In no particular order, here are the top trends and product directions I’ll be watching out for in 2025.
Web automation and AI browsers
Is this the year AI agents break out? I won’t go that far, but the piping is underway and the web will be a key part of it.
Using ChatGPT today still feels like talking to someone who wakes up to answer a question, then goes into hibernation until you ask it something else. This stateless flow was a great way to kick off the AI hype: easy to use, scalable to host, and relatively safe insofar as a single text response can only do so much damage. However, it’s left people wanting something far more autonomous, which is why software agents quickly followed ChatGPT, e.g. BabyGPT and AutoGPT. Those were cool in concept, but tools like this need to interact with the internet to do useful things, and that’s where they often came up short due to anti-bot mechanisms. It’s now a couple of years later, and industry leaders are working on ways to grant agents internet access, safely and with consent of both users and service providers.
Web browsers look like they’ll be a big part of this, since they share common patterns across services and are easier to augment than native apps, as browser extensions have long exemplified. There’s already been various efforts to make browsers AI-capable, such as Copilot being bolted onto Bing. Startup Arc has gone further by organising tabs and downloads with AI, and performing searches by going out to the web and building a summary result.
OpenAI has already baked search into ChatGPT and is rumoured to be contemplating a browser. It would make sense in this context. While I’m sure they would love to build out an entire hardware and app ecosystem longer term, a browser doesn’t require experimenting with entirely new form factors or convincing developers to port their apps across to a new OS. Thanks to the open nature of the web and the open-source foundation of Chromium, startups such as Brave and Arc have been able to gain traction in the browser space.
Google is also pushing the frontier in this space. They recently announced they’re testing a browser extension that supports Gemini-powered automation. Watch this space … or ask your AI agent to watch it for you.
Synthetic podcasts
Text-to-speech is not yet perfect, but getting to be good enough for long-form consumption. It’s becoming a powerful enabler for fully automated podcast production when combined with the capacity of (a) LLMs to generate content and (b) AI image generators for artwork.
Soon after ChatGPT launched, there were already fully automated podcasts in production, such as Hacker News Recap. I subscribed for a while and found it quite handy for a catchup; my main issue was the language could get quite repetitive. That’s the “ChatGPT voice” that professors could use to detect plagiarism with certain idiosyncratic expressions. However, this effect will dissipate as LLMs improve and prompts are better engineered to inject some more randomness and distinct personas.
While HN Recap is a single-voice narration, Google’s NotebookLLM generated excitement this year when people discovered its ability to produce podcasts using two synthetic hosts on any topic, covered even by CNBC. There’s not yet flexibility on the voices, but there’s certain to be startups as well as Google working on a whole set of options for anyone to spin up a podcast.
The immediate implication will be a lot of junk filling podcast catalogues, which I fear will lead to a breakdown of the long-lasting open nature of podcasting. Within that bucket of AI content, I believe some will turn out to be interesting enough if they can just inject a little bit of human guidance and editing in the mix. The software can still handle much of the grunt work of gathering content, setting up a plan, recording, and editing, but humans can still touch up each of these. In a few years, the current production model, with humans micromanaging everything, will feel positively quaint. There will certainly be demand for human talent doing “real recording”, but even this audio or video will be editable with the touch of a keyboard or the utterance of a command.
Video podcasts
Video is becoming a key part of the mix for modern podcasts. While video has always been possible in the 20-odd years since podcasts were invented, it was far from practical or desirable for publishers to bother with. Today video is easier to monetize, particularly because of the ubiquity of YouTube on smart TVs and mobile devices, and the bandwidth available for billions of people to access video at low cost. Where hosting costs used to be a major burden for publishers, YouTube makes it free and then adds a revenue share on top. Spotify is also working aggressively to handle video podcasts and with hosting costs now much cheaper than the 2000s, community platforms like Substack are able to host video on behalf of premium subscribers.
While video production has also become easier, it’s still a big investment and not something every creator can do well. Part of the appeal of podcasts is how open they are to anyone who can speak passionately about a topic. However, audiences only have 24 hours a day and their attention is divided between all players. Within any niche, there’s a winner-takes-all dynamic taking place as podcast budgets continue to grow. As technology makes video as readily available as audio, we can expect to see video becoming almost a requirement for aspiring podcasters. When you consider how easy it’s becoming to AI-generate a half-decent audio podcast (see previous section), there’s even more reason to differentiate with a video format.
Many users will no doubt continue to listen to podcasts rather than actively watch them. However, there’s still going to be increasing pressure to have the video available to those users as part of the complete package.
Neatfakes: Authorised replica media
Deepfakes have gotten some bad press because — quite rightly — no-one wants to be impersonated without their consent and there are serious security concerns. However, the underlying technology of digital twin replication has genuine applications when it’s used in an authorised context.
One use of authorised digital media is to have your “digital twin” act as yourself on phone calls. To be clear, I’m not sold on that idea and not referring to it here. It’s deceptive to use such a voice if the other side is not clear they’re talking to a machine. And if they are clear they’re talking to a digital version of someone, what are we achieving here exactly? Why couldn’t it be a different voice?
Where this tech is more practical is in media production. Dubbing movies and TV shows can now be done via AI. AI speech preserves the voice of the original actor, while AI video smooths over their face movements as this speech takes place. This will dramatically lower the cost of translation while improving quality. In the future, we can expect any given production to be available in dozens of languages.
The production ease and speed means we can expect to see translations in contexts we have not yet seen them before. Lex Fridman’s recent interview of President Milei showcased how translation software can be used for an interview as he was able to release it in both full-English and full-Spanish, while preserving either speaker’s voice and expressions. Expect to see more use of this technology in podcasts and mainstream news media.
