Early in 2023, we decided to delve into the field of artificial intelligence because we sensed it would become very big. Traditionally, we had always been very conservative, when picking new technologies, but this one gave us a special feeling we couldn't describe at the time.

Pretty early on, our clients requested help in AI, and without even publishing it on our website until this year, we landed a few contracts through word of mouth or from existing clients.

I wanted to give a small recap of how this first year of AI has played out for us. 18 months in AI is an entire lifetime, but we'll dive into this later on in the post.

Is AI here to stay?

That AI is transforming our lives is undeniable. While most businesses and VCs are squandering money left and right pouring millions into expensive generative AI tools and copycats of copycats just to generate more automated content for their marketing purposes, others are silent killers.

Fake news and deepfake videos continue to proliferate on social media, with bots increasingly menacing networks and messaging apps, becoming more credible each day. Banks and payment platforms are leveraging AI to detect fraud and suspicious transactions with greater efficiency. Militaries worldwide are training AI-powered drones to identify and target moving objects to shoot the shit out of them with surgical precision. Financial and credit institutions are using AI to optimize portfolio management, amassing larger sums daily and impacting the global macro-economy.

According to Monterail: 40% of businesses already use AI. But are they really using it for something useful? How many of them are using it to make profit? How many will still use it in 12 months from now or - more importantly - how many of these businesses will still exist 12 months from now?

What have we tried?

Well, it'd be more accurate to ask "what haven't we tried?" but it would be an extreme understatement. The exaggerated amount of tools popping up every day is overwhelming.

Back in late 2022, we spent a lot of time playing with generative AI. Midjourney, Stable Diffusion, DALL-E and the like. As it happened to pretty much everyone, there weren't many useful use cases for that. We mostly used them to generate fun images for our internal keynotes and company update presentations, before they became acceptable to be used in blog posts. And still, they're not as good as they should be, but they're good enough.

A one-of-a-kind service like ChatGPT has done wonders for our company, though. I, for one, have been able to upgrade myself by having it as a sidekick. I use it daily for data transformation, generating placeholders for projects, debating ideas with myself and mostly to discard potential side-projects that would go nowhere.

Across the company, ChatGPT has been used widely for small, repetitive tasks, and to redundate other third-party services like Grammarly. We use it to spell-check our texts, to look for better pieces of code, read long stacktraces and stuff like that.

Later on, we decided to give everyone Raycast Pro because it allowed us to prompt an AI directly from the Mac native app, only to learn that not everyone needed it. I explained this mistake in my last blog post about our SaaS spend. Some of us still use Raycast with AI with custom plugins we've built to do more accurate translations, for instance, or to fetch information from external APIs.

Over the course of the last 12 months, we have also tried AI for coding. Everyone went bonkers for Github's Co-pilot but we weren't too impressed by it. Maybe one or two people in the company use it, if even, but we're now trying Cursor, which definitely feels like a huge step-up in the right direction.

For marketing, I have been using Buzzsprout's own AI tools to generate content related to the podcast episodes I'm uploading, and I use Auphonic to improve the quality and the outcome of the audio files for the benefit of our listeners. For video, Capcut offers a great set of AI-powered tools and I'm generally happy with it.

Nowadays, I'm tinkering with translation tools and lip-syncing libraries to see if it's feasible for me to translate our podcast to other languages with AI, so I don't have to manage two feeds, the Spanish and the English one, but I think this needs a bit more time.

Last, but not least, I switched over from Google to Perplexity as my main search engine. Its main drawback is its slowness, but the results are pretty great.

But, enough of it! Let's speak about our projects!

Show me the projects

Here's a summary of the five projects we've done:

Project #1: Scenario simulator

We have built a scenario simulator app using VR and AI for a big financial corporation from the US.

This application will be used to simulate business scenarios so they can take more informed decisions afterwards. To avoid breaching confidentiality, we can only say that we generate visual avatars (using virtual reality) personifying different experts within the company (using agents) so the user can engage in single and group conversations with them.

Project #2: CV matcher

For a European HR company, we have built a tool to help them increase their conversion percentages.

When users of the platform upload their CVs to build their online resume, our component matches the resume automatically to the most relevant jobs available in the platform, using more advanced & complex algorithms than just searching for their last job title.

