Is Meta's AI really open-source?

Welcome to the latest edition of AI Odyssey!

In just 621 words, we've packed essential AI insights, breakthroughs, trends, and strategies tailored for founders and executives.

Latest: xAI's next model drops in August, reportedly using 20,000 H100 GPUs for training. Grok 3 is rumored for year-end, potentially needing 100,000 H100 GPUs.

1 big thing: Is Meta's AI really open-source?

Why it matters: Meta’s chief AI scientist, Yann LeCun, claims Meta’s large language models (LLMs) are open-source. Some praise this move, while others call it misleading.

The big picture: Critics argue that Meta’s models are open but not truly open-source. One commenter even called it "marketing spin for data laundering." This debate is more than semantics; it impacts how these models are perceived and used, especially in academic circles.

Between the lines: The European Union’s AI Act exempts open-source models from certain regulations. If classified as open-source, developers disclose less about their model's development. The Open Source Initiative (OSI) is working on a clear definition for open-source AI, focusing on criteria like free redistribution and source code availability.

Zoom in: True open-source requires transparency and peer review. Many AI models, including Meta’s Llama 3, have restrictive licenses and lack detailed training information. OSI hasn’t approved Meta’s license, suggesting it doesn’t meet open-source standards.

Digging deeper: Radboud University researchers propose a 14-criteria framework to grade models as open, partial, or closed. Meta’s Llama 2 scored as closed due to lack of source code, while BigScience’s BloomZ was graded open. This highlights a trend of overselling AI openness.

The Transparency Index: Stanford’s Foundational Model Transparency Index evaluates AI models on 100 indicators. Meta scored high on basic transparency but low on data and labor transparency.

The bottom line: Stanford’s Percy Liang emphasizes the need for truly open models with accessible training data and code for scientific and research integrity. As the debate over open-source AI continues, the industry must ensure clarity and trust to avoid misleading claims.

Do businesses really need GPT-5?

💬 The GPT-5 obsession: Rumors swirl weekly about GPT-5, the next big model from OpenAI, potentially under a new name. But at Eric Newcomer’s Cerebral Valley conference, a contrarian view emerged: Silicon Valley might be overhyped about the next generation without fully leveraging today’s models.

Focus on the present: At the AI Engineer World’s Fair, web developer Simon Willison highlighted how current models like ChatGPT have unresolved issues that don’t need GPT-5 to fix. He cited usability problems with ChatGPT, such as difficulties processing tables and diagrams in PDFs. Surprisingly, taking screenshots of these images and using optical character recognition can solve these issues.

Practical enhancements: Instead of speculating about future capabilities from OpenAI, Google, or Anthropic, the emphasis should be on maximizing the performance of existing models. There’s significant potential to improve today’s AI applications with clever engineering tweaks and practical solutions.

The bottom line: Before getting swept up in the hype for GPT-5, businesses and developers should focus on what can be achieved with current models. There’s still a lot of untapped potential in the AI tools we already have.

Prompt of the day:

Generate a LinkedIn bio

Craft a compelling LinkedIn bio that showcases my expertise as a personal branding expert. Highlight my key skills in [List 3-5 key skills]. Share notable experiences or accomplishments that demonstrate my value. Express my goals for networking, job search, and personal brand building.

💣 Latest AI hits: quick recap

• Clay, an AI-powered sales and marketing startup, raised $46 million in Series B funding led by Meritech Capital, with participation from First Round and Sequoia Capital.

• Runway, an AI-powered image and video generation startup, has been speaking with investors about raising $450 million at about a $4 billion valuation.

• AI deals lift US venture capital funding to highest level in two years, data shows.

• OpenAI’s ChatGPT Mac app was storing conversations in plain text.

• Google will now generate disclosures for political ads that use AI.

• Meta ordered to stop training its AI on Brazilian personal data.

• Apple’s Phil Schiller is reportedly joining OpenAI’s board.

That’s a wrap! See you again for a fresh dose of AI Odyssey soon. 😎