AI

AI models keep getting better

by

This September was a particularly busy month for AI. Fable 5.1 already feels like a distant memory, with GPT-6 Astra, Opus 5.5, Sonnet 5.5, and then GPT-6.1 Sol released shortly after. And to cap it off, Google announced Gemini 4 Argon, which at least by its shared benchmarks, suggest a return to the frontier.

To me as a consumer of AI models, this has been a (mostly) joyous and productive period. AI has been genuinely useful for me at work, especially in the realm of data analysis, as I could self-service a lot of analytics. Questions that would have been shelved a couple of years ago – due to bandwidth prioritization of scarce data analysts – now have a directional answer in 30 minutes. Having different models / harnesses independently tackle the question, and/or do adversarial reviews of others’ work, provide some reassurance that I’m not being blindly led on.

I’ve also been amused (and somewhat excited) by the discovered new capabilities of these new models. First it was GPT-6 Astra and Blender. And then Opus 5.5 showed its knack at pixel art and programmatic video generation. And of course, there’s been a flood of one-shot “make a video game” examples, which looks better and better with each new model.

For video games, there is a big difference between “looks good” and “plays good”. To quote Simon Willison (whose writings on AI I’ve greatly enjoyed):

Something I’ve realized about game development is that you can vibe-code something that looks like a computer game, and that’s easy.
Building a game that’s fun, has a good gameplay loop, and is challenging and interesting and keeps people coming back for more… that’s still beyond me, and beyond any of the agents I’ve tried.
This ties into the Deep Blue thing. Just because we can make something that looks like a game does not mean that we are game developers.

(If you are curious what Willison means by “Deep Blue”, you should read the rest of his post.)

I’d like to push this thread one step further though – what is stopping AI models from being great game developers right now? Let’s think through this.

Is it a lack of domain knowledge and theory? Likely not. The frontier models are a compression of the internet (not to mention being literal devourers of large volumes of books). On any game dev discipline, they are probably already better equipped in knowledge than the average person (or even professional game developers not of that discipline – i.e. AI is a better game artist than a typical human game programmer).

Is it due to AI outputs being too derivative? I’d give this a qualified yes. Without proper guidance, AI outputs do tend to look too familiar or over-index on established tropes (based on its training data), but even this longstanding point of criticism is weakening when the base quality of the output keeps getting better.

How about limitations in tool-use? This is where I’d argue that we’ve seen the biggest leaps in model capabilities recently, but also where it’s still lacking the most. AI is already better at web search than humans, and they can decently use browsers for various automation. More recently, we’ve been wowed by their abilities to utilize Blender, three.js, or vanilla JS (for pixel art), as well as python + ffmpeg for video. But if you’ve observed their workflow in these tasks, you’ll notice still that we’ve barely made any improvement on how AI review their work – by and large, they are still limited to taking screenshots, and then inspecting these screenshots.

Processing visual info frame-by-frame is computationally expensive, and this is easily one of the biggest areas where your token budget on frontier models get devoured. It also means that models are currently woefully ill-equipped at iterating on moment-by-moment game feel. Remember this is not just a visual task – a combat animation can look right but plays wrong, so it is the orchestration of the physical button presses and the responding audio-visual frames.

Some of this can be mitigated by making the task easier to tackle for AI. One specific path is being data-driven. If we could structure the data, such that models can see perfectly through data a comprehensive representation of the game state, then it would be akin to models having access to the firehose in the Matrix. For some tasks, this is already the established best practice – for example, having AI run lots of headless simulations to rapidly and exhaustively verify the effect of game balance changes is a powerful tool that many teams already employ.

But even if we had data abstraction such that models obtain “realtime vision” (however that’s achieved), models would still need to internalize taste to be able to evaluate and iterate towards the desired game feel. And thus we finally mention taste – perhaps the most cliched word for me in AI discussion this year. I now loathe to say it, but at least for now I can’t work around it. As long as humans are the consumers (scary thought if we aren’t), human taste is part of the equation to solve for commercial value. Any commercially successful game by definition has a commercially validated game vision. Such game visions don’t come from thin air; they are the culmination of the developers’ past experiences, judgment and taste. And materializing a game vision from a design doc requires lots of further judgment and applications of taste. In our AI-native world, we can feed AI models our design docs, and (through lots of iteration) help it refine its own heuristics to know “what good looks like for this specific team and game vision”. But it’s hard to imagine a workflow where AI models independently arrive at that end point without lots of human intervention.

So, I guess where I stand is this: can AI models be great game developers independently? No. Can they be great game developers working alongside human developers? Absolutely, and they already are.

