Ten predictions for 2025
What will the software engineering industry look like at the end of 2025?
Note: I will be taking a break on December 26th for Christmas, and the next newsletter will be December 30th.
This is the first of two final posts for 2024: predictions for 2025, and the year month in review for the Client/Server newsletter.
These ten predictions are split into two groups: “AI and politics”, and “everything else.” I could have easily made 10 predictions just about AI, but I wanted to force myself to assess several parts of the industry.
This is a “just for fun” post. Imagine that we’re having drinks at a holiday party and you ask me “So what do you think will happen in 2025?” and I ramble for 5 minutes. I didn’t research it that heavily, I’m not bothering to cite sources, some things may be factually incorrect, etc. To encourage more aggressive predictions, I am challenging myself to get exactly 5 out of 10 right.
Predictions for AI and politics
1. The gap between AI software engineering benchmark performance and real-world impact will widen.
LLM systems will become cheaper, more accessible, and more powerful. But their stated performance will sound superhuman and we may even see the first AGI claim.
First, there are the “good faith” reasons. The industry is still pouring resources and effort into researching how to produce larger and more effective LLM architectures. We started to see hybrid architectures like OpenAI’s reinforcement-learning-powered O1 model. They may find more effective ways to run larger architectures, efficient ways to run ensemble models, or even release viable alternatives to transformers with fewer parameters or cheaper training.
Then there are the “bad faith” reasons. There is just too much to gain from stating that you’ve topped some latest benchmark. You did well on some coding benchmark? We trained on the entirety of GitHub, it’d be a shame if someone had published solutions to the benchmark questions there.
If this doesn’t happen, it’ll be because the 2024 hype translated into real 2025 results.
2. Some company will go viral for discriminating against non-AI developers
Somebody will become the Main Character on their social network of choice for stating that they explicitly test developer skills using Claude/ChatGPT/Copilot/Cursor/etc during an interview, and fail anyone that can’t code using an AI assistant.
I think some company, in an effort to get “all press is good press”-style coverage, will publicly announce that developers who don’t code with LLMs will be rejected from their hiring funnel. This will generate every thinkpiece from “This is discrimination” to “This is good, actually.”
3. AI agents will increase in popularity, and might even start to not suck
AI agents (long-lived programs that can both gather data and act on it) are in a relatively sad state right now. For example, despite all the hype, Devin is not the first AI software engineer. Instead, it is fairly limited and works best when it’s doing a menial task that you would have done yourself if you had the time.
But I predict that AI agents will become better and better as they are hooked up to stronger models. Just like how AI agents conquered a limited version of DoTA 2 and Starcraft II, they will be able to be fine-tuned and live inside their own OODA loop, performing useful work on a task for long periods of time with limited intervention.
I will consider this useful if we see “more Devins.” I.E. more companies release their own AI agents and the technique starts to find niches.
4. Justice Department actions against most big companies will go away
In 2025, I predict (based on past performance) that cozying up to the President will provide benefits for companies that are under investigation. Control of government institutions seems to be a focus of the incoming administration. So the administration will have an outsized influence over the Justice Department that is not typical for presidential administrations.
This means that I expect that companies will wave their checkbooks and make the Justice Department go away. Google won’t need to divest Chrome anymore, app stores will remain locked down, RealPage will keep raising rents across the US, etc.
So I think that the TikTok ban/divestment will still take effect, but any ongoing efforts — like the Google trial — will end with a slap on the wrist if anything.
5. GitHub Copilot will introduce a Cursor-style product
GitHub Copilot will get an incremental release — that costs more money — that acts more like Cursor.
If you’re not deep in the weeds on LLM-assisted coding, the difference between Copilot and Cursor is pretty simple. Copilot exclusively acts through autocompletion suggestions. Cursor acts through a series of suggested edits. Cursor’s model is clearly the correct one, since it excels at autocompleting code blocks but struggles at other common tasks like refactorings.
Since GitHub and Visual Studio Code are owned by Microsoft, they have everything necessary to make a new Copilot product that outperforms Cursor. If they make a new tiered subscription model and provide Cursor-style prompting and suggestions on the more expensive tier, I’m sure a lot of existing Copilot users would upgrade. I sure would.
This will be a huge miss for GitHub if they don’t address it. I use Copilot extensively in my day-to-day work, and it’s painfully clear that software is done through edits and not exclusively additions.
