Housekeeping
I'm going to be at the AI Engineer World's Fair next week! Leave a reply/comment if you're also attending!
We hit 10K subscribers! I'll have a longer recap next week, but there’s 50% off paid subscriptions through the end of the month. Thanks for reading!
The new new Claude
Yesterday, Anthropic released Claude 3.5 Sonnet, the first in a new and improved family of Claude models.
Why it matters:
The model handily beats out Anthropic's previous flagship model, Claude 3 Opus (and even nudges out GPT-4o on some benchmarks). It's also twice as fast and 80% cheaper.
It accepts text and images and can work with Artifacts - a beta feature that gives Claude a dedicated sandbox to run code, edit documents, and preview designs.
But perhaps most importantly, it's proof that we're still seeing improvements in frontier models - if not in terms of raw reasoning power, then certainly in terms of capability and speed.
Elsewhere in model milestones:
China's top-valued AI startup, Moonshot AI, recently launched both a role-play chat app and a music making app in the US.
Runway unveiled Gen-3 Alpha (a model capable of creating highly realistic 10-second video clips), while a new AI video app Dream Machine has been going viral on Twitter.
Nvidia announced Nemotron-4 340B, a family of models designed to generate synthetic data for training other commercial LLMs.
And while not a new model, it feels worth mentioning that Ilya Sutskever (former cofounder and Chief Science Officer of OpenAI) has a new company: Safe Superintelligence Inc.
Perplexing behavior
Perplexity, an AI-powered search engine, has found itself in (more) hot water with internet publishers, with new accusations of deceptive scraping practices.
Between the lines:
Last week Perplexity was accused by Forbes of plagiarism via its new Pages feature. Now, it seems the company's web crawlers misrepresent their origin and violate scraping best practices.
Wired also reports that despite marketing itself as a search engine, Perplexity is fully capable of hallucinating answers, the same as other leading LLMs.
The AI space is still navigating the legal gray area around training and copyright. Some firms like OpenAI have struck licensing deals, but other proposals imagine new business models that can reward content creators and publishers.
Elsewhere in AI anxiety:
A study of 10 AI chatbots, including Perplexity and ChatGPT, found that they repeat Russian disinformation and cite fake local news sites.
Hacker News asks: Could AI be a dot com sized bubble?
And Ted Cruz has proposed a new bill that would make social media platforms liable for deepfakes.
Left brain, right brain
It's been over a year since Alphabet merged its Google Brain and DeepMind teams into Google Brain. Since then, there have reportedly been some growing pains as the new outfit races to keep up.
Between the lines:
What's made matters worse is that Google is trying to move from being mostly research oriented to having a more product and commercial focus - not an easy thing for many researchers.
But despite Gemini’s snafus, DeepMind continues to put out research breakthroughs - such as AlphaFold’s DNA improvements and new math research.
More recently, the company has published its experiments with AI generating comedy for professional comedians and generating soundtracks and dialogue from video footage.
Elsewhere in the FAANG free-for-all:
Apple has held talks with Chinese tech giants to try and offer Apple Intelligence overseas, as ChatGPT is banned in China.
Snapchat previewed a new model that will turn prompts into custom Lenses.
Meta detailed how it maintains large-scale AI capacity in a technical blog post.
And TikTok is expanding its Symphony ad suite with AI-generated avatars based on popular creators.
Things happen
A look at tech companies' investments in AI startups. SoftBank’s new AI makes angry customers sound calm on phone. Inside the players and politics of the modern AI industry. Apple may not be bringing Apple Intelligence to the iPhone 15. "I will fucking piledrive you if you mention AI again." AI took their jobs, now they get paid to make it sound human. How AI is revolutionizing drug development. AIs are coming for social networks. With Microsoft's acquisition, a DeepMind cofounder gets a look into OpenAI's code. AI images in Google search results have opened a portal to Hell. Japan goes light on AI regulation to court investment. Users jailbreak AI video generator to make porn. Rebind, a startup that turns authors' voices into AI reading companions. Black founders are creating tailored AIs, like ChatBlackGPT. Making my local LLM voice assistant faster and more scalable with RAG. AI in finance is like ‘moving from typewriters to word processors.’ The R in RAG stands for "royalties." AI experimentation is high risk, high reward for low-profile political campaigns. A generation of AI guinea pigs. Software companies may be getting left behind in the AI boom. LLMs that can call multiple tool APIs with one request. 8 UK train stations scanned unwitting passenger faces with AI for two years. Turning the tables on AI. The limits of LLM scaling based on human-generated data.