These 6 things happened in AI today

Happy Friday everyone. It’s another day of ground-shaking philosophical debates and historic break-throughs in AI! I won’t comment otherwise but if you’re in the mood for some prompt experimentation, check out this new paper titled Metacognitive Prompting Improves Understanding in Large Language Models. It explores how asking your AI to do things like clarify its understanding, critically assess its initial responses, and discuss its level of confidence all make a big difference. It worked well in my experiments this morning.

And now today’s stories:

"Large Language Models Understand the World," Claims Andrew Ng

Andrew Ng, one of the world's utmost AI experts, believes Large Language Models (LLMs) like Othello-GPT understand the world by constructing complex models, not just replicating words. Critics argue that LLMs only generalize from data, lacking true understanding or reasoning. The debate centers on whether LLMs' predictive responses equate to real understanding. [DeepLearning.ai]

Waymo Secures Paid Driverless Permit in California

Waymo has obtained a paid driverless permit from the California Public Utilities Commission (CPUC), allowing it to charge for rider-only trips in San Francisco soon. Its autonomous service, Waymo One, already offers over 10,000 weekly rides in San Francisco and Phoenix. There have been no reported injuries or collisions involving pedestrians or cyclists in its first million miles. This permit is a key milestone in the autonomous vehicle industry's transition from prototypes to a market-ready product. [Waypoint - The official Waymo blog: Waymo’s next chapter in San Francisco]

AI + Expertise = A lot of Y Combinator Startups]

Paul Graham notes a trend among Y Combinator startups using a AI to address persistent issues in specific domains. Many experts across various fields are using AI to solve these problems for the first time. Graham suggests the increase in AI startups is a response to now solvable problems, not a fleeting trend. [via @paulg]

Shortage of NVIDIA GPUs prompts Nvidia GPU optimization

A Reddit post describes an effort underway from a group called the Machine Learning Compilation to make AMD GPUs work better for running large language models. This is big because NVIDIA GPUs are very hard to get right now. The team is using ROCm software to get models running on AMD cards like the Radeon RX 7900 XTX. They say they have made the RX 7900 almost as fast as top NVIDIA cards like the RTX 4090 and RTX 3090Ti for these models. They also enabled the Vulkan system so the models can run on other AMD devices like the SteamDeck gaming handheld. The goal is to make AMD hardware a good option for large language model inferencing when NVIDIA cards are scarce. [Reddit]

ChatGPT Responses Found Inaccurate, Yet Preferred Over Stack Overflow

A study shows 52% of ChatGPT's responses are inaccurate and 77% verbose. "Nevertheless, users still prefer ChatGPT's responses 39.34% of the time due to their comprehensiveness and articulate language style." This highlights the need for ChatGPT's error correction and user risk awareness. [Who Answers It Better? An In-Depth Analysis of ChatGPT and Stack Overflow Answers to Software Engineering Questions]

StackOverflow Adopts Semantic Search Solution

StackOverflow explains in a blog post how it has improved its search function by incorporating Weaviate, an open-source platform, for semantic search. This uses a pre-trained BERT model from SentenceTransformers to create embeddings for natural language searches, boosting efficiency and result quality. The updated system combines the benefits of semantic and lexical search. [Ask like a human: Implementing semantic search on Stack Overflow]

That’s it! Know someone who could use these news summaries? We’d love for you to share them!

thanks,

Marshall