Josh English

“Was AI involved?” is already the wrong question for games. Modern development does not happen in a clean room. The real questions are where AI entered the pipeline, what shipped, who approved it, and whether the player experience depends on it. The industry does not need a purity test. It needs a vocabulary. @jengas/was-ai-involved-is-the-wrong-question-for-games-3e613619f9a5" target="_blank" rel="nofollow noopener">medium.com/@jengas/was-ai…

@haider1 I asked it to do something stupid, and it (correctly) refused. I'm very impressed with this model

GPT-5.6 Luna deserves far more attention much smarter than 5.5, not a yes-man, and will actually correct you, and a major reason is that 5.6 Sol autonomously post-trained it so it is not just a smaller model, but its training was improved by a more powerful model, not only by human researchers

@kimmonismus I love this model, but tokens seem to completely evaporate in just a few minutes

GPT-5.6 Sol is 54% more token-efficient for agentic coding tasks, according to Sam Altman. GPT-5.6 Sol is great. But token efficiency is the one thing I just can't seem to confirm. I've never used up my rate limits faster than I do now - even on the "Pro" tier.

play it yourself, free & open source: josh-english-2k18.github.io/webgl-games/as…

Grok 4.5 dropped, so I gave it my one-shot test: build a self-contained Asteroids game in WebGL. It didn’t just reproduce Asteroids. It turned it into third-person combat around the ring of a gas giant, where you magnetically tether rocks, sling them into each other, and turn their mass into a weapon. The interesting benchmark is no longer whether a frontier model can make a game. It’s if it can find the game inside the prompt.

@BrainEngineer @sammysintech 😅

@JoshSeriesAI @sammysintech Something tells me “let our customers have six beers and sit naked in a co-ed sauna with unrestricted access to the adjacent freezing sea” doesn’t read well on an SF permit

We are going to see a huge rise in sauna experiences and related companies coming to SF. There’s a few in Canada popping up (new) and around the US that I think would do very well here. They are also often fully booked and well funded in the cities they are in. Back to basics and human connection.

@BrainEngineer @sammysintech I doubt the permits would be attainable in a reasonable timeframe and budget

@sammysintech Every time I’m drinking a beer at Löyly in Helsinki I think, “I’d kill for this in California” and, “oh my god the insurance costs in California would absolutely destroy this business”

@Nerdrotics If they make less than $600M, it's clearly in loss territory. The question is: will that be counted as a loss?

Will the Odyssey FLOP?

@kimmonismus Optical memory is a cool idea, I haven't seen it really work in a frontier scenario yet

pxpipe is interesting because it turns dense text context into images to cut Claude Code costs. But the idea is not entirely new: DeepSeek explored something very similar with "optical context compression" end of 2025 (if you remember); rendering text as images so models can process long context more efficiently. DeepSeek treated it as a model/architecture problem. pxpipe treats it as an infrastructure hack: a local proxy that exploits the fact that frontier models can already read images well enough. That makes pxpipe less elegant, more lossy, and probably risky for exact strings, hashes, IDs, or code details. But still a cool idea and i really like the approach.

@haider1 Is it exponential thinking or simply a lack of abundance mindset?

another example of how terrible humans are at understanding exponential progress in 2022, AI experts gave AI a 50% chance of proving publishable math theorems by around 2050 and winning the Putnam math competition by around 2033 LLMs did both this year

@signulll culture is the arbiter of how technology is used, and X is the jury deliberating it in real time

posting about culture is often more important than posting about technology because culture is where people decide what technology means. tech posts often explain what’s possible but culture posts decide what’s desirable, threatening, cringe, inevitable, or worth funding. when you marry the two well, it’s steroids for influence.. combined with this platform it’s the most potent worldview distribution machine ever built.

@tunguz Greatest? Lol

I saw the greatest minds of my generation get one-shotted by the AI psychosis.

