- The Daily Update
- Posts
- Anthropic discovers how AI thinks
Anthropic discovers how AI thinks
Insight: Tech companies struggle to turn AI hype into profit
Welcome back!
We have lots of big news on deck today. Anthropic is conducting groundbreaking research in hopes of understanding how AI thinks, and a new GPT-4 competitor has hit the market for free.
We’ll also take a look at why companies are struggling to turn AI hype into profits. Let’s go.
In today’s Daily Update:
🗞️ Tech companies struggle to turn AI hype into profit
🤖 LLaVA challenges GPT-4
📸 Anthropic wants to understand how AI thinks
🚨 AI Roundup: Three quick hits
Read time: 2 minutes
🗞️ Tech companies struggle to turn AI hype into profit
Source: Ideogram
The AI boom has presented consumers with a plethora of new tools that can generate everything from computer code to press releases, but companies are still figuring out how AI products will generate a profit.
The details:
Generative AI tools require robust computing power and expensive chips to operate. Building and training AI products can take years and hundreds of millions of dollars.
According to the Wall Street Journal, Microsoft has lost money on several of its generative AI products, including its popular GitHub Copilot for programmers.
The chief executive of Amazon Web Services, Adam Selipsky, says customers are unhappy about the cost of running AI models.
High costs can also be attributed to the strain caused by overuse of potent AI models. Essentially, using GPT-4 to summarize an email is like delivering a pizza in a Ferrari.
The response:
Microsoft is going with a higher price for its next AI software upgrade. The company will charge an additional $30 a month for the AI-infused version of its Microsoft 365 suite.
Google, which is releasing a similar AI assistant feature for its Workspace apps, will also charge $30 a month on top of the regular subscription fee.
Tech companies hope that higher monthly charges will cover the cost of running AI technology.
Microsoft has also been exploring less powerful and cheaper AI tools for its Bing search engine.
The bottom line: Companies and consumers will soon have to decide how they want to use AI and what they are willing to pay for it.
🤖 LLaVA challenges GPT-4
Screenshot
Researchers from UW-Madison, Microsoft and Columbia just launched LLaVA, a new open-source AI model that rivals GPT-4 for visual and language understanding.
Key points:
LLaVA combines CLIP (an AI system that understands images) with LLaMA, another open-source LLM.
The model can understand and have conversations about images.
It’s open-sourced and completely free to use.
LLaVA’s code and training data are freely available online.
50-word review: If you’ve already found an LLM that suits your needs, LLaVA probably isn’t going to take its place. Still, this is a strong model despite having far less training data than GPT-4.
📸 Anthropic wants to understand how AI thinks
Source: Ideogram
Anthropic has developed a new method to interpret the individual neurons inside LLMs like Claude, helping researchers understand and decode AI models’ reasoning.
Key points:
Neurons typically respond to many unrelated concepts, making them hard to understand in isolation.
Anthropic found a way to separate groups of neurons into simpler “features” with clearer meanings.
The team decomposed a layer with 512 neurons into more than 4,000 features, which separately represent capabilities like legal language, HTTP requests, Hebrew text and much more.
Anthropic hopes this research will eliminate a “serious roadblock” to a mechanistic understanding of language models.
Why it matters: This breakthrough makes it easier to understand what’s happening when AI is thinking on a small scale. As Anthropic scales this approach up to larger models, it will enable improved control over AI systems. This is a huge development from an AI safety perspective.
🚨 AI Roundup: Three quick hits
Source: Ideogram
Fear Factor: ‘Godfather of AI’ tells ‘60 Minutes’ he fears the technology could take over humanity.
Revenue Race: ChatGPT’s mobile app hits record $4.58 million in monthly revenue.
Heating Up: AMD to acquire AI software startup in effort to catch Nvidia.
How was today’s newsletter? Your feedback helps me create better emails for you. Thanks for reading The Daily Update!
Jack
(P.S. If you want to share The Daily Update with a friend or colleague you can find it here.)