I’ve been asking myself these questions a lot lately and I’m sure we’ll hear more and more AI related questions from startup founders, who enquire about our startup development services, in 2024.
As a foundation for answering these questions here is some background on our brush with AI so far.
Github CoPilot for developer productivity
GitHub Copilot is an AI tool that improves developer productivity.Copilot is especially good for generation of code with repetitive patterns, such as test code. It boosts productivity by handling a great deal of mundane work, allowing our developers to spend more time focussed on higher-order problems such as code architecture and design.
ChatGPT for code generation / code correction
ChatGPT is extremely versatile as a development tool.We use it for work on top-down design patterns, such as organising different levels of abstraction within a project; exploring alternative code variations (e.g. I know X will work, but are there more effective or efficient alternatives); and translating between languages (e.g. I have code in programming language X, generate an equivalent in programming language Y).
Bard for code generation
Google’s Bard seems like the unloved step-child compared to the bullet train that is ChatGPT (at the time of writing). However, we’ve tested it and it has helped with some fairly basic tasks including:- Surfacing answers from API documentation (eg: we recently had an uncommon use case for the Docusign API and had many questions about their API). The results were accurate, helpful and were delivered faster than skimming through documentation.
- Generating low level HTML & CSS. We’ve used this mostly when creating HTML tables for use in Webflow websites. While writing the HTML for a table is exceptionally easy, it is a mundane task that can be eliminated with a simple prompt that saves 10 - 15 minutes of manual work.
Midjourney for image generation
We’ve recently started paying to use Midjourney for image generation so that we can quickly generate images for use in our startup web design projects. The results can be hit and miss but with a bit of learning and trial & error we found our prompts became better and thus the results too.
So far we have used images generated by Midjourney in 2 websites that we’ve designed. I’ve also used Midjourney to spark ideas for highly creative B2C website designs.
Unfortunately the delivery is a flat image so if the concept is good it still needs to be re-created by a human. I’m sure we’ll soon get to the stage where I won’t need to re-create concepts I generate using the AI … it will simply generate the concept and the code.
The image used at the top of this article was generated by Midjourney. The 2 images below were too. I struggled to choose which of these 3 images to use as the main image in this article so added them all.
I decided to use the more optimistic image in the hope that humans can keep AI safe and we lead it and it doesn’t lead us, or at the very least we work side-by-side.
The alternative is the more ominous theme portrayed in the 2 images below.
I’m sure you’ll agree that the quality is incredible.
We have also tested other image generation tools like Dall-e before. Currently Midjourney delivers better results than Dall-e for our specific use cases, but we’ll continue to keep an eye on Dall-e and all tools released by OpenAI.
Adobe background remover
This is really low-level stuff, but it makes life so much easier than the days of having to manually deep etch images in Photoshop to get rid of a background.If you need the background removed from an image simply upload it to Adobe background remover and abracadabra … background removed.
I’ve also tested Runway for a complex background removal task. I was able to get a better result than Adobe but it became a time sink. That’s no slight on Runway. The task was extremely complex and non-standard.
We’re keeping a close eye on Runway for all their AI magic tools which include image generation, video generation, motion brush and more.
Another tool in this space which is worth watching is Clipdrop by stability.ai which provides image generation and image editing tools.
Amazon Q
Amazon Q is new off the block, and still in preview, but we’re already using it to investigate ways for optimising AWS cost effectiveness for our clients.LLM API to determine email intent
We use an LLM API in a B2C startup Launch Lab built to determine whether emails the software receives are about a specific subject.This may seem simplistic but this feature is one of the biggest selling points for this startup. It helps solve a major pain for their customers and the accuracy of the results is exceptional.
While the API integration work was quick to do and not challenging from a technical perspective, the challenge was in the prompt engineering. We spent a long time manually typing prompts, and uploading emails, to the LLM’s UI to determine the optimal prompt that delivered the best results.
Once we had determined that, we then figured out the best way to reduce the amount of characters used in the prompt as well as in the response from the AI. This is important as the cost of your usage of the API is based on tokens, which are essentially tied to the number of characters. At scale, getting the AI to respond with a ‘y’ as opposed to ‘yes’ saves cost.
Machine Learning for Biotech
We’ve completed machine learning work for an Australian biotech company using their own extensive dataset. You can read more about the Tensorflow work completed on this project on our ML page.LLM API for email generation
While you can use LLMs to write blog posts, or most other forms of text, we use an LLM API with inquiries we receive through the contact form, on our website, to generate subject lines for return emails.Thus, when someone submits our contact form we use AI to ‘read’ the intent of their enquiry and generate text for the subject line of the email. When we reply to the enquiry, via email, a highly relevant subject line will have already been written for us.
Again, this is a fairly simple use case but saves us probably 10 - 20 seconds on every enquiry we reply to (approx. a couple of hours a year saved for a very small cost).
A 20 year AI journey
Our CTO (David Pretty) developed a fascination with computational neural networks 20 years ago during his doctoral research.He was captivated by the concept that a basic unit neuron, when scaled up with a combinatorially large number of connections and weights adjusted through back-propagation, could provide a vast space for the appearance of complex emergent behaviour.
However, at that time, neural networks weren't suitable for Dave's research, which involved using unsupervised machine learning to classify magnetohydrodynamic fluctuations in nuclear fusion plasmas.
This period also predated the significant advancements in GPU technology, which later revolutionised neural network training.
The advent of powerful, more accessible GPUs and the development of algorithms optimised for parallel processing enabled the efficient training of deep neural networks, marking the beginning of the deep learning renaissance which paved the way for today's AI technologies.
Dave has been closely following these developments, and now incorporates AI into client projects using both TensorFlow and PyTorch, as well as via foundational LLMs.
The type of AI integrated projects we’re hoping to work on in 2024:
- ChatGPT API for startups: similar to some of the API examples provided above, we’re certain that many of the startups we build next year will require some level of AI API integration to either improve the product or to speed up internal processes.
- AWS Bedrock integration: if you have data that you don’t want to share with the likes of OpenAI (ChatGPT) then AWS Bedrock is a great alternative. You can choose which LLM to use (Claude, Llama etc) and ensure that your company’s, and users, data remains within your AWS environment. While we foresee using Bedrock for a lot of the startups and scale-ups we work with, we're hoping to work with established companies too to implement AI solutions.
- We are always interested in other bespoke machine learning / AI projects, however the cost effectiveness of LLMs for startups often can't be beat.
Human vs AI written blog posts
An important fact about this article is that a human wrote approximately 90% of it. ChatGPT was used, in some areas, to expand from bullet points into structured paragraphs as well as for fact checking.It may or may not be the last time we write articles for our startup blog in this way. Future posts might simply be the musings of an LLM!
If you are using an LLM to write blog posts for your startup, for content marketing purposes, be sure to keep an eye on Google’s approach to AI generated content in search results.
Currently, Google states that they reward high quality content regardless if it was written by a human or an AI (Read what Google says about search and AI content).
If you do use an LLM to generate content for your blog remember that it still needs to be original and helpful to the user if you want SEO benefits.
If you’re blogging for thought leadership purposes then using an LLM, other than as an editor or co-writer, is potentially a slippery slope.
2024 and beyond
As a final aside, I can’t wait to read this blog post in 12 months time and see how rapidly it has aged. I’m sure the use cases written about here will seem basic, and potentially redundant, and Launch Lab as a business will offer a very different service to the current web development services we offer. Time will tell!Help using AI in your startup or product
If you’re a startup founder or established business and would like to discuss how you can leverage LLMs and machine learning in your own products, applications, or with existing data sets please contact us.