Another more controversial use is to “revive” the deceased in video form. People can’t give consent after they die (although their digital twin can). They can, however, establish rights agreements and wills with that technology in mind, and I would imagine this will start to become standard practice in future even for non-celebrities. Netflix recently released a Churchill documentary along these lines, with the following preamble making it clear the text has been authorised to the extent that’s possible:
“With permission from the Churchill Estate, this series utilizes voice enhancement to present many of those written and reported words aloud. For the first time.”
Prediction markets
I’m make a prediction that prediction markets will rise in consciousness among the pundit and journalist class. This includes dedicated prediction markets such as PolyMarket, Kalshi, and Manifold; as well as conventional betting sites that are expanding their scope to cover broader topics than sports such as politics and current affairs.
When it comes to politics, there’s a rich history of betting on US presidential elections dating back to the 1800s, with trading volumes at times exceeding stock and bond volumes. Before the polling industry became more sophisticated, price action on candidates was seen as a useful gauge during election campaigns. However, the activity was eventually swept up in anti-gambling legislation and a century later, activity is still very restricted in the US and many other jurisdictions.
However, the notion persists that such markets can be a valuable indicator on politics, technology, or just about any other topic. The economy already benefits from cues provided by prices of stocks, bonds, currencies, and other instruments, allowing businesses and governments to tap into the wisdom of the crowds and plan accordingly. Thus, the thinking goes, why not also use the greed motive to inform businesses and governments about the odds of elections, product launches, or Nobel Prize winners. Compared to polls, prediction markets are a more direct way to estimate probabilities at any given time; and they are not susceptible to selection biases or respondents who are too shy to disclose their thoughts truthfully.
This has been a start-and-stop industry for decades. There’s not just concerns about gambling (which a cynic might associate with heavy funding from gambling lobbyists), but also worries about insider information and manipulation of results. In the extreme case of manipulation, detractors have pointed out the risk of “assassination markets” where tragic events are betted on and then brought into reality by an actor with a stake in the outcome.
Before 2024, most conversations about prediction markets were limited to very niche communities, e.g. political data wonks like Nate Silver, “rationalist” writers, some economists. The recent US election saw conversations break out into broader media. As one example, the founder of PolyMarket appeared on All-In’s election night livestream, where he pointed out in real-time how much traditional media was lagging the market’s forecasts state by state. On social media, there was widespread speculation around one “Théo”, a French trader who bet over $70 million on Trump. By way of contrast, it was frequently mentioned after the Brexit result that hedge fund traders had made fortunes shorting the pound, but no-one ever mentioned prediction markets in all this. The volumes would have been too minuscule for anyone to care.
Why now? I believe the timing was due to the growth of those prediction market platforms in the past few years. That begs the question, why did those platforms grow all of a sudden when anyone could have made a smartphone app to do this in 2010? I think the primary answer is crypto. Decentralised currency made it easy for some platforms to skirt the regulations in allowing trading with real money.
While there will be ever-growing opposition to this trend, there’s a powerful viral element in speculation like this that should continue to see interest rise. Through social media and podcasts, the public is becoming aware that these markets can forecast events in ways that polls cannot. And it doesn’t hurt that they’re hearing about big winners like Théo.
Defence-Tech
Defence-tech is becoming popular with investors and entrepreneurs from within mainstream tech industry. Workers and investors have often avoided the industry due to moral objections as well as a general perception that the industry is too dull, too slow-moving compared to the familiar Just Ship It mentality.
The objections reached a peak a few years ago, when (a) the threat of superintelligent AI seemed at least decades away (b) enough tech employment for sufficient portions of the tech industry to feel empowered enough to lobby their bosses against that line of work (c) there were less hot wars happening to bring about a sense of urgency.
Two companies in particular have been working to combat the notion that defence-tech work has to be slow and boring: Palantir and Anduril. Both take a product-oriented approach to their work, going against the tradition of years-long waterfall projects built to spec and billed hourly, thus creating an incentive for contractors to drag their heels and creep the scope. Both have also benefitted from the recent AI boom as they were already anchored in that world before it was cool.
Palantir was founded in 2003 and has built a suite of data analysis products for customers spanning from secretive intelligence and defence organisations to high-profile enterprises and government departments. By sharing common components between all of these customers, Palantir is able to provide highly competitive products and timelines to defence that would otherwise have been custom-built, even if there’s a degree of customisation required.
Founded in 2017, Anduril represents an even newer and more radical approach than Palantir (where one of its co-founders had previously worked). The goal is to build vertically-integrated defence hardware in the style of a tech startup, taking an approach that is agile and embraces a software-first model that would not be out of place in makers of physical consumer products such as Apple, Tesla, or Peloton. Going a step further, Anduril has also developed an operating system, Lattice, with the intention of leading an entire ecosystem of software-focused companies that can plug in and communicate with each other. Remarkably for a project of this nature, the company recently made public the SDK and docs just like any startup with an API would do.
Severance
To round out the list, here’s a fun prediction and frankly a recommendation too. Severance season 2 is due in January. The first season rates alongside Black Mirror in the pantheon of sci-fi productions over the past decade. I recently re-watched the season again in anticipation of S2 and found the second time round just as engaging. Seriously, go binge if you haven’t already. I’m expecting this will be a series that accumulates a fan base as time goes on, in the mould of Breaking Bad or White Lotus. The unique premise and the cast of characters in this universe lend themselves to an intriguing set of story arcs ahead.