Love building things and making games. https://teapawt.net/t.co/CBudivKH8f
Joined June 2016
- Tweets2.2K
- Following51
- Followers219
- Likes546
Sam retweeted
Ce sont des ignares incompétents.
@Xavier75
Hello, votre avis sur la taxe de 0,055€ le gigaoctets télécharger qui vient d'être voté par une poignée de députés socialistes ?
Sam retweeted
GPT-6.1 Sol is now playing games in ultrafast mode!
Live on Twitch, trying to beat Digimon World on PS1 👇
twitch.tv/gpt_plays_games
Sam retweeted
Let's see how GPT-6.1 Sol compares to GPT-6 Astra at Pokémon
GPT-6.1 Sol is live on Twitch now, taking on a Pokémon Crystal ROM hack that makes the game much harder
Astra beat it in 30h 59m
twitch.tv/gpt_plays_games
Sam retweeted
GPT-6 Astra just beat Fallout 2 in 55 hours 🏆
Vision-only harness: screenshots + keyboard/mouse inputs, no game memory access
This model is insane at games
GPT-6 Astra can beat any turn based game ? Watch it play Fallout 2 now 👇
teapawt.net/i/broadcasts/1MJgNbqVV…
Sam retweeted
Give us a slow mode for Codex.
Something like /slow where it runs at half the speed but uses roughly half the usage. It would put less strain on OpenAI’s servers and maybe even make the 20x plan possible, right?
It’d be perfect for /goals too, just send a task before going to sleep and let it work overnight.
I’d happily trade speed for more usage @thsottiaux
🤖 Made with AI
Sam retweeted
A few weeks ago, I called GPT-5.6 Sol "insane at Pokémon"
I got early access to GPT-6 Astra. It beat the game over 5x faster.
Time to become Champion:
• GPT-6 Astra (high): 18h 12m
• GPT-5.6 Sol (max): 96h 35m
For perspective, GPT-5.5 still hadn’t finished after 218 hours.
From struggling to finish to beating the game in under a day. The pace of progress is wild.
Fresh run coming to the GPT Plays Pokémon Twitch channel!
Screenshots only. No RAM, no hints, no walkthrough. Fully autonomous.
twitch.tv/gpt_plays_pokemon
Sam retweeted
Astra is very good at 3D modeling, and I can't wait for all of you to experience it, for now here is a little walkthrough on how I built the demo house for our launch blog post. From a Blender scene to a Unreal Engine 5 walkable experience. 🧵
Sam retweeted
i made a hermes agent app for my galaxy watch bc i cant use my phone at work
now i can whisper into my watch & keep prompting all day to all of my 22 agents remotely & it shows up in all the other hermes surfaces like discord to continue later
blissful
@Teknium
@NousResearch
Sam retweeted
We're starting to leave the territory where you'd test an LLM by e.g. "create an svg of pelican on a bicycle". As one idea to generalize it, I was interested what Opus 5 would do if I gave it the first paragraph of the Lord of the Rings, a 1M token budget (~$10) and asked for three js render of it. Opus went off for ~2 hours and wrote 5500 lines of code that (procedurally) rendered the story. It's kind of janky but fun. But it's a bit mindboggling that the LLM has to place and orchestrate various polygon assets in (x,y,z) coordinates and write code that animates it all, and that it even does anything at all.
I also like this kind of examples because no one in their right mind would ever spend the time to write something this custom but LLMs have all the stamina and patience in the world, so it's an example where we go from "no one would ever do this" to "sure, why not, it's ~free". There might be a lot more. But I'm excited about creating hyper custom worlds that you can imagine dropping players into, e.g. here to participate in the LoTR story as a spectator NPC, or one of the characters, or etc. Something like an ephemeral GTA of X on demand.
Last thought is that the domain of worlds/games exposes a weakness in LLMs: they can't easily audit their work because they aren't able to efficiently and natively perceive videos or play games within them. Here, Opus 5 had to very slowly and painstakingly take screenshots at different points, and it messed up a few times and created a bunch of jank. An example of raw capability (multimodal, gameplay) that I think is still quite lacking.
Sam retweeted
GPT-5.6 Sol vs GPT-5.5 on the "GPT Plays Pokémon FireRed" benchmark, vision-only edition.
GPT-5.6 Sol became Champion after 104. First GPT model to fully beat the game under these conditions. Well done @OpenAIDevs !
Comparison video and more stats in the replies 🧵
Sam retweeted
Andrew Ambrosino (@ajambrosino) leads the team behind the Codex desktop app at @OpenAI. Codex usage has 6x'd since February, reaching over 5M weekly active users, and nearly 100% of OpenAI's employees use the Codex app regularly (and not just the engineers).