I should stop here – this post has gone down a branch longer than I anticipated. But back to the main trunk of observing AI models getting better rapidly – there is one more thing to mention. I believe we are seeing real world examples of the “lightspeed leapfrog” trope from sci-fi. AI models are gaining specific capabilities faster than people building harnesses for specific use cases. The best-in-class pixel art workflow 6 months ago might involve a mix of 2d image generation and feeding images into video generators (seedance etc.); with Opus 5.5 and beyond, maybe the workflow has migrated to programmatic generation. So it’s quite possible that a team that started 6 months ago is facing the prospect of being leapfrogged by teams starting last week. This does not mean you can afford to wait – if you want to ship a game, you work with what you have. But it is nonetheless a fascinating phenomenon to live through.

AI, taste and judgment

by

Time for another AI post. For brevity, let’s agree that the following discussion is centered on the games industry (though a lot of the points are applicable to broader domains). This is a bit of a free-flowing post, but the central theme is be about the work relationship between human talent and AI.

These days we talk a lot about how AI will or will not replace jobs (and what kinds of jobs). One of the common optimistic opinions goes something like this:

  • Games are entertainment products serving human consumers.
  • The commercial viability of a game – whether human consumers prefer it (over alternatives) – depends on the game’s ability to identify and satisfy complex human motivations and needs. Put more simply, human taste is key to the creation of fun.
  • Taste itself cannot easily be replaced by AI; thus, the future of game dev is not humans replaced by AI, but rather, humans directing AI and humans exercising taste.

This is a train of thought that I largely agree with. But it is possible to poke some logical holes. For one, from a “jobs to be done” theory perspective, there are lots of occasions where the job a game is “hired” to do does not require sophisticated taste in execution – for example, if the primary need is to “kill time”, then the consumer is often not seeking an intricately crafted experience. But, we can also argue that the key insight here – applying the “jobs to be done” framework to discern consumer needs – is itself an exercise of taste (or more precisely, judgment).

Before going further, let’s distinguish taste from judgment (and briefly discuss how they are formed). For me, taste is subjective. It is “beauty in the eye of the beholder.” I associate it with creative expression. There is no objectively good or bad (right or wrong) taste, but there are certainly societal opinions around “good” or “bad” taste. For commercial concerns, it is critical to understand how popular / niche a certain “taste” is (and this is often where we are surprised).

Regarding how taste is formed, I strongly believe that you can’t skip the experiential part. It’s not just about media literacy (playing lots of games if you want to make good games). You have to feel it for yourself, and your emotional journey is part of the taste alchemy in your head. Watching a gameplay video instead of doing a full play-through yourself (or say, watching a video essay about a film instead of the film) are not the same. Your emerging taste, formed from passive consumption, is then refined when you exercise creative expression; the community reaction to your work then completes the feedback and learning loop for your taste.

Judgment is more objective. It is exercised in decision-making and STEM-related problem-solving. We can often verify with measurable data whether the judgment was sound or not (however, beware of pitfalls – a reckless decision can get bailed out by random luck).

Since I’ve defined problem-solving as part of judgment, the typical learning model shows one path to forming good judgment in a specific domain – understand the fundamental concepts, study previous examples to see the concepts in practice (and learn derived principles / mental shorthands), and then work through new problems to reinforce the lesson.

Going back to the train of thought earlier: how defensible is “taste / judgment cannot easily be replaced by AI”?

Here, judgment seems like the weaker plank to discuss first. Chess is an example of a “solved” problem where humans are no match to machines, even prior to modern day AI. (For those of you out of the loop on chess these days – 30 years after Kasparov vs Deep Blue, the world’s best players simply have no chance against your average laptop running the open-source Stockfish chess engine.) And for domains that involve hidden information and randomness, as long as we can articulate a rational decision-making framework, it seems plausible that we can hand it off to software (AI or not) to ruthlessly exercise rationality.

But – this sounds like one of those economics theories (people are rational actors / efficient markets etc.) that fail in practice. As long as the domain concerns humans as stakeholders, and humans act with varying irrationality, good judgment without human intervention remains out of reach.

To give a trivial example: I recently tried to automate some meeting scheduling with an AI agent. After jumping through some technical hurdles, I ended up in a place where I realized I was a frivolous stakeholder – I was constantly nitpicking the agent’s proposed scheduling. Don’t schedule these two meetings back-to-back (but those other 2, that’s ok). Don’t book me 8 meetings a day. Mondays and Fridays are off-limits… Some of these criteria can be turned into explicit rules, but getting from 90% to 100% seems impossible (because it’s up to my whims).

How about taste then? Shouldn’t this be inherently harder for AI to replace since it’s subjective and personal? Maybe… But as a counter-argument: doesn’t the fact that so many people are forming digital relationships with AI show that AI, as-is, is already quite adept at emotional engagement? And so, while they may not be proven capable of predicting or defining a new popular taste, they seem to have some innate ability to understand existing taste.