Predictions for everything else
6. Tech hiring will increase at least 10%
For a while, tech hiring has been slowing down. I predict that we are at the bottom of that, and that hiring will pick up in 2025.
During the zero interest rate years during the pandemic, companies were incentivized to hire relentlessly. There was less focus on profitability because money was so cheap to borrow. However, once everyone acknowledged that inflation was not transitory and interest rates spiked, reality settled in. There were huge rounds of layoffs. Since then, the tech hiring market has been tight.
I predict that we have hit the bottom of this slowing market, and hiring demand will increase in 2025. Tech layoffs have been declining for 2 years, so I believe that companies have finished compensating for the overhiring of the zero-interest-rate years. That means that any company that can make a growth case will be looking to bolster their ranks.
7. React will remain supreme in the frontend web world
React may decline in marketshare, but there will not be a single obvious alternative.
React has many detractors for several reasons. It has a nontrivial download size. It encourages developers to change simple sites into single-page apps, which is potentially overkill. The virtual DOM feels like an obvious smell that shouldn’t have to exist (and doesn’t in basically every other frontend language and library). React itself is not opinionated on your underlying state management, so developers need to find their own way to ensure that server and client state remain synchronized for long-lived sessions.
React has an incredible number of alternatives. There are traditional framework alternatives like Vue and Angular. There are also newer Web Component libraries like Lit and Svelte. Additionally, there is also the HTMX approach of revitalizing the server-side rendering approach.
I predict that none of the alternatives will gain significant marketshare over React. At the moment, React is the unified frontend library for the industry. Development is faster because of JSX. It’s the thing that everyone is learning. Despite a lot of unrest towards React and a lot of sentiment that many use cases should be trivially replaceable with things like HTMX, its fall will be much like JQuery’s: gradually, then suddenly.
8. The percentage of Rust code in the Linux kernel will climb
The percentage of code (by LoC) in the Linux kernel will be greater at the end of 2025 than at the end of 2024.
My case for this is simple:
The US Government is strongly urging large companies to adopt memory-safe languages.
This is clearly targeting languages like C and C++.
So big companies with systems experience will develop in-house
And since they have the Rust system developers anyways, it’s the most obvious language for their kernel contributions.
9. An indie game will be nominated for Game of the Year
This one is simple. I think that in 2025, another indie game or new unaffiliated studio will be nominated for the Game Awards’ Game of the Year category.
To me, the awards seemed overindexed on AA and AAA games. In the past 10 years, there were plenty of indie games that arguably could have “provided the best experience across all creative and technical fields.” But there are some notable indie nominations, like Balatro in 2024 and Celeste in 2018. But where are the Hollow Knights? Is every Resident Evil really hitting that sweet spot of “always nominated and never winning?”
I don’t think that an indie game has a chance of winning. How is a roguelike going to compete with a game like Elden Ring? But indie games keep getting stronger every year and I think that the gaming industry will continue to see “small game with fun mechanic” outperform “10 minutes of credits” by virtue of being passion projects.
10. The market for vector databases will be consumed by the incumbents
Vector databases have made a great niche for themselves. They allow rapid searches of embeddings using techniques like HNSW, and even allow the vectors to be associated with other data that can be used for filters.
But do you know what else has solid vector support? Postgres and Elasticsearch. And these database are infinitely more flexible.
My theory is that if you have a simple site that can tolerate a limited implementation, pgvector has everything that you need. If you have a complicated use case (aggressive prefiltering, bidirectional preferences, geographic filtering, complicated conditionals, scripted behavior) but you don’t need ACID compliance, then you would simply use Elasticsearch. But the Elasticsearch point isn’t important, it’s the Postgres one. If you’re already using a mature well-understood database like Postgres, with a mature, well-understood plugin like pgvector, why would you also include a separate vector database outside of your primary stack?
Predictions I’d like to make, but they’re either boring or I don’t have a good way to validate them:
Offshoring and nearshoring will go down. Don’t have a good way to measure this and I’m not comfortable vibes-ing it.
Python will be the top language in GitHub’s 2025 survey by a wider margin than 2024. Too boring.
VPS will see an increase of popularity. I don’t think I have a good way to measure this and I don’t trust my vibes check.
I will regret making a list of predictions. Too obviously true.
Someone will tell me one of these predictions is already true. Too obviously true.