. . AI: GPT-4V Makes Better Websites Than Humans (3.6.24)

Elon Musk, AI Education, Front-end Engineering

Author

Marshall Kirkpatrick
March 06, 2024

Friends, today's edition covers more dirty laundry from the OpenAI and Elon Musk saga. More interestingly, new research showing AI models are very capable at front-end engineering (often beating the original reference websites provided). Conveniently, for those of you who might be adversely affected by this, our final story is about a new training to uplevel your tech skills in open source AI.

These are the stories being discussed most by the AI community, summarized each day like we do.

Speaking of the AI community, newsletter reader Eric Norlin is putting on an AI conference this year and it will probably be the kind of thing people talk about with joy a decade from now, as many of his conferences are. Called SW2 Con (“devoted to the next wave of software development”), early bird tickets are on sale now. Tell them “impact10” sent you for 10% off.

Now here’s today’s news,

Marshall Kirkpatrick, Editor

First impacted: Front-end developers, web designers
Time to impact: Medium

Generative AI has shown promising developments in front-end development by converting visual designs into code, according to a recent research paper. The study, which tested models such as GPT-4V and Gemini Pro Vision on a benchmark set of 484 diverse real-world webpages, found that GPT-4V was deemed superior by human evaluators, with 64% rating the webpages it generated as better than the original ones. "We hypothesize it is possible that the model has more access to modern and popular webpage design principles such that it can automatically improve the original design based on these best practices." Despite the successful performance, the study identified areas for improvement as well. And to be fair, the tests were performed on static, not dynamic and interactive, webpages. [Paper page - Design2Code: How Far Are We From Automating Front-End Engineering?] Explore more of our coverage of: Generative AI, Front-End Development, GPT-4V Model. Share this story by email

First impacted: Open AI, Elon Musk, Open Source Community
Time to impact: Long

A blog post authored by OpenAI's executives refutes claims in the lawsuit brought by Elon Musk and instead offers the other side of the story in which Elon actually offered OpenAI to merge with Tesla. The original lawsuit by Elon was that OpenAI breached the contract by becoming a for-profit entity. OpenAI says the emails imply that a for-profit pathway was on the table all along. Click the link to see the redacted emails, if that's what you want to spend your time doing. [OpenAI and Elon Musk] Explore more of our coverage of: OpenAI, Elon Musk, AI Development. Share this story by email

First impacted: Prompt engineers, AI developers
Time to impact:

Amanda Askell, a member of the technical staff at Anthropic, provided a breakdown of the 'System Prompt' for Claude. System Prompts are specially crafted inputs designed to instruct or guide the model in performing specific tasks or generating content in a desired format or context. Interesting read for those interested in the prompt engineering space. [via @AmandaAskell] Explore more of our coverage of: Anthropic AI, Bias Correction, Political Sensitivity. Share this story by email

First impacted: AI Researchers, Data Center Managers
Time to impact:

Reka AI says its experience training large language and multimodal models has been a "hardware lottery" due to variations in the performance of different compute providers' hardware. They reported unexpected differences in performance, even within the same type of hardware, with some clusters found to be ineffective due to frequent breakdowns and file system issues. The post also highlights the benefits of using TPUs instead of GPUs and the advantage Google has in this space. [Training great LLMs entirely from ground zero in the wilderness as a startup — Yi Tay] Explore more of our coverage of: AI Hardware Performance, Reka AI, Compute Providers. Share this story by email

First impacted: AI developers, Job Seekers
Time to impact:

The Hugging Face team have launched a short course on how to use open source models from the Hugging Face Hub. The course, which is available for free for a limited time, covers things like automatic speech recognition and text to speech conversion, and how to package code into easy-to-use, cloud-based apps using Gradio and Hugging Face Spaces. [Open Source Models with Hugging Face] Explore more of our coverage of: Hugging Face Hub, AI Education, Cloud-Based Apps. Share this story by email

That’s it! More AI news tomorrow!