Ben's Bites

Is China really behind in AI? Deep Seek AI just replicated OpenAI’s method of getting models to think before answering. The resulting models even beat the o1 models at some benchmarks. What's going on here? Chinese AI companies are shipping open-source models that match or outperform western counterparts (closed/open) despite GPU restrictions. What does this mean? DeepSeek's new R1-Lite model matches o1-preview's performance on AIME & MATH benchmarks—both tests of complex reasoning and mathematical ability. More importantly, like o1, it shows its thinking process in real-time, and performance improves as the model spends more time reasoning. Unlike o1 though, DeepSeek plans to make this completely open source. Why should you care? If you're using AI from web products like ChatGPT, Claude or using their cloud APIs, stick with it. The product/developer experience is worth it. But if you’re even thinking about local deployment, DeepSeek and Qwen models (from Alibaba) should be at the top of your mind. Read our tutorial on using local models with LMStudio: https://lnkd.in/dmYQsufK

Wordware (YC S24) is building the tools to really make prompting the next programming language They just raised a $30M seed round and are already being used by Instacart and Runway. (We're an investor) More info: https://lnkd.in/dy4Tn6t7

Top new AI tools we discovered this week

Microsoft's Ignite event dropped some numbers that caught our eye: Nearly 70% of Fortune 500 companies use Microsoft 365 Copilot now They also announced "Copilot Actions" - think of it as IFTTT but for your Microsoft workspace. Set up custom automations right inside your usual apps. https://lnkd.in/d_4nNKjy

Gemini finally got a feature to save information about you (they had a prototype when it was called Bard—RIP) It’s called “Saved Info” and is similar to Memory in ChatGPT But unlike ChatGPT’s auto-saved memories, you explicitly tell Gemini what to remember about you It’s less magical, but we are cool with it—it doesn’t inflate expectations Gemini Release Updates: https://lnkd.in/dqdHirpF

Stripe launched an agent toolkit It will allow AI agents to handle payments, billing, and financial tasks The toolkit also integrates with major AI agent frameworks making it easier for developers to build AI apps that can charge users based on metered usage, process refunds, and handle billing More info: https://lnkd.in/dDvJTffV

Mistral AI just copied ChatGPT's homework and made it free Their chat app now has multiple things that ChatGPT Plus offers—web search, image generation, agents (like GPTs) and even a Canvas feature for co-editing with AI Thanks to their new 124B model Pixtral Large, it also understands documents (with images and charts too) You can try it at chat.mistral.ai (plus there’s API and weights access to the model for devs and researchers) Full story: https://lnkd.in/dV2bumKy

YouTube is testing restyled soundtracks for Shorts Creators in the test group can now use AI to reimagine official songs into unique 30-second Shorts soundtracks They can change the genre or mood while keeping original vocals, with clear attribution to the source song More info: https://lnkd.in/dumDh2bC

?? Top new tools from this week

You might have noticed ChatGPT search had an underwhelming response We believe this because people are using it as a standalone search engine The better way to use it is to remove the back-and-forth to a Google search tab in your workflows But how? We got you: Learn how to master new software using ChatGPT Search Struggling with a new tool? Instead of drowning in scattered documentation and outdated forum posts, use ChatGPT as your personal tech mentor. It can search, understand, and explain technical steps in plain language—turning that intimidating learning curve into manageable steps. Step 1. Gather your learning resources Start by asking ChatGPT to search for relevant examples and documentation about the software you want to learn. Be specific about your business context while keeping the language natural. ChatGPT will find and compile resources from across the web, including visual examples. Step 2. Get personalized guidance - Describe your specific needs in plain language - no technical jargon needed - Share screenshots of features you like from example implementations - Ask ChatGPT to explain the thinking behind different approaches - Get click-by-click instructions customized to your requirements - Share your progress and get help with troubleshooting - Request simpler alternatives if something seems too complex Tip: Don't worry about using precise technical terms. ChatGPT understands the context from your descriptions and screenshots. Phrases like "that dropdown thing" or "the coloured labels" work just fine. This walkthrough is a condensed version of this tutorial: https://lnkd.in/dK4eGreg Check it out and see how easy it is to prompt AI to master any new tool