Humanity has prevailed (for now!)

I'm completely exhausted. I figured, I had 10h of sleep in the last 3 days and I'm barely alive.

I'll post more about the contest when I get some rest.

(To be clear, those are provisional results, but my lead should be big enough)

staying at an airbnb
took me 10 minutes to find the switch for the lamp above the bed

this is some escape room-ass level design

when streaming Order of the Sinking Star, I try to explain my thought process whenever I find a solution interesting, here's one example

notes:
- that was near the end of my stream and I was running very low on energy, so you might want to watch this at 1.5-2x speed; sorry for all improvised
- if you don't know the rules: the goal in this level is to remove all of the goblins, they kill each other then they're on adjacent tiles; we can move them around by teleporting them through the mirror; our hero is rogue and unlike in the classic sokoban, rogue pulls objects instead of pushing them

a small announcement: I'm almost done with a puzzlescript-based benchmark; you can think of it as ARC-AGI-3.5, maybe even ARC-AGI-4; I expect it to be the most reliable way of assessing agents' abstract reasoning skills in a game environments

I went through over 300 games, handpicked the most interesting and diverse set of 40, fixed a ton of bugs, went through multiple iterations of prompt design, etc

most games can be completed in around 1h, some of them are significantly harder; models are tested using standard harnesses (codex/claude code) with no code generation allowed; action limits are very generous and unfortunately the benchmark is already somewhat saturated, but the models are still insanely action inefficient

FYI, I won't release the game set to the public, but I guess I might share runs from some of the rejected games.

btw, if anyone knows, what do I have to do in order to not pay $10K out of my own pocket for testing, I'm all ears

if you're wondering why suddenly I was motivated to finish an ambitious benchmark: I'm procrastinating doing other things in my life; turns out I'll do anything except for preparing for an interview

Remek is currently working on a decision model (i.e. "system one" model, similar to Jev / OpenAI's Decisions API) and he's currently ranked #1 in open weights (#2 overall!) on JevBench, essentially beating almost every commercially released product, including Jev itself that went viral some time ago

what's even crazier is that he's working on it alone, as an after-hours side project, he started a few (two?) weeks ago and he's paying for compute out of his own pocket

I suspect that one of the biggest reasons why almost no one pursued LLMs despite GPT-2's overnight success was surprisingly simple: almost no one was excited to work on language

even at OpenAI, people didn't want to leave their cool game environments behind

you can laugh at this, but I'm pretty sure that if Dendi had won, OpenAI would've raised far less money afterward and AI progress would've been slower

back then, OpenAI was a relatively unknown research lab (~50 people) and this was the first time (after the initial launch) OpenAI made it to the mainstream news

a few random facts from the development:
- the bot was still not good enough to beat the pros a few days before the showmatch; we were still making last-minute changes during The International; in particular, I was making a ton of changes to feature extraction, while Jakub & Szymon were surgeoing old weights to the new model - we didn't have time to test our changes, so we literally pushed everything directly into "production" because we only had enough time for one training run; it worked
- we were aware of multiple exploits you could do vs the bot (our training env wasn't general enough and you could do a lot of weird things like luring the bot to other towers); that being said, I was sure that it's nearly-impossible to win vs the bot in a straight up micro battle - as long as Dendi tried to play vanilla 1v1, I thought we have 98%+ chance of winning
- a lot of pros thought that we have "solved" the items that heroes should buy in 1v1; in particular the bot was buying a ton of consumables; AFAIK pros were later inspired by the bot and the meta shifted to buying more consumables in the midlane; the truth is, we just hardcoded the item order and never had enough time to explore other options

a few months ago I was extensively testing codex/claude code on the public demo of ARC-AGI-3; then astra came, obliterated the benchmark and I lost the interest

originally, I started testing it, because I was tired of the misleading "Humans score 100%, AI <1%" narrative from the ARCmpletely failed at tracking the pace of the AI progress

