2024 marks a pivotal moment in the evolution of engineering organizations, driven by the accelerating influence of Artificial Intelligence. This technological leap is not merely an advancement— it's a transformation that redefines the very essence of engineering workflows, decision-making, and innovation. 

‍
As AI becomes more integrated into the fabric of engineering it will reshape the tools and processes, and perhaps more importantly, the strategic outlook of organizations. The AI landscape in 2024 will be a complex mix of accelerated development capabilities, cost considerations that defy traditional scaling laws, and ethical implications that necessitate careful navigation. 

‍
With that in mind, here are AI developments that tech leaders should prepare for as we head into the new year. 

‍
AI code assist picks up momentum

‍ 
The first time most developers tried tools like Github CoPilot, they expected them to hilariously fail. In reality, most were shocked at how good they actually were. Similar quality results came out of ChatGPT and Bard. None of these tools are as good as humans yet, but they are making development faster: Github reports developers can be 55% faster with CoPilot. And while Github has a vested interest in these numbers being good, it’s worth noting that I’ve seen the gains with my own eyes and heard similar reports from other CTOs. 
‍

In 2024, developers will start to find the right balance in when and how to use code assist. For some developers, the tools will be best deployed in areas where they’re rusty, like when they have to use a language they haven’t touched in a year or two. For others, it will be in moving quickly through rote work so that they can focus on unique challenges where training models don’t have the best solutions. Expect junior developers to reap the greatest gains, if they know enough to properly vet the generated code. For developers in specialized industries or in cutting edge work, where there’s less training data available, expect less significant gains. 

‍
While there will still be concerns around copyrighted code and IP scraping, for better or worse, most companies will remove obstacles to enable employees to use code assist (or deploy enterprise solutions that claim to protect IP). 
‍

Also worth noting that while developers may get 55% faster, that won’t necessarily mean they’ll get 55% more work done, it might just mean they do the same amount of work in a 5 hour day… and at the moment, short of some very Big Brother-esque monitoring, it’ll be hard for managers to know. As always, happy, well-incentivized developers are likely the ones you’ll see hitting those 55% higher outputs. 

‍

Startups punch above their weight 
‍

CTOs at large companies are slowly getting AI tooling into their teams hands, but given copyright concerns around AI output, and that AI services may be vacuuming up proprietary code for training data, many big companies have understandably been moving slow. Startups have none of these qualms, and are embracing generative AI with open arms. They are finding the limits and potential of AI faster than big companies, and are adapting AI to their own needs with no red tape. 

‍

Less talked about than code assist is the ability for generative AI to radically strengthen a developer’s weaknesses. If a developer is highly proficient in front end development, but only so-so at back end, with AI support they could potentially raise their back end skills up to a level that’s passable in a startup environment. Does that mean that their company won’t ever need a backender? No, but it might mean that they can get by with ‘good enough’ while they keep their company small, and only hire full time when they really need ‘perfect’. This logic can be extended out beyond development to product, marketing, management etc. All of which means individuals can cover a wider range of functions at critical stages in startups. 
‍

While these startup advantages are industry agnostic, expect startups to excel in any areas where incumbents are saddled with bureaucratic technology management. That said, startups need to be wise and take advantage of this moment as incumbents will eventually figure out how to leverage AI. 

‍

AI costs reach new highs before becoming manageable
‍

Lost in the wonder of what generative AI can do is the fact that all of its output is so wonderful because it’s 1:1, and not 1:many. The assumption is that AI will get faster/cheaper before there is a scaling problem. And I expect this will be true – eventually. But supply chain constraints + epic demand + customer demand to use the latest/best models will equal continued high compute costs. All of which will make AI monetization tricky. Most content on the web is write once read many, enabling highly effective and cheap caching. Generative AI output has relatively high compute costs for every user, and the longer the conversations are the higher the costs. 
‍

It's arguable that with the value of programmatic ads being as low as they are these days, coupled with high compute costs, most AI businesses will lose money if their only monetization is advertising. Keeping in mind that it’s ads that have traditionally, if controversially, sustained most new technologies (check out Tim Hwang’s Subprime Attention Crisis for an excellent overview on this). That’s why you’re seeing so many AI companies push subscriptions. But subscription fatigue is real, and only so many companies will be able to get users to add another recurring charge to their credit card. 
‍

Given the insatiable demand for all things AI, compute cost could be a significant and unexpected speed bump to AI adoption in 2024, and businesses should be thinking about how to approach these costs now, not waiting for scale and users first. (And I’d be remiss if I didn’t call out some solutions I’ve been working on in the micropayments space that could help better match consumption with compute costs 😊). 

‍

Blurred lines around when to attribute AI output lead to trouble 
‍

As people, specifically managers, get comfortable using tools like ChatGPT, two concerning issues will become more prevalent: 
‍

Managers are dramatically adding more red herrings into problem solving. Case in point, a few weeks ago I was asking my dev ops team to provision an EC2 instance with some very specific services, and they were having some issues. While I waited for them to debug, I copy pasted the problem into ChatGPT to see if it had any solutions. And it did! Seven of them… I was about to send off the most plausible answer to dev ops, when I realized that while I had very little information about what the problem was, my dev ops team would have to consider the suggestion I’d sent over, not only because I was the CTO sending them a suggestion, but because ChatGPT would make the solution sound intelligible, even if it was garbage. This is especially troubling because ChatGPT can generate such convincing sounding solutions. I guarantee a lot of execs, while trying to be helpful, will in fact delay solutions. And the kicker is that in a lot of organizations, telling your manager to stop sending over unhelpful suggestions is a non-starter. 
‍

Which brings me to a more serious issue: Smart answers from ChatGPT are going to enable poor managers to look competent, especially in big organizations where they can more easily fake it. The danger here is that when managers are giving opinions on things they don’t understand, bad decisions are going to be made. And don’t think that they’ll get caught easily; anyone can copy/paste a Slack exchange into ChatGPT and ask for a reply that leverages training data that approximates their way of writing. The responses won’t be perfect, but in many companies B+ will get you by. This may take several years to really identify as a real problem, but I guarantee it’s already started. 

‍

Excitedly and exhaustedly looking forward to 2024… 
‍

As we end 2023, the integration of AI in engineering organizations signifies a pivotal shift, offering remarkable opportunities for growth and innovation while also presenting unique challenges. We’re going to see enhanced efficiency in engineering processes, a competitive edge for agile startups, and a need for new monetization models in AI services. The success of engineering organizations in this new era will hinge on their ability to adapt to these changes with strategic foresight and tactical experimentation.

‍

If your looking for more reading on AI history and predictions check out these articles:

GitHub Copilot gets a new ChatGPT-like assistant to help developers write and fix code

AI is the great equalizer

Here’s how AI startups can beat the big guys

Subscription fatigue: More companies are charging monthly fees. How much can consumers take?