Bot Eat Brain
Posts
Anthropic launched its new Computer Use feature.

Anthropic launched its new Computer Use feature.

PLUS: Boost your productivity & creativity with AI

Michael Parrish
October 23, 2024

In partnership with

Good morning, human brains, and welcome back to your daily munch of AI news.

Here’s what’s on the menu today:

AI completes your tedious tasks 💻 🧠
Anthropic launched Computer Use.
1 intern, $1M damage, 0 regrets 🫣 📲
An intern got caught sabotaging TikTok’s AI.
Boost your productivity & creativity 🤖 🎨
Automate tasks, enhance decision-making, and more.

New here? Subscribe! 😎

Sponsor Bot Eat Brain and reach over 20,000 readers.

(Now 50% off) 🤯

Peep today's ‘What Would You Do?’ at the bottom. 👇

MAIN COURSE

1 AI just hacked your computer 💻 🧠

Yesterday, Anthropic launched Computer Use. It lets Claude 3.5 Sonnet control your computer, see your screen, move your cursor, and more.

Source: Anthropic

What can it do?

It can look at your screen, move your cursor, click buttons, and type text. It aims to mimic how a human uses a computer.

Is it perfect?

Nah. Even Anthropic claims it is “cumbersome and error-prone” and says it can miss quick actions due to its “flipbook” view of screens.

Will it do anything you ask it to?

Nope. Claude won't touch social media, election activities, domain registration, or government websites.

How can I take it for a spin?

It's now available in public beta through the API. Here’s how to use it.

Teach me more about Claude.

Sure. In August, I compared ChatGPT, Claude, and Grok-2 in a livestream. I tested the AI models’ skills across various beginner-friendly tasks.

I also still had long hair. 💇‍♂️

The next day, we covered Anthropic’s system prompts for Claude models. These are AI's basic instructions that shape its behavior/personality. Most companies keep this secret.

Last month, we reported on Anthropic’s Contextual Retrieval. It aims to significantly improve the performance of Retrieval-Augmented Generation systems used with Claude 3.

SIDE SALAD

How to lose millions with 1 intern 🫣 📲

Over the weekend, a ByteDance intern got caught deliberately sabotaging AI training runs.

ByteDance owns TikTok btw. 🤷‍♂️

Source: @YouJiacheng

What did they do?

Exploited vulnerabilities to insert delays, kill processes, and reverse training progress.

How sneaky were they?

Pretty bold. They even attended troubleshooting meetings to adjust their attacks and avoid detection.

Why'd they do it?

They got upset because GPU resources were diverted from their research project.

How much damage?

Millions of dollars worth.

Was this a typical intern mistake?

Nah. This was deliberate sabotage.

Teach me more about ByteDance.

Okey dokey. Last year, we covered TikTok’s AI model, Tako. It’s a chatbot that serves tailored videos & ads to TikTok users.

In January, we reported on Tiktok’s Depth Anything. It’s a depth perception model that achieved state-of-the-art performance.

In February, we covered TikTok’s Boximator. It’s a new motion control technique for AI-generated videos.

YOUR DAILY MUNCH

Cool Tool 🛠️

Mammouth

Mammouth gives you access to the best LLMs: o1 Preview, Claude 3.5 Sonnet, GPT-4o, Llama 3.1 405b, Gemini Pro, and Mistral Large.

Together with AI image generators: Midjourney, Flux Pro, DALL-E, and Stable Diffusion.

And more for only $10.

Learn more here.

Think Piece 🧠

How does an AI model “think?” How knowledge of token prediction and limited memory can help you better use AI for work and creative endeavors.

Which method is best when data is scarce? EPFL researchers discovered that in-context learning matches instruction fine-tuning performance for AI models.

Startup News 💰

IBM released Granite 3.0. It’s a suite of highly efficient 1-bit large language models tailored for enterprise AI applications.

Former OpenAI CTO Mira Murati is launching her own AI startup. She’s seeking over $100 million to develop proprietary models.

Research 👨‍🔬

World-Model-Augmented (WMA) — a web agent that uses AI models enhanced with world models to better predict the outcomes of actions during web navigation.

The Curse of Multi-Modalities — a comprehensive benchmark to evaluate hallucinations in large multimodal models across language, visual, and audio inputs.