I've been testing frontier model performance with a one-shot Asteroids benchmark. Nate grabbed one of my test results and instantly made it online real-time multiplayer. The question of who can make entertainment with AI is obsolete. Now it's all about who has the best idea, and how fast can they get it playable with real people?

@Devon_Eriksen_ Approaching the problem from lower in the stack strips away symbolic references (that are useful for humans) and optimizes for performance in a machine-centric way. Fascinating idea, especially given current model's verbosity bias

Okay, looks like we're going to need some Carmack-to-ordinary-mortal translation. So I'm going to take a few pages to explain what he said in a paragraph. When LLMs read code, they parse it as a linear stream of language nuggets. But computer languages are basically designed to describe a sort of tree-like structure of nested meanings called an "abstract syntax tree". An LLM can emulate an understanding of this by its trained ability to parse specific characteristics of the language, but that only goes so far due to the limitations of the model. Humans, by contrast, have the ability to actually construct pieces of the tree in our heads, giving us a much more exact understanding of code. But we're limited, because we're slower, especially in typing, and because we can remember fewer things at once. Carmack suggests that it might be worthwhile to train an AI type that reads abstract syntax trees directly. Such an AI would first invoke the front end of a language compiler, to parse code into an abstract syntax tree, just as in the process of compiling code to binary. It would then operate on the AST itself, rather than on human-readable code, and then use a reverse-compiler like process to turn its result back into code, for human reading and review. How transformer architectures might work on a tree is a little bit of a head-scratcher. I'd want to play with graph transformer type models, but I'm reluctant to say anything more than that off the cuff, because my knee-jerk approaches to the problem might turn out to be egregiously stupid. Would this be useful? I don't know, but Carmack thinks it might be worth a try. However, it might turn out that linear stream transformers have an emergent "understanding" of tree structure if you throw enough GPU cycles at the problem. A highly correlated question is how human brains code. I haven't looked deeply into this, but Grok claims that fMRI studies show that human coders do not primarily rely on language centers (Broca's area, etc) for coding, but show a lot of deep-brain activation. If it isn't hallucinating, Carmack may well be right.

@pmarca the operative term in "...catch up soon" is: soon = several years

Interesting.

@pmarca So many good things from the 90s aren't around anymore

You kids can’t possibly imagine what WIRED Magazine used to be.

Was AI involved in making this game? Of course it was. Modern game development already happens inside an AI-saturated toolchain: search, email, documents, spreadsheets, code assistants, asset tools, localization pipelines, analytics, and even the game engines themselves. No serious studio is building clean rooms where developers are forbidden from using modern search, productivity software, IDEs, or engine tooling. So tagging a game on Steam does not reflect reality. Instead of asking if AI touched the game, the better question is where AI entered the chain, what it produced, whether that output shipped, who approved it, and whether the player experience depends on it. An AI-assisted workflow is not the same as AI-generated shipped content. AI-generated shipped content is not the same as live AI runtime generation. And live AI runtime generation is not the same as building an entire game around AI as a design pillar. Those are different creative, legal, ethical, and commercial questions. We do not need AI disclosure. Because “was AI involved?” is already obsolete. Of course it was.

@AndrewCurran_ @IamEmily2050 Certainly been leaks that support this

@IamEmily2050 If the rumors are true Google has had a version of Veo that could gen 20 sec+ clips with better quality for a very long time now, but didn't want to serve it due to cost. But yes, they intentionally slowed down and it looks like bytedance is going to benefit from it.

The people who decided to slow down the development of image and video models in the West because of Hollywood and backlash from some policymakers will pay the highest price in the generative AI era. History will not be kind to you, and your excuses in the future will not even matter.

@beffjezos Don't rock the product roadmap boat

In my experience at big G, nothing will piss off leaders as much as actually shipping something fast.

@tunguz It directly (out)competed with Google products, q.e.d.

Yeah, Big Tech doesn’t like it when you take initiative. Not one but. And if you tweet about it you are screwed.