Claude 3 Unveiled, Elon Musk vs OpenAI, Midjourney vs Stability AI, & Reproducing Sora

Anthropic unveiled the Claude 3 model family, Elon Musk sued OpenAI, Midjourney banned Stability AI employees, and there's a project reproducing Sora.

Anthropic Unveils Claude 3, A Trio of AI Models

Anthropic introduced the Claude 3 model family, featuring Claude 3 Haiku, Claude 3 Sonnet, and Claude 3 Opus. Each offers a unique balance of intelligence, speed, and cost to suit various applications.

With the now-available Opus and Sonnet models, and Haiku soon to follow, these models excel in a range of tasks including complex analysis, content creation, code generation, and multilingual communication. Claude 3 sets new standards with its impressive ability to understand and process vast amounts of information rapidly, delivering near-human comprehension and fluency across tasks. Moreover, Claude 3 boasts strong vision capabilities, reduced refusal rates, enhanced accuracy, and long context understanding with near-perfect recall.

Before showing Claude 3's incredible recall accuracy, let's revisit Claude 2.1, courtesy of Greg Kamradt.

As you can see, LLMs had trouble recalling information in the middle of a context window, and Claude 2.1 was no different.

Now consider Claude 3 Opus, which is all green, including in the middle.

That's a pretty remarkable change, and makes you wonder how much larger these context windows will become over time.

Elon Musk Sues OpenAI, OpenAI Responds

Ben Thompson of Stratechery provided an insightful summary of Elon Musk's lawsuit against OpenAI, complemented by OpenAI's response.

The lawsuit centers around three principal accusations: breach of contract, promissory estoppel (reneging on a promise that led to substantial action), and breach of fiduciary duty.

However, the contention arises from a misunderstanding of OpenAI's Certificate of Incorporation. Contrary to the expectation of open-sourcing all technology, the document specifies that the corporation intends to "open source technology for the public benefit" with the critical caveat "when applicable." This language introduces ambiguity, compounded by the absence of a formal contract.

This lack of a concrete contract, coupled with the conditional commitment to open-source technology, casts doubt on the success of Musk's legal challenge.

And perhaps Musk is not interested in winning the case, and instead hopes he wins the court of public opinion. OpenAI responded publicly to Musk's lawsuit, showing that Musk is being duplicitous, as he wanted OpenAI merged with Tesla in 2018.

But adding to the complexity, they shared an email from OpenAI co-founder Ilya Sutskever to Musk in 2016—merely two months post-OpenAI's inception—suggesting a flexible stance on sharing scientific advancements. Ilya implied that while complete transparency might be advantageous for recruitment in the short to medium term, it was "totally OK to not share the science."

This statement seems at odds with OpenAI's initial pledge to encourage widespread publication of research and the sharing of patents.

As the situation unfolds, the court of public opinion will undoubtedly play a significant role in shaping the narrative. The discrepancy between OpenAI's founding ideals and subsequent actions raises important questions about the balance between openness and strategic discretion in the rapidly evolving field of artificial intelligence.

Midjourney Bans All Stability AI Employees

Another AI war may have been overshadowed by Elon vs OpenAI. Pirate Wires reported that Stability AI employees were banned from Midjourney.

According to Nick St. Pierre, Stability AI employees caused a 24-hour outage by grabbing all of Midjourney's prompt and image pairs in the middle of a night on Saturday.

Nothing official has come from Midjourney or Stability AI, but there is enough supporting evidence suggesting this happened. How Midjourney can prevent others from trying this in the future, though, seems up in the air. Simply banning Stability AI doesn't address the root problem.

GitHub Repo Reproducing Sora

This GitHub repo aims to reproduce OpenAI's text-to-video model Sora. They’re currently building out the basic components and searching for interested open-source contributors to assist.