Where am I going with this? I don’t know, exactly. I started out in one direction, but I seem to have argued myself into a slightly different space. But zooming out to the macro level, I think “humans directing AI and exercising taste & judgment” is still a solid-enough paradigm to work towards (for now).

There is one last bit to talk about regarding human talent and AI. Many have called out the risk of AI to the development of junior talent. The old ways of on-the-job learning is obsolete; and even worse, in the eyes of an experienced dev who has successfully navigated the transition to AI, adding a junior talent may be strictly worse than adding a few agents to the fleet. So there is a question of whether our human capacity in a domain will wither as new talent are not grown.

I think this risk is real, but it is overblown. The real impact is it skews the economic opportunity towards a more specific type of talent. The type of person who has deep natural curiosity and the drive to do stuff, and whose experience is self-taught. These kinds of people have always been in high demand, but in the AI age they are even more valuable. Yes – organizations still need to figure out how to turn fresh college grads into veterans, but these unicorns who’ve been tinkering with AI since their teens will also have a decade of experience by their early twenties. Curiosity, innate drive and self-learning mindset – are you born with these attributes, or can we nurture them?

2025 year in review

by

Previous yearly reviews: 2024, 2023, 2022, 2021, 2020, 2019, 2018, 2017, 2015.

This is my 10th year-in-review post – I just feel old writing that. Anyway, let’s talk about some themes of the year.

AI accelerates

Skip to the next section if you are tired of talk of AI; but for me, AI’s continued advancements remained the biggest topic. This was something that I experienced and felt first-hand. Early in the year, I dabbled in vibe-coding and did some small demos in office to spread awareness; at the time, professional programmers’ response was generally “this is cute, but not ready for real game-dev work.” (I was happy that they did start tinkering.) Fast forward to December, the most aggressive adopters’ mindset has already shifted to “I’m not writing any code anymore; how do I effectively manage many AI agents?”

It’s clearer than ever that we are in the beginning phases of an asymmetrical revolution to how all software (including games) is developed. Despite the current lack of mature tooling (and best practices / workflow know-how), we can already confidently predict that it will be far easier to scale AI agents than to hire and manage a large corp of developers. Each human programmer is now theoretically capable of performing as an engineering manager, managing a team of AI devs. And these AI devs will be working 24/7.

This revolution is and will continue to be controversial. Not all programmers want to be managers; many will continue to cherish handwriting code. Soft-skill attributes such as curiosity and bias for change will determine who survives (and thrives in) this turmoil. Quite possibly a lot of jobs will be gone. I’m not taking a moral stance here, but rather making a simple economics-based prediction: we won’t (and can’t, really) put the AI genie back into the bottle.1

Back to the practical implications. When I wrote “asymmetrical” earlier, what I meant is that many things that we previously considered as competitive advantages are going to be flipped upside down:

  • The organizational capability to hire and manage large dev teams is a questionable asset now. Smaller teams have far less human communications (and alignment) overhead, and can drastically scale up production capacity with AI.
  • Similarly, prior investments in optimizing human efficiency (in the old dev workflow) are also possibly rapidly depreciating in value. For example, game engine editor features / tooling that made it easier for manual editing (and enhanced productivity previously) could now be impediments to a full AI workflow (since they may be less AI-friendly to direct data manipulation).

It will still be years before the dust settles on the new software development paradigm; 2026 will be a busy year. At the same time, we are also seeing a brewing consumer backlash. I feel it’s partly a continuation of existing cultural wars (just a new battlefront), though this is clearly a tricky issue for developers to navigate. It’s also interesting how sentiment on this differs by country. There will likely be further conflict/drama before a realignment.

The market stagnates

Last year I spent some time sharing an analysis on the mobile games market. I was going to post updated macro charts (I already populated the data from Sensor Tower), but really the lines for market consolidation barely moved. Meanwhile, we have to put an even bigger caveat on not over-interpreting this data due to the growing fog of war: Apple and Google’s share of spending is rapidly decreasing thanks to factors like web-stores (DTC) and the Chinese mini-app platform.

With that said, let’s take a look at the top of the charts:

Rank2024Revenue Growth %2025Revenue Growth %
1MONOPOLY GO!93%Last War: Survival36%
2Honor of Kings-9%Whiteout Survival44%
3Royal Match48%Royal Match-6%
4Roblox20%MONOPOLY GO!-21%
5Last War: Survival7535%Honor of Kings5%
6Candy Crush Saga-7%Candy Crush Saga0%
7Whiteout Survival249%Coin Master2%
8Coin Master-4%Roblox-42%
9Dungeon Fighter Mobile3191%Peacekeeper Elite9%
10Peacekeeper Elite-6%Pokémon TCG Pocket137%
  • Illustrating the fog-of-war issue I just mentioned, Roblox mobile revenue did not decrease by 42%, but rather, they likely executed a successful shift to their own payment channels
  • The Chinese 4x (or SLG, as Chinese devs would call them) games continued to scale up, all the way to the top of the charts
  • 9 out of the 10 titles are the same, with the only change being Dungeon Fighter Mobile (its decline unsurprising, in hindsight) replaced by Pokemon TCG Pocket

The market continues to be difficult for games to break out. Scanning the top 100, we have the following 11 new entrants (as in, appearing in the top 100 for the first time):

  • Gossip Harbor (no. 11) – an “old” game released mid-2022, but having a breakout year scaling up with 239% revenue growth, 137% downloads growth, and 198% DAU growth
  • Kingshot (no. 15) – another Chinese 4x, globally released in 2025
  • Delta Force (no. 28), globally released in 2025
  • Last Z: Survival Shooter (no. 29), globally released mid-2024, scaling up in 2025
  • Royal Kingdom (no. 31) – globally released late 2024, so 2025 is its first full calendar year
  • Dark War: Survival (no. 39) – guess the genre and which country this is from? Globally released mid 2024, revenue grew 840% in 2025
  • Archero 2 (no. 67). Globally released Jan 2025
  • Color Block Jam (no. 73). Globally released mid 2024, scaling up in 2025
  • Disney Solitaire (no. 84). Globally released Apr 2025
  • Seven Knights: Re:BIRTH (no. 87). Squad RPG, globally released Sep 2025
  • Where Winds Meet (no. 94). Chinese MMORPG, globally released 2025

7 of the above 11 are from Chinese devs; 1 Korean; the rest are all casual genre from the west. Similar trend as last year.

A few other worthy callouts – not new entrants to the top 100, but impressive growth (>50% revenue growth) in 2025:

  • Clash Royale (no. 13, 152%)
  • TopHeroes (no. 41, 135%)
  • Seaside Escape: Merge & Story (no. 49, 67%)
  • Wuthering Waves (no. 56, 56%)
  • CookieRun: Kingdom (no. 82, 83%)

Game devs, what is your China strategy?

(This could have been its own post, but I got lazy.) With the increasing dominance of Chinese devs, it’s worth being a bit provocative towards game devs elsewhere: do you have a China strategy, from the following perspectives?

  • China as a market opportunity – is this something you care about? I’m very opinionated here. On mobile, I remain bearish for the “average” foreign studio. The regulatory barriers to entry remain very high (only 95 licenses granted to foreign games in 2025, compared to 1,676 domestic), while the home turf competition is astronomical. It’s a hugely lucrative market for the winners, but it’s arguably not even worth the effort for the rest. Steam is a whole other story, but we should remain vigilant about the continued legal uncertainty
  • Chinese talent. Chinese devs have accumulated strong know-how in specific genres and mobile. Is this capability of relevance to you? If yes – unlocking it is not straightforward (build/buy/partner?), but perhaps worth a deeper investigation.
  • Chinese games going global – from a defensive perspective, how are you stopping them from taking your lunch? Or stated neutrally, what do we think the stable equilibrium will be of Chinese games’ market share in western markets? Will this play out more like the film industry (where Hollywood productions collectively dominate the box office everywhere, aside from a few specific markets), or the auto industry (with a lot more geo-market share variance, associated with regulatory protections)? A common critique I hear is that Chinese devs still don’t have the cultural taste for the west – how strong is your conviction that they will never acquire this (or, can they acquire it faster than you can catch up in capacity deficits elsewhere)?

Personal bits

I didn’t play a lot of new games in 2025. I remember spending a lot of time on Hades II while it was still in early access, but ironically I still haven’t experienced the ending since its 1.0 release. In December I jumped on the extraction shooter train, but no, not Tarkov or Arc Raiders, but rather Escape from Duckov, the PC game made by a tiny Chinese team. It is very good, especially the first 20 hours. It’s clear the devs took to heart key design principles from not only Tarkov, but also games like Dark Souls and Diablo, and fused them together in a truly addictive package with a whimsical (and smart – as in cheap to produce) art style. And on mobile, more than anything I played a lot of Merge Tactics in Clash Royale. (It’s great to see the former Clash Mini team take another swing at mobile auto-chess.)

During the summer I started binging The Expanse. First the TV series, then the books – all of them. One thing that left a strong impression was how I felt the writing noticeably improved after the first book. Perhaps it’s just a subjective feeling, or, this is a good example of craft growth.

I kept up decent active habits (for my standard), averaging 10,882 steps per day which was a small 2.5% increase over 2024 (10,608). I kept up sports climbing, continuing at a fairly beginner level but somehow managing to sustain some over-use injury on my fingers (so I’m taking a few weeks rest currently).

  1. I’m not endorsing a laissez-faire position either: sensible (consumer-oriented) regulation of AI is needed, similar to how there is growing appreciation for regulating social media. ↩