This module has increased profile completion and provides a great example that, in a lot of cases, AI is better suited for a small micro-functionality to make processes more efficient, cheaper or less error-prone.

Project #3: Translation for minority languages

Since 2015, we have been working for Naiz, one of the leading media platforms in the Basque Country. Compared to other languages, Basque has got a very small number of speakers (according to the Euskal Kultur Erakundea the number is around 750000 speakers), yet in the last years, its usage is on the rise.

Because of many historical/societal reasons, Basque hasn't been used much in technology nor in business, and therefore most translation tools don't have it or are very inaccurate, as the corpus they've been trained with is more scarce than other languages'.

For years, we have been using AI to translate from Spanish and French to Basque and viceversa because Naiz has content in the three languages, so we are welcoming new tools and progress in this field with open arms, as our options had always been extremely limited - to just one tool!

Lately, we are comparing different LLMs to see which one is more cost-effective and accurate when translating to/from Basque.

Project #4: More attractive job offers

Back to the HR sector, for another project, we developed a module to see if we could improve the generation of job offers for large announcers and headhunting agencies.

Because most job offers' summaries and highlights are auto-generated using rudimentary methods (mostly copy&paste), they all look alike and aren't perceived as original content by web crawlers like Google's. Most go like these "Company X, from Connecticut, is looking for a web developer, with technologies X, Y and Z...", such and such and all that jazz.

By using different LLMs to generate better-looking summaries of the posted jobs, we have increased the conversion rates in this employment platform by 10% of the traffic coming from SERPs.

Project #5: Image recognition for marine biodiversity

We have been developing the MINKA platform for a couple of years now. MINKA is a citizen science platform that plays a crucial role in monitoring and documenting marine biodiversity along the Mediterranean coast. Since its launch in 2021, MINKA has surpassed 200,000 verified marine observations on the Mediterranean coast, marking a significant achievement in citizen science efforts with over 2,500 different species documented, 34 invasive species and 40 protected species identified.

The platform has a functionality for users to upload their findings via a photo uploaded. These photos are, in turn, scanned using AI to help to identify the correct species and tag them accordingly, to speed up the process and reduce human error by over 90%.

Final thoughts

After barely 18 months using AI and working in AI-related projects, I am still forming my opinion and reshaping it on a weekly basis, if not on a daily basis!

However, some thoughts I've had simmering for the last months still stand true today, so I can share some quick parting ideas:

• Too much hype: It's hard to separate the wheat from the chaff. While AI seems to be experiencing a slow decline, now in August 2024, new tools and models still grab all of the attention and we're seeing humongous VC rounds every week, so the hype won't stop anytime soon.
• Tech isn't stress-tested: Most AI tools don't go past the attention-grabbing demo, a few videos made by influencers and some basic usage. Once you start needing them for more advanced needs, they don't produce the desired outputs.
• AI will produce more and better coders: We're very close to having the required toolset to make more and better programmers, yet not everyone will be able to become one. No-code brought new people to the party, and AI is bringing even more people, but there'll still be people who won't join until the next thing comes out and makes it even more accessible.
• Coding co-pilots: Cursor + Claude takes away the fun part but doesn't help with housekeeping. Until it's able to resolve gem conflicts correctly, read stacktraces and error logs correctly and point to the right answer without too many iterations, people will still get stuck fairly easily and give up trying. I'm fairly optimistic about this technology, though.
• Training is a full-time job: The landscape changes rapidly, there's new stuff everyday, the FOMO is too much to handle and as a result of all of that, it's hard to dive deep into anything without being lured and blinded by the next shiny thing.
• AI is making me be a better boss: AIs do one thing better than humans, which is to tell bosses that their instructions are rubbish. Humans tend to avoid this kind of conflict lest they upset their bosses, because their next salary raise, or their job altogether, might be at stake. Give poor instructions to an AI and it either comes back to you saying that it needs more context, examples or sources, or it might output a terrible job. I am learning how to give better instructions thanks to this reinforced training. So, hey! an AI has trained me, in the end!

I know for a fact that these thoughts won't stand the test of time and that in less than a year, I will have to eat up my words, but I am definitely looking forward to working more with AI and testing more exciting stuff.