GPT-4 Released

+ Emergency London Hackathon

Hey friends 👋 ,

First things first, I'm hosting an emergency hackathon in London, UK for the release of GPT-4.

When you think of the London tech ecosystem, you think of a tortoise racing against cheetahs, a typewriter in a room full of smartphones, and a cup of tea at a high-speed espresso party.

I'm changing that.

Interested? Drop your details here. Let's build together.

For today's piece, we're diving under the hood of GPT-4 and how you can use it right now.

Today's Through the Noise is brought to you by The Diff.

Ever feel like you're lost in the tangled web of finance and tech?

It can be hard to "cut through the noise" (yes, I'll see myself out).

Meet "The Diff," a newsletter that's like having your own personal finance and tech sherpa, guiding you through the economic wilderness.

Posts range from in-depth company profiles to applied financial theory, strategy breakdowns, and macroeconomics.

That’s why the newsletter is read by founders, VCs, hedge fund managers, and more.

You can join 49,000 other readers here:

Leaving ChatGPT in the dust

GPT-4 / OpenAI

I've been testing GPT-4 for the last 20 hours. Needless to say, it demotes ChatGPT to a clear second place. This is what I've found so far.

First off, you can access GPT-4 if you subscribe to ChatGPT Plus ($20/mo). Whilst default GPT-3.5 outperforms GPT-4 on speed, I'm using GPT-4 from now on as it can solve difficult problems with greater accuracy.

Let's take a complex example. I'll ask ChatGPT to modify each word in a riddle, to begin with the letter "P".

GPT-3.5 Output

GPT-3.5 tried its best, but it didn't quite hit the mark. Firstly, it failed to make each word begin with the letter "P". Secondly, it condensed the riddle to a shortened version. It then went on to answer it—unprompted.

GPT-4 Output

GPT-4 is a different beast. Not only did it (almost) make the first letter of each word the letter "P", it followed the structure of the original poem. GPT-4 also provided additional reasoning after the riddle. It initially stated the answer but acknowledged that because the riddle is now changed, the answer could be different. This extra layer of reasoning and creativity is what makes GPT-4 superior to its precursor.


GPT-4 is a multimodal modal. This means it accepts both image and text inputs for text output. Here's an example of asking what's funny about an image—this really puts its reasoning capabilities to the test.

GPT-4's ability to pick up on nuances—and humour—is completely mindblowing.

What's more, it can now iterate with you on creative writing tasks: composing songs or writing screenplays. The best part about this is learning your writing style, something which GPT-3.5 struggled with.

Wild Stats

Here are some crazy stats about GPT-4 so far:

  • It can now handle 25,000 words

  • GPT-4 scored in 90th percentile for the uniform bar exam (ChatGPT: 10th percentile)

  • GPT-4 scored in 99th percentile for the biology olympiad (ChatGPT: 31st percentile)

  • It is 82% less likely to respond to requests for disallowed content

  • It is 40% more likely to produce factual responses than GPT-3.5

What are your favourite use cases? Reply to this email and let me know.

Let me know what you think

Login or Subscribe to participate in polls.

That’s all for today friends!

As always feel free to reply to this email or reach out @thealexbanks as I’d love to hear your feedback.

Thanks for reading and I’ll catch you next for some Sunday Signal.


If you liked this piece, subscribe below: