BB Digest: First agent of 2025

Digest #527 → Subscribe | Upgrade to Pro

Hey folks, another brief crazy season of AI is here. Last weekend saw the first agent from OpenAI and there’s much more lined up for this week. No worries, we’ll recap it here.

TLDR; inside today’s newsletter

• OpenAI's Operator launch, Canvas updates, and Meta's massive AI investment plans

• ICYMI: A deep dive into using AI effectively with Slack

• Highlights from our community conversations (and a new channel for builders)

• 6 of the best new AI tools including Pika 2.1 and Perplexity Assistant

• 5 interesting posts about AI model selection and historical research

• to-do’s

Alright, let’s get to it!

 🔎 Trends & news

• OpenAI’s first agent is called Operator. It can spin up a browser and do tasks on your behalf. A new version of 4o tuned for computer use powers Operator. You can take control from the agent if you want or it’ll let you know when it need your help (like for logging onto some website). Currently available to only ChatGPT Pro ($200/m) members in US only, Operator is a research preview. It can use Replit, kinda replace PMs but fail whenever you ask it something meaningful. Here’s a detailed review of Operator’s capabilities.

• ChatGPT Canvas got a long requested update from OpenAI (and something that Plus members can use). It now renders HTML and React Code plus works with OpenAI’s o1 model. The direction is clear—with o3-mini coming soon, ChatGPT wants to be the place for quick, on the fly apps (and make us forget Claude Artifacts)

• Claude gets better at citing sources. Anthropic's new API feature, Citations, tells you which parts of your PDFs and documents Claude referenced for its answer with pinpoint accuracy. It's available on Claude 3.5 models and is more reliable (and cheaper) than DIY prompts.

• DeepSeek is attracting conspiracies (and we don’t love them). The facts: R1 is an impressive model you can run locally, while their app raises the usual China-based security concerns. I liked these reads on the topic: Ben Thompson’s technical FAQ, Karpathy’s explainer on RL (secrete sauce of R1) and Scott Belsky on impact of cheap intelligence.

• Reid Hoffman has created a new drug discovery company, Manas AI with $24.6M in seed funding.

• Zuck is planning $65B investment for AI infrastructure and a data center of the size of Manhattan in 2025.

👀 ICYMI: New stuff we’ve launched

• Slack is where work happens for most of us and bringing AI to it can be confusing. There are many official and unofficial products—some that work, while many disappoint. We put together 4 clever ways to use AI with Slack and get more done.

💬 Inside the community this week

Become pro to join us building and learning together.

1. We opened a #build-apps channel to chat all things Cursor, Bolt, Replit AI, Lovable, Create etc. This is where people early in their building with ai tools journey are sharing their tips and helping one another. Alexander shared his tips for setting up new projects using these tools. (link)

2. Daniel is hosting an event on building reliable LLM apps at Serverless Toronto (link)

3. Wyatt shared links to a few good reading materials to understand the impact of DeepSeek AI’s models. (link)

4. Zhao introduced Memobase, an open-source project for user memory in GenAI apps. (link)

5. Aishwarya tried Rox, an AI sales tool, and shared her thoughts. (link)

6. Cameron built a GPT for education leaders to handle AI-related questions. (link)

7. Jorge is working on an AI-powered travel inspiration tool. (link)

8. Wyatt recommended the BRXND LA 2025 event for AI marketers in LA. (link)

9. Jesus's search for free Slack AI alternatives led to suggestions using Slack API with Cursor or Replit. Ben shared a ben’s bites blog post on building an automated Slack summary tool (which we use, for this!). (link)

10. Jason got recommendations for LLM courses in real estate starting with basic prompting. (link)

11. Frank sought advice on using AI for code refactoring from PHP to JavaScript. (link)

12. Paul and others discussed concerns about data storage on Chinese servers for apps like DeepSeek, with some suggesting alternatives like using it via Groq. (link)

13. Keshav discussed the potential of experimental models from Gemini, comparing them to OpenAI's offerings and highlighting Google's strengths in reinforcement learning. (link)

Join the conversation, plus full access to courses and workshops by becoming a pro member today!

⚙️ Top new tools

• AssemblyAI: Speech-to-Text API with unmatched accuracy and diarization.*

• Pika 2.1 - New video generation model from Pika Labs. Create 1080p videos with seamless motion and lifelike human characters.

• Perplexity Assistant - Allow AI to work with apps on your phone. (examples)

• Missio - Save hundreds of hours on repetitive tasks and workflows in product management.

• Wepost - Create on-brand marketing content for social media.

• Gemini Deep Research is available on Android now.

More tools here →
*sponsored

 📜 Interesting posts

• An updated guide for which AI you should use in early 2025.

• Google's NotebookLM will change my life-maybe yours too.

• I got OpenAI o1 to play codenames and results were suprisingly good.

• The leading AI models are now good historians.

• How AI is leading to a collapse of PMF.

📌 To-dos

1. Let us know if you have any tutorials or courses you want us to create—we’re always open to ideas.

That wraps things up for today! See you again on Thursday. 👋

Ben