Claude: The Sequel
I like to say that today's LLMs are "the worst they'll ever be." And we saw plenty of evidence of that this week, with multiple models getting significant upgrades.
What to watch:
Anthropic released Claude 2, with improved performance on language and programming tasks.
Google updated Bard with 43 more languages and new customization and productivity features.
While not yet official, Meta reportedly plans to release a commercial version of LLaMa for companies to fine-tune.
And a look at Google's Med-PaLM 2, a modified version of PaLM 2 designed to answer medical questions.
中国人工智能
While I tend not to cover news coming out of China (there’s only so much I can pay attention to), a tremendous amount of AI progress is happening overseas. This week, the Chinese government published the first version of its generative AI guidelines.
Why it matters:
The guidelines are looser than the draft rules published in April, signaling a more flexible stance from regulators. They go into effect on August 15.
Generative AI tools must register with the government and conduct security reviews. But there are no longer harsh financial penalties for companies that break the rules.
Overseas companies that target Chinese users are expected to comply, making the future of ChatGPT in China unclear.
Elsewhere in Chinese AI:
Alibaba unveils Tongxi Wanxiang, an AI-based image generator to compete with Midjourney and DALL-E.
Huawei releases Pangu Model 3.0, a new cloud-computing model for operations and product R&D.
JD.com launches ChatRhino, an LLM for enterprise users that's being tested internally and will be rolled out early next year.
Deal or no deal
OpenAI inked two deals this week to license content for model training - the AP signed a two-year deal for access to its text archive, and Shutterstock announced a six-year deal for access to its video, image, and music libraries.
Between the lines:
Today's text and image generation models have been trained on publicly available internet data, but that approach faces new challenges.
Multiple lawsuits, including new ones from Sarah Silverman and a class-action firm in California, allege OpenAI and others scraped data without user content and trained on copyright-infringing material.
And sites like Reddit and StackOverflow are beginning to require AI companies to license their content to train future models.
Elsewhere in no deal:
Hollywood actors go on strike and are aghast at a proposal for studios to own actors’ likenesses for future AI-generated content.
A Stability AI co-founder sues the startup after being “duped” into selling his 15% stake for $100 to the CEO.
The FTC sends OpenAI a demand for records about risks tied to its models and a March security incident regarding payment information.
Things happen
Elon Musk forms xAI to "understand the true nature of the universe." NotebookLM, an AI notebook. "Modern day Robert Oppenheimers" - Anthropic's obsession with AI safety. The last "Attention is all you need" co-author leaves Google. PoisonGPT, an "AI supply chain" attack. A playbook for creating and hosting your own AI companions. AI junk is starting to pollute the internet. CEO roasted after laying off 90% of support staff for AI chatbot. A look at the often overpaid and underworked contractors auditing Bard's answers. Meta unveils CM3leon, a 5x leaner image-generation model.
I'm not like super eager to play with Claude 2, but I do hear that it does certain things better than GPT4. That's not too hard to imagine, honestly.