GPT-4o fine-tuning is here

Howdy, wizards.

If you’re an entrepreneur, indie hacker or working in a startup, I warmly recommend checking out the interview with Pieter Levels (serial solo-founder currently focused on AI startups) on the Lex Fridman podcast that dropped yesterday. Lotta gold nuggets.

Dario’s Picks

The most important news stories in AI this week

‍

1. GPT-4o gets fine-tuning. OpenAI's flagship model, GPT-4o, just got a neat upgrade for developers. Using fine-tuning, developers can now train the model on their own data or examples, which can make it better for handling specific tasks. OpenAI shared examples of early testers which have built products powered by a fine-tuned version of GPT-4o, including Genie (AI software engineer) and Distyl (AI solutions for enterprises). The cost is 25$ per on million tokens, but is free until September 23.

‍

‎ Why it matters‎ ‎ This is likely to get a lot more businesses experimenting with fine-tuning, and we'll get better task/domain-specific AI applications. The offer to use the feature for free for the first month shows how hard AI companies are working to win over developers to use their tech; Google also announced they're giving out 1.5 billion tokens for free in the Gemini API.

‍

2. ‎How a 2 million token context window can help you automate stuff at work. Ever wondered how the ability of some AI models to process hours of video, thousands emails, dozens of reports, etc can be used to make work more effective? Here's an article on some practical ways a DevRel and UX research professional uses Google's Gemini 1.5 Pro (and it's loong context window) for that.

‍

‍

‎ Why it matters‎ ‎

‍

The author uses it to (spoiler alert):

1) analyse changes between versions of documents,

2) analysing & prioritising customer feedback,

3) turning written content into scripts for YouTube videos,

4) analysing lengthy videos of users trying to accomplish tasks to find friction points.

‍

My favourite has to be 2 & 4: analysing IRL feedback like this used to be a highly manual process, but can now easily be done at scale with LLMs.

‍

3. An AI bot runs for mayor in Wyoming. At least kind of. Candidate Victor Miller says that if he's elected, his AI bot (a GPT-based chatbot named VIC, short for Virtual Integrated Citizen) will help make the decisions. This is a first for US political campaigns, and the candidate isn't saying the bot will autonomously make decisions, but rather that it'll be a "hybrid approach" to Governance.

‎ Why it matters‎ ‎ I don't think Wyoming is ready for a custom GPT as its next mayor just yet. However, while the idea of AI bots single-handedly calling the shots politically seems pretty ridiculuos at this point, there's likely room for AI-assisted decision making in politics in the near future. AI has the potential to help make decisiosn by bringing a breadth and depth of knowledge no single person has accesss to, and in some cases more objective and balanced assesments.

‍

4. Airtable works like a spreadsheet but gives you the power of a database to organize anything. Imagine a super-organized toolkit that lets you update information in one place, and it automatically updates everywhere it’s used; huge time saver.

‍

Try Airtable for free (this is a referral link, I personally love Airtable and will get some free credits if you sign up).

‍