I've never shared the results of those tests so here you go

what I tested:
- openai models with codex or ant models with cc
- agents were sandboxed, had no internet access, all games were independent
- agents were free to use the tools (e.g. create code)
- prompt is a variation of PRO-LONG
- agents could batch multiple commands per call; this drastically speeds up testing and reduces the costs; this has a small downside that the agent has in practice fewer turns to "think"
- I've used the same scoring (RHAE) and the same action limit (5x of human baseline) as the official ARC-AGI-3 does

some caveats:
- runs were performed at various moments, using different versions of codex/cc; this means that quite often, used agent harness might be much more recent than the model
- at least since Opus 5 / Astra 6, models generally have a very good understanding of ARC-AGI-3 env;
- variance is high, but the relative order of models should be more or less stable
- remember that this is the public demo and the verified benchmark uses private games; that being said, those two should be very close and in general I saw no signs that models have memorized the public games

graphs:
#1: RHAE vs release date; you can see the steady improvement of the models here; ARC-AGI-3 was released soon after gpt 5.4
#2: RHAE vs cost (per game); official testing had astronomical costs (20-100x higher than mine); honestly I don't know how they generated such costs - batching & code execution doesn't fully explain the difference
#3: RHAE vs official RHAE; both values are on the public demo; you can see how little correlation there is between ARC Prize official "minimalistic harness" and a realistic use with a regular coding harness (e.g. codex/cc)

one thing maybe worth mentioning is that I had a ton of additional runs testing variations of prompts, context limits, reasoning level and few other properties; pretty much all of those things made no significant difference on the results

imho the only thing worth mentioning was that apparently a big part of the generated code was used for parsing the visual input from games; I experimented a bit with creating a special limited language for parsing the visual input and when agents were limited to using it, they were nearly on par with runs that had full access to code generation

unfortunately some notable models are missing (fable, opus 4.6 & 4.7), but I've moved on to testing other things so I'm definitely not revisiting this benchmark

I decided to make this livestreaming a semi-regular thing:
- I will continue with my playthrough of Order of the Sinking Star
- the next stream is in 2 hours from now (18 CET / 12 ET / 9 PT); tomorrow I'll do another one at a similar time; maybe another one on Friday and/or Sunday
- so far the streams are a mix of me not being able to solve trivial puzzles and mini-AMA about AI
- I don't want to spam twitter with my stream announcements; if you want to join any stream, the most reasonable way is to give me a follow on twitch: twitch.tv/fakepsyho
- unfortunately, twitch is being an ass and doesn't allow creating non-recurring schedule without being an affiliate (which gives ads to viewers, so I'm not doing that); I'll put my planned schedule somewhere on the twitch page
- I'm treating these streams as a lifehack to practice my English; after the stream, I get the automatic transcriptions from my mic output, feed it to AI and then get personalized lessons
- not sure if twitch was the best choice for the platform, but at this point, I'm probably sticking to it

protip: if you want to use your recordings for AI-generated English tutoring, make sure you don't mix your mic output with the game audio

Next week I'm speaking at two conferences in Poland:

Game Industry Conference:
there will be a fireside chat with me about (I presume) the future of AI and it's impact on the gamedev industry

Warsaw AI Safety Day (first AI Safety conference in Poland AFAIK):
I will be a part about less obvious AI risks (so I guess no x-risk / concentration of power / misuse)

come and say hi to sleep-deprived me if you're around

Links to the events:
- gic.gd
- warsawaisafety.day

POV: your twitter feed after you interact with a single post

I'm tired that for the past several months, 80% of the twitter feed is posts on the same 2-3 topics that you've recently interacted with

would be nice to have an "exploration" slider that you can use to adjust how much information you'd like to see outside of your bubble

no matter how many (controversial) predictions Dario gets right, I'm sure he will still get a ton of shit every time he makes another one that people don't want to hear

not particularly happy with the caption, considering he hasn't made that many predictions (and thus it's hard to eval his accuracy)

my more general gripe is that there's surprisingly little correlation between people's forecasting track records and how seriously we take them

Order of the Sinking Star was released 2 days ago and I decided to continue streaming my playthrough. First stream in ~2h from now (21:00 CET, 15:00 ET, 12:00 PT).

I've upgraded my setup with a greenscreen, so hopefully my camera won't cover important parts of the levels as when I streamed the demo:
- I'll talk about puzzle solving / game design while playing
- feel free to ask questions, they don't have to be related to the game
- my main goal is to practice English and to decouple the parts of my brain responsible for talking and performing any mental work (right now I can either talk or play the game, I have no clue how people can do both at the same time)

I will resume where the demo ended. The only thing I know about the full release is that they haven't fixed the walking animation, so expect me to complain about it a lot.

Link to the stream: twitch.tv/fakepsyho

low-stakes conspiracy theory: what if Anthropic believes that in the future superintelligence might judge us based on how we treated early AI models? hoping the machine god will spare them

for the record, I don't have strong opinions on AI welfare