AI Roundup 057: Claude 3
March 8, 2024.
Housekeeping
I'm hosting a reader meetup in San Francisco with Timothy Lee from
on March 25th! If you're near SF and want to hang out, RSVP here: https://lu.ma/0p3mc5bx.
Claude 3
Anthropic released Claude 3 this week, the latest version of their chatbot. If you're interested in a deep dive into what's changed, check out Wednesday's post.
TL;DR:
Claude 3 comes in three sizes: Haiku, Sonnet, and Opus. Prices range from "cheaper than GPT-3.5" to "more expensive than GPT-4."
All three models have a 200K token context window, vision and document support, and can speak several languages.
The Opus model beats GPT-4 on various benchmarks and has near-perfect recall on a "needle in a haystack" test.
Elsewhere in foundation models:
Stability AI released a research paper on the technology behind Stable Diffusion 3, including its Multimodal Diffusion Transformer (MMDiT) architecture.
Inflection AI launched Inflection-2.5, a new model powering the latest version of its chatbot, Pi.
And Haiper, an AI video startup founded by two DeepMind alums, dropped its first text-to-video model.
Google got Dinged
On Wednesday, the DOJ charged Linwei Ding, a former Google software engineer, with stealing AI trade secrets while secretly working with two China-based companies.
Why it matters:
The US government has, for many years, accused China of intellectual property theft. It was one of the major sticking points of President Trump's trade war.
One incredible example is from 2013, when Huawei engineers stole an entire arm from a secret T-Mobile robot named Tappy.
And when it comes to AI, US officials have taken steps to prevent both IP and hardware from getting to China. This week, AMD was once again stopped from selling its AI chips in the Chinese market.
Elsewhere in the FAANG free-for-all:
Microsoft launched Copilot for OneDrive, which can fetch and summarize your files.
Google reported that Gemini Nano won't be coming to the Pixel 8 due to "hardware limitations," but plans to bring the LLM to other devices shortly.
And Facebook is working on an AI system for video recommendations across all its products.
Begin forwarded message
After Elon Musk filed a lawsuit against OpenAI last week, the company fired back with a blog post countering his arguments, complete with redacted emails from Musk himself.
Between the lines:
Outside the company, some seem to think the case doesn't hold much merit. Inside the company, at least one exec has attributed the lawsuit to sour grapes.
I'm stating the obvious, but this won't be the last lawsuit against OpenAI. Sam Altman told employees, "the attacks will keep coming."
And the Wall Street Journal looks at the long breakdown of Altman and Musk's relationship.
Elsewhere in AI anxiety:
Studies continue to find copyright infringement and election misinformation problems with leading AI models.
Researchers created a generative AI worm that can spread between AI agents.
And Microsoft filed a motion to dismiss the NYT's copyright suit, accusing it of "doomsday futurology."
Things happen
Public trust in AI is sinking across the board. India changes course, requires government approval for new models. Apple is right not to rush into AI. “AI will cure cancer” misunderstands both AI and medicine. AI startups require new strategies. Elliptic curve 'murmurations' found with AI. "AI, no ads please": 4 words to wipe out a trillion dollars. AI prompt engineering is dead. Adobe finds AI hype is a double-edged sword. Optimizing technical docs for LLMs. Training LLMs from the ground up. TikTok spammers and the AI tools that enable them. Amazon's shopping assistant is mostly useless. Cloudflare's Firewall for AI. Can AI therapists do better than the real thing? OpenAI's CTO played a role in Altman's ouster. Something like fire. How AI is being used to authenticate paintings. An interview with the CEO of Groq, the AI chipmaking startup. How LLM-powered games could kill the dialogue tree.
