summit on AI safety. The summit aims to bring together key countries, leading tech companies, and researchers to discuss, evaluate, and monitor the most significant risks from AI. Shortly after the announcement, Palantir's CEO, Alex Karp, argued that the West should stay ahead in the AI race.
Forbes article on Stability
Forbes recently published an article criticizing Emad Mostaque, the CEO of Stability, claiming that he has a history of exaggeration. The allegations are based on interviews with current and former employees, investors, collaborators, and colleagues. Emad responded to the Forbes report on his blog, describing the comments in the article as false accusations and misrepresentations.
Logic and Reasoning in Bard
Bard, an AI model by Google, is getting better at logic and reasoning, according to a company blog post. Using a technique called implicit code execution, Bard identifies prompts that might benefit from logical code and executes it 'under the hood.' This has led to approximately a 30% improvement on internal challenge benchmarks.
Conversations Around AI Risks
A conversation took place between two prominent AI researchers, Yoshua Bengio and Andrew Ng. Both agreed that it's crucial to articulate concrete scenarios where AI can lead to significant harm. Meanwhile, OpenAI's Sam Altman continued his world tour, answering questions about the challenges of competing with OpenAI.
Faster Sorting Algorithms
DeepMind published a paper in Nature outlining the use of Deep Reinforcement Learning in the development of AlphaDev. This new approach led to the discovery of smaller sorting algorithms that outperform human benchmarks. However, the work stirred some criticism within the community. A popular post on Hacker News called the improvements into question, and a Twitter exchange pointed out that GPT-4 can also identify the superfluous instruction removed in one of the algorithms (though it was unclear whether this resulted from a hallucination).
Why AI will Save the World
In an article titled "Why AI will Save the World", Marc Andreesen offers a rebuttal to fears about AI's potential dangers. His view? AI will not only not ruin society, but it may well save it. Andreesen outlines a plan allowing AI companies to build AI as quickly and aggressively as possible, focusing on mitigation strategies rather than technology bans.
Progressive Learning and Complex Reasoning with Orca
Microsoft Research has developed Orca, a 13-billion parameter model that imitates the reasoning process of Large Foundation Models. Orca has surpassed Vicuna-13B by over 100% in complex zero-shot reasoning benchmarks, edging ahead of ChatGPT on BigBench-Hard, although it lags behind ChatGPT on professional and academic exams. The details of this work are available in the paper "Progressive Learning from Complex Explanation Traces of GPT-4".
Uncertainty and AGI
Lauro Langosco argues in a LessWrong article that uncertainty about the future of AGI doesn't necessarily imply that AGI will be beneficial, and notes that the key question is rather whether AGI is a default-success or default-failure scenario.
Exploring Instruction Tuning
Despite recent claims that open models can match proprietary ones, a comprehensive evaluation of instruction tuning resources suggests more work is needed according to the study "How Far Can Camels Go? Exploring the State of Instruction Tuning on Open Resources". The work suggests further investment in building better base models and instruction-tuning data is required to close the performance gap.
Maths: To Crack or Not to Crack?
Ferenc Huszar raises the question: We may finally crack Maths. But should we? He reflects on potential dual-use developments, accelerated AI deployment, and the potential loss of meaning for mathematicians.
Enhancing Truthfulness in AI
Researchers from Harvard have proposed a technique to increase the truthfulness of large language models. The work, titled "Inference-Time Intervention: Eliciting Truthful Answers from a Language Model", shows promising results in enhancing AI truthfulness.
Resource Highlights
Two resources for this week: the Prompt Engineering Guide by Elvis Saravia, which gathers the latest papers, models, lectures, and tools related to prompt engineering, and the lighthearted, bite-sized machine learning videos from AICoffeeBreak with Letitia.
Samuel's Book Recommendation
This week, I'm recommending "The Hard Thing About Hard Things" by Ben Horowitz, an enlightening read full of strategies, anecdotes, and advice for when things get tough.
Project Spotlight: filtir
Lastly, I'd like to mention a project I'm involved with called filtir, focused on fact-checking ChatGPT outputs. We're developing new techniques to do this efficiently. If you’re interested, join the discussion on our Discord server.
If you prefer videos, you can find a video version of the news here: