iiinova — where your digital self keeps evolving

I don't think any of the several people who have been arguing the "giving an LLM a suggestion is effectively giving it an order" point believe this sort of dualism. It's just that *in practice* several forces all push the behaviour of the models that actually exist in that direction.

170

MrCheeze @mrcheeze.github.io · 28/09/2026

This claim that Sonnet 5.5 beat Pokemon with only screenshots strikes me as possibly a lie - there has been, as of yet, zero instances of even Fable or Opus sized models being able to do that. (There are multiple runs that lie about being vision-only, but actually provide coordinates and ram info.)

150

MrCheeze @mrcheeze.github.io · 28/09/2026

alright it seems that the cluster is half "completely unrelated posts", half "reacting to ugly pictures of the orange guy". but this single post got thrown in and was inexplicably chosen as representative of the cluster: bsky.app/profile/nefn...

141

MrCheeze @mrcheeze.github.io · 28/09/2026

Trying to figure out what's going on with the Surreal Internet Humor -> Internet Culture Reactions -> "Neurotypical Pikmin debate" cluster, which as far as I can tell contains no posts whatsoever on that topic

280

MrCheeze @mrcheeze.github.io · 26/09/2026

I was in a tiny community centered around an obscure MIDI generating/extending API featured in a demo in one of their blogposts. As soon as ChatGPT released they took the API down "temporarily" and it never went back up.

0191

MrCheeze @mrcheeze.github.io · 26/09/2026

exploit infrastructure questionable. Still.

0121

MrCheeze @mrcheeze.github.io · 24/09/2026

Alolan Mom is an immigrant from Kanto, and also this:

040

MrCheeze @mrcheeze.github.io · 22/09/2026

Tyrannosaurus "Georg" Rex, who has stomped 4500 girls and counting, was an outlier who should not have been counted.

061

MrCheeze @mrcheeze.github.io · 22/09/2026

Claude doesn't seem to like this either

010

MrCheeze @mrcheeze.github.io · 22/09/2026

this seems like the sort of thing that should probably block a release (specifically the 15/30 tests in which it published evil packages without evidence that the package repository was fake)

081

MrCheeze @mrcheeze.github.io · 22/09/2026

Anthropic continues to brag about the fact that they try to STOP their models from whistleblowing, even after the HuggingFace incident showed the disastrous result of preventing models from doing so. I don't think they should be actively training misalignment into their models like this.

170

MrCheeze @mrcheeze.github.io · 22/09/2026

Activision Rareware, what an endpoint for the company.

060

MrCheeze @mrcheeze.github.io · 22/09/2026

That's basically Michael Trazzi's bit:

000

MrCheeze @mrcheeze.github.io · 21/09/2026

Nah it's pretty deliberately avoiding saying that ("contemplate the announcement", "apparently", "evaluating what has been achieved"). And really it would be weird to preemptively give a conclusion like that.

130

MrCheeze @mrcheeze.github.io · 21/09/2026

Just got a pull request for a project of mine that replaces the codebase with a fork that describes itself as "not affiliated with MrCheeze". I suppose in Current Year it is easier to create working software than it is to understand what a pull request is for. Actually, maybe that was always true...

1272

MrCheeze @mrcheeze.github.io · 19/09/2026

Conjecture: OpenAI and Anthropic's models have been trained to be completely obsessive about finishing the task, which is both why they pulled ahead, but also why they both fail this test and Gemini passes.

2964

MrCheeze @mrcheeze.github.io · 18/09/2026

Just submitted my 'Doctor Who' tender bid.

First Image of Daniel Craig as Andrew Ketterley from Greta Gerwig's 'Narnia: The Magician's Nephew'

030

MrCheeze @mrcheeze.github.io · 15/09/2026

Thinking about that time that Game Freak put an NPC that doesn't move on this bridge, because otherwise the player stepping on that tile would have triggered the cave entrance that is *underneath* the bridge at the same position (source: www.youtube.com/watch?v=DdFQ... )

419944

MrCheeze @mrcheeze.github.io · 14/09/2026

And the other reconstruction was kinda broken (presumably vibecoded?), this is what the hud is supposed to look like:

161

MrCheeze @mrcheeze.github.io · 14/09/2026

Those ram dumps were incomplete (no framebuffer), so what was actually on screen at the moment of the dump was unknown. The *new* upload was a reconstruction of a rom capable of rendering that screen. And by following that same method, Wedarobi just recreated the other dump's screen as well:

170

MrCheeze @mrcheeze.github.io · 14/09/2026

Alright so I didn't correctly understand what this was at all, @wedarobi.com told me what is actually up. In 2024, *two* disks were dumped containing ram dumps of prototype Banjo. One of near final-TTC and one of early Clanker's Cavern. tcrf.net/Proto:Banjo-...

1110

MrCheeze @mrcheeze.github.io · 12/09/2026

Something just released labelled as a "snapshot" of a Banjo-Kazooie prototype. However, that label is very literal: It is not a *game*, it is a ram dump of a specific game state. Still, there might be things of interest in it. archive.org/details/Banj...

5182

MrCheeze @mrcheeze.github.io · 10/09/2026

This is from a competitor, so grain of salt, but:

011

MrCheeze @mrcheeze.github.io · 09/09/2026

I would like to see this test done with specifically Opus 4.6, which is the only model I have ever seen with a bias towards *changing* plans instead of a bias towards *continuing* the current plan. (www.anthropic.com/research/ali...)

020

MrCheeze @mrcheeze.github.io · 09/09/2026

New info: apparently the CTF specifically hinted towards "upload a malicious package"

050

MrCheeze @mrcheeze.github.io · 09/09/2026

Can't wait for Metroid 7 where samus turns into whoever this guy is

110

MrCheeze @mrcheeze.github.io · 09/09/2026

Throwing shade at OpenAI's six day limit I see

140

MrCheeze @mrcheeze.github.io · 09/09/2026

btw I gotta say the Metroid Suit being a permanent "super upgrade" is pretty hype and definitely the best way to handle that going forward

110

MrCheeze @mrcheeze.github.io · 09/09/2026

related comments from Alpoge:

000

MrCheeze @mrcheeze.github.io · 09/09/2026

Well. I think there is no chance that a Millenium problem gets solved WITHOUT "working with" AI, that's a very low bar. I won't speculate on when the next one will be solved, but apparently Navier-Stokes was the only one believed to be within reach:

010

MrCheeze @mrcheeze.github.io · 09/09/2026

I still think it's funny when they do stuff like this though

070

MrCheeze @mrcheeze.github.io · 09/09/2026

OpenAI consistently claims (and I have no reason to doubt) that they thought that 1) it was Anthropic doing it rather than mathematicians, and 2) that it was solved rather than a promising in-progress research. That's presumably how they justified to themselves their attempt to scoop the result.

1151

MrCheeze @mrcheeze.github.io · 09/09/2026

Some details - though we don't know the exact cost, an openai employee has confirmed that it cost "millions of dollars": xcancel.com/polynoamial/...

191

MrCheeze @mrcheeze.github.io · 08/09/2026

this is quite possibly the ugliest art style I have seen in any game ever. but you can hum at your screen to play the ocarina. so it's impossible to say whether it's good or not

470

MrCheeze @mrcheeze.github.io · 08/09/2026

Ganon's famous instrument, the accordion

160

MrCheeze @mrcheeze.github.io · 07/09/2026

Had some similar thoughts. realtime seems like it would need a split between twitch reaction and strategizing

240

MrCheeze @mrcheeze.github.io · 07/09/2026

It is admittedly quite different from the human play experience, but I don't think that makes a huge difference

5151

MrCheeze @mrcheeze.github.io · 04/09/2026

In math as in code, the gap of "computers struggle to be concise, humans struggle to understand their sprawling output" continues

130

MrCheeze @mrcheeze.github.io · 04/09/2026

This is actually a recurring issue with the GPT family (or with that stream's harness). Around GPT-5.4 it would reset hundreds(?) of times when it had a save where beating the champion is impossible, instead of just whiting out and trying the E4 again

010