BB Digest: Secrets of o1 revealed

Digest #525 → Subscribe | Upgrade to Pro

Hey folks, my Twitter feed is full of memes from Trump’s inauguration and AI folks going crazy over Deepseek’s new model. The second one got my attention and I have some thoughts in the newsletter today.

TLDR; inside today’s newsletter

• SOTA open source AI from China, AI pricing updates from Google/Microsoft

• ICYMI: top Claude features and making landing pages with AI

• snack n skill - learn to customize AI writing to your brand voice

• 6 new tools including LinkedIn's Job Match

• Quick recap on how local AI models are finally becoming practical

• 5 interesting posts including the environmental impact of AI

• to-do's

Alright, let’s get to it!

 🔎 Trends & news

• DeepSeek AI’s new reasoning model DeepSeek R1 matches OpenAI’s o1 in benchmarks and is open-source. Their paper reveals a ton of secrets behind reasoning models. Another interesting bit: they also use this R1 model to improve other small open-source models from the Qwen (by Alibaba) and Llama (by Meta) family. This can make local AI models finally viable. (more thoughts below)

• Google increased the price of its workspace plans ($12 → $14) but bundles AI features for free. On a similar beat, Microsoft is relaunching Copilot for Business with a free chat plan and pay-as-you-go agents. We’ll reach AGI when it fixes these terrible naming decisions.

• In ChatGPT land, the Tasks UI is finally working for me. They’ve also upgraded how you set custom instructions with more presets. I’m scared that GenZ is an option 💀.

• Adobe adds Bulk Create in Firefly - With two editing options “resize” and “remove background” available in beta, it can edit thousands of images in a single go.

👀 ICYMI: New stuff we’ve launched

• AI tools are adding new features every day. Some are useful, others not that much. We put together 4 features in Claude that you might have missed (with tutorials to use each one of them).

• Landing pages are your gateway to your potential customer’s mind. AI now helps you make one faster and improve the ones you already have.

🍽️ Snack ‘n’ skill

Learn how to create content in your brand voice using Claude Styles

Ever struggled to make AI-generated content match your brand's tone? Claude's new Styles feature lets you customize its responses to match your voice, but only if you know how to use it right.

Step 1. Get started with preset styles

Access styles by clicking the "Choose style" button at the bottom of your prompt input in Claude. Try out preset options like Concise, Explanatory, and Formal to see how they alter Claude's responses for different content types.

Click "Create & Edit Styles". Instead of describing your style, add 3-4 writing examples that consistently represent your desired tone. Keep examples similar - multiple styles work better than one catch-all style.

Step 2. Create your custom brand style

• Click "Create & Edit Styles" then "Create Custom Style" to begin customizing

• Choose "Add Writing Example" and paste 3-4 samples of content written in your target style

• Ignore the automatic preview tests - they rarely match what you need

• Jump straight into a new chat and test your style with a simple prompt

• Fine-tune by selecting "Edit style manually" from the Options menu

• Keep testing and tweaking until Claude consistently matches your voice

Tip: Create separate styles for different content types (emails, social posts, documentation). The more focused your examples, the better Claude will match your style.

This walkthrough is a condensed version of this tutorial. For more AI writing tutorials, check out our content creation collection.

⚙️ Top new tools

• Job Match by LinkedIn - Find how the qualifications from your profile and the job description match up.

• Frames by RunwayML - Improved style control in your image generations, available now in their unlimited plan.

• /extract from Firecrawl - Get structured data from any website with just a prompt, without needing a web scraper.

• Stella by FastTracker - Automate tasks like meeting invites, emails, to-do lists, reminders, and note-taking with AI.

• Windsurf Wave 2 - Memories, search and other improvements to the rising coding editor by Codeium.

• Browser-use - Make websites accessible for AI agents by extracting all interactive elements.

More tools here →

🗣️ Everyone’s talking about… local AI models

I have long kept the belief that local AI models are not usable for practical work. The two big reasons for this have been:

1. They are much dumber than the models behind ChatGPT or Claude.

2. They are too massive to run on consumer hardware—think your Macbook Air.

But a series of improvements in the open-source AI space over the past few months has solved both of these problems to the extent that local AI is now viable.

What's going on here?

Open-source AI models (esp. ones from China) match state-of-the-art performance while being cheap and laptop-friendly.

What does this mean?

There’s a ton to unpack when it comes to local models. So we need some guiding principles:

• Models with 7B parameters can run on older MacBooks. 14B and 32B is a stretch. 1.5B models can run on today’s powerful phones.

• I’ve found GPQA Diamond to be a reliable benchmark (vs maths or coding benchmarks) for understanding a model’s intelligence and general world understanding.

Now, look at the performance of these small models compared to the latest closed-source models. The 7-8B models match the original GPT-4o performance on GPQA Diamond. They aren’t too far from Claude 3.5 Sonnet too.

Why should I care?

This means you can have near-frontier intelligence on your laptop. The next big hurdle is making these models usable for regular folk. A lot of people want to use AI to work with their data, but many hold back because they’re scared about how secure their data is once it’s uploaded. A good local model would solve that problem. I’m keen to see someone build a good product that uses these local models instead of OpenAI or Google’s APIs.

 📜 Interesting posts

• Solving 64.6% of issues (current SOTA) on SWE-Bench-Verified with o1.

• Using ChatGPT is not bad for the environment.

• Amazon races to transplant Alexa’s ‘brain’ with generative AI.

• Replit’s move to serve first-time coders with Agent in the wake of AI.

• The arrival of AI is prominent in the work change report by LinkedIn.

Here’s Why Over 4 Million Professionals Read Morning Brew

• Business news explained in plain English

• Straight facts, zero fluff, & plenty of puns

• 100% free

See for yourself

📌 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