2001zhaozhao

This fairness vs hardware control dichotomy isn't unsolvable. I believe that I have personally found an extremely good technical solution to this problem which will debut in my upcoming FPS game supporting the browser and other platforms, allowing things like reliable client-side anti-ESP (yes you heard that right).

It isn't as good as kernel level AC in some sense but is equal in others - it's easier to cheat at all on my AC because you don't need hardware, but it limits cheaters' opportunities enough that my AC + server-side detection is equal to kernel AC + server-side detection at catching any given cheater.

This seems like a big deal, no?

If Qwen had released a Qwen3.8 MoE of similar size, with a naive extrapolation its capabilities would roughly be similar to this model. So basically someone has caught up to Qwen (albeit Qwen from 1-2 months ago) in parameter efficiency.

Part of me would like to add tracking of human/AI provenance for every line of code in my future game the way I want to do it for in-game art. So when a human hobbyist modder writes a line of code by hand and open sources it, my platform shows that it's human (or it's someone disingenuously and deliberately writing off AI work as human, ideally with no in-between).

But even the very design of modding APIs itself is changing when expecting coding agents as the main consumer. So it might not even be possible to cheaply support human programmers at some point. For example perhaps some trusted APIs are inaccessible to human developers and must be programmed by AI with particular prompting to never generate code that abuse said APIs.

This is interesting to me, as I was thinking of adding a Git server to my AI orchestrator when I eventually need to split agent containers off of the main machine (to make agent CPU horizontally scalable).

Maybe I should just bite the bullet and use a different VCS system entirely once I get there, which has most of the Anthropic system's properties (including being much faster) but doesn't need to be horizontally scalable.

Yeah I think we are slightly ahead of the prediction on coding capabilities right now. AI 2027 predicts Superhuman Coder in March 2027 but I think my expectation for it is 1-2 months earlier right now

Yep we're basically inventing all of the RSI loops that AI 2027 thought we'd invent lol.

Pretty cool. This has a fine tuning step so it's not just a harness optimizer. I think something like this loop is the end game of "automatic RL environment generation"-adjacent efforts, but I expect the challenge will be how to properly keep the training on rails and not Goodhart into a stupidly irrelevant metric.

What does this do that giving the agent a wayland session in Linux doesn't?

Tbf this is THE way to get cost-effective AI into video games (think enemy AIs and NPC AIs that are much better than current ones and runnable on consumer GPUs), i'm really looking forward for decision models maturing.

Agreed, the state of social media today is basically a result of market failure.

I think the point of this product is not the quality of the games, it's the accessibility and distribution and the fact that it's free to use.

Big tech already controls the distribution of traditional platforms so they are building AI platforms to make sure they keep control of the distribution of new forms of content, including that for AI games now that they're easier to make than a YouTube video.

But inevitably these products (and the competitors they will inevitably spawn) put us closer to a future of arguably dystopian, AI-driven hyper-entertainment world, even if no one inside the companies meant to do that. I can only hope in that future many people still appreciate human-made games, and even on these AI platforms genuine human ideas (assisted by AI) can still bubble to the top.

As for game quality, these look close to Opus 5.5 one-shots to me, maybe even a little worse.

Roblox has already been pushing AI-generated games heavily.

For what it's worth, I'm thinking of some version of this for my upcoming game's modding tools, although the backend processing will be on a web server unlike this demo, and there is much more of an AI agent focus.

I don't think your take is counter to the article. It is precisely the "founder giving a shit about everything" that will be amplified the most by a capable AI harness, because AI reduces the power of capital (e.g. hiring a lot of good engineers or buying a lot of data or ads) as a moat by amplifying good judgement.

(This is temporary until the AI gets better judgement than humans, then capital will therever be the most powerful moat in a market full of dystopic, consequentialist, incredibly long-sightedly-greedy companies)

> Distribution is one of the smaller ones left, but the personal software trend might eat that too

Being a platform for personal software is gonna be valuable, but it needs a lot of trust. (I have a nonprofit idea around this right now)

Btw, I think distribution might temporarily become less important (because with better AI you can actually pull so far ahead of competitors quality-wise and therefore succeed despite a distribution drawback), but long run it actually becomes more important because of AI persuasion and commodification? If you are the super app then, well, you are the super app

Yep private scientific knowledge is a massive new moat that you can pull if you simply invest in scientific discovery methods in general and throw enough resources and tokens at the problem. I think there are already startups specifically trying to do this. The main obstacle is whether you're actually able to pull significantly ahead of (AI-enabled) public science to make a difference, but I guess the math works out if you're sufficiently AGI-pilled.

I agree this is a kind of data moat, but it's also arguably distinct enough to be its own thing.

You are assuming that getting the AI to generate correct (or accurate, representative) data will be very difficult. I would agree, but I think it will become possible in the long run.

(Edit: alternatively you just use AI to get rid of the need for data to solve a problem, like Jev did for traditional classification models)

I think current incentives definitely go against any efforts to build this. It's very hard to build this and be rewarded for it by, say, investors or your boss, because you can't really prove that your system is non-sloppy while your competitor's is (even if being non-sloppy is all that matters), because by definition your novel results are not verifiable or else the model labs will have already trained it into their model.

But the same is true for high-quality AI systems in general. In general, I think AI model advancements will make the systems easier and easier to build until some small guy accountable to no one but themselvs can build it, and then it will actually be built.

This couldn't be more true. Companies will be driven by AI harnesses (as defined by this article) that automate decision-making and prioritization. In the medium term, may the company with the best harness win.

In the longer term, the downstream impact is massive commoditization of software and invalidation of most existing moats. Data moats are gone if you can simulate the data with AI. Even platform effects can be sidestepped if AI replaces one side of the platform.

In addition, while right now agile startups have the advantage, at some point the balance will start tilting towards whoever has the most tokens (OR perhaps durable moats will trump even near-infinite tokens; we will have to see). Startups have a limited time window to have whatever impact in the world they are hoping to have, or to build a moat that won't be disrupted by AI, but there are few of them left in the world.

The upside is that when there is a lot of commoditization, then the consumer benefits.

Is this like a spring boot for typescript or something like that? Looks kind of interesting

I've personally been using a Phoenix LiveView-like approach extremely successfully, combined with LLM coding. It is extremely, stupidly fast. We are talking about loading an entire multi-window desktop environment in the web on spotty 5G networks in half a second, transferring less than 1MB of data.

I think traditional web frontends might be dead for any applications with servers that can afford to keep server-side state of a logged-in user.

(Notably this includes almost any kind of AI application, because the LLM costs dwarf the web hosting costs anyway.)

And for the stateless ones, there's always htmx

FINALLY, I can support Pi in my orchestrator system that relies on integrating agents with my custom MCP server.

I think there will be a lot to be learned if we can scale up the high resolution measurement methods here to map exactly how brainwaves travel when performing specific tasks.

Maybe they should recruit experienced meditators who can give accurate introspective self-reports to be compared to the measurements, then we would start to have a relatively confident map between brain physiology and psychology. (e.g. X brain pattern in Y region specifically corresponds to the Z step of reasoning when solving the problem)

I'm pretty convinced now that grading take-home assignments is pointless with the prevalence of AI cheating. (There's surely already "innovation" going on targeted at letting people cheat on assignments while staying undetectable.)

It's probably best going forward to grade only in-person proctored exams using paper, or offline air gapped computers in the case of programming courses.

Afaik the usage quota nerf to $200 Pro is true. Twitter is all over them right now. (Also you haven't been able to buy the $200 Pro plan for the last two weeks or so)

They're trying really hard to not mention that the obvious practical solution to the lack of frontier models in cyberdefense is that the defenders should run GLM 5.3 themselves.

Perhaps they should do something like remove dual-use cyber safeguards on older models as soon as open weight models of a similar capability are released.

> It's also in their best interest to do so, because if they don't and these provider hosted agents become the norm, demand for independent inference would drop.

I disagree with this part. AI companies will want to try their damndest to control distribution of AI, so that they can enshittify later.

Consumers conscious of this will want an alternative, of course. Might be niche similar to how Kagi is in search because big tech will always have a AI inference cost advantage + making users the product (extra $ from ads & purchase cuts) + the good old strategy of dumping.

Lock-in is easy with these agents because they NEED all of your data to be useful, and they will continue to learn internally about you.

But ultimately the AI company CAN choose to just make all the data exportable and open source their product for self-hosting. (The mainstream ones won't, of course, they want to lock you in and hide their AI prompts and algorithms.)

I think what is sorely needed is a version of Dots/Muse without lock-in risk but is still accessible to regular people unlike Openclaw.

...did you just describe openclaw?

If the AI they use is reliable, this is actually pretty cool.

> indie game developers have stopped the traditional practice of updating a dev log as a form of marketing and community building because people will point their LLm at it and create your own game before you do

And that's why we can't have good things...

(This is just gonna keep happening more and more until eventually we'll need something like a patent system for ideas)