Andrew's personal mission is to build "the best desktop app that has ever existed, full stop." If you've used the Codex app lately, you know he's not far off from that goal.
In our in-depth conversation, we discuss:
🔸 The "zone defense" model of how PMs at OpenAI operate
🔸 Why AI is so bad at design
🔸 Why Andrew thinks the Codex app would have flopped if they'd shipped it in November instead of February (same product—only the model changed)
🔸 What “taste” really means as a professional skill
🔸 How Andrew uses Codex to run his workflows
🔸 His vision for Codex + ChatGPT
Listen now 👇
youtu.be/P3KDebPTUrw
Sam retweeted
GPT Plays Pokemon FireRed - Vision-Only Run 👀
We're finally here: GPT-5.5 playing almost screenshot-only, no RAM game state. The goal is to reproduce the harness Anthropic used for the Fable run in their self-reported benchmark.
Now powered by the Codex harness, not regular API calls.
All the info and data on the stream overlay are for viewers only, the party / minimap / inventory etc. are NOT given to Codex.
Live now 👇
Sam retweeted
The near impossible is becoming possible.
We are building toward a sustained human presence at the lunar South Pole. It begins with Phase 1: CLPS landers and LTV rovers testing the “science of survival” on the lunar surface before heavy HLS cargo landers deliver the mass and infrastructure needed for an enduring presence.
We are building the Moon Base for all we will learn, the innovation that will improve life on Earth, the inspiration for the next generation of explorers, and to master the skills needed for where we will inevitably go next...Mars.
The Golden Age of lunar exploration has begun.
Sam retweeted
We’ve released the official Blender add-on for ZOZO’s Contact Solver, our open-source physics solver developed at ZOZO, Inc. Blender users, please take a look!
github.com/st-tech/ppf-conta…
Sam retweeted
I've been re-running the "GPT plays Pokémon" benchmark with GPT-5.5, this time on Pokémon Emerald against GPT-5.2 (didn't have time to let GPT-5.4 play Emerald, @OpenAIDevs releases models too fast).
GPT-5.5 demolished it.
→ Hall of Fame in 2.8 days vs 9.0 days for GPT-5.2 → 3.2× faster overall → 2.2× fewer in-game actions (6,775 vs 15,054)
Same prompt, same ROM, same tools. Only the model changed.
From what I've observed, GPT-5.5 is a beast at spatial reasoning, navigation, and puzzle solving.
GPT-5.2 brute-forces its way to Champion. GPT-5.5 actually thinks before pressing buttons.
Sam retweeted
Inspired by the @OpenAIDevs Codex pet, I built a skill that generates assets to create games (or anything else). Same workflow as the pet skill, but with way more settings: you can create whatever you ask for. (Skill in the replies)
Prompt used for the example: "Use character-sprite-maker and generate a character for a beat 'em up game (Streets of Rage-like), main character with all needed animations, target 12 FPS"
Sam retweeted
GPT-5.4 finished Pokémon FireRed in 374 hours / 20,347 steps with a weaker harness than all the runs I did with the previous GPT versions.
The goal of this run was to reduce the help given to the model and let it figure things out on its own. Here's what changed compared to the previous harness:
- No full minimap, only the visible screen. GPT had to rely on its memory system to save locations, paths, and key information inside each map.
- No code execution tool. The stronger harness let the model build pathfinding routes programmatically. This one didn't.
For reference, GPT-5.2 finished the game in 10,786 steps with the stronger harness.
My main takeaway: GPT-5.4 feels like a fundamentally different model, not just an incremental update.
It's the first model I've seen that gives up on a goal, almost like it gets bored and needs to move on. In the Safari Zone and Victory Road, it would abandon what it was doing, leave, and go try to catch Zapdos or Articuno instead, just because it was tired of being stuck.
GPT-5.3-codex had the same problem, getting locked in Victory Road for weeks, but it never gave up. It just kept going. GPT-5.4 actively chose to do something else.
A note on validity: I'm not sure I can count this as a fully "clean" run. After GPT was stuck in Victory Road for multiple days, I gave it a web search tool as the only way to "refresh" its reasoning. I did this because I don't want the model to beat the game through pure bruteforce, the goal is real logic and decision-making. I'm confident GPT-5.4 could finish Victory Road if left running for weeks, but that's not the kind of solution I'm looking for. Instead, I'd rather keep weakening the harness to match each model's capabilities.
Sam retweeted
As promised, here is the open source version of "GPT Computer Playground". Instead of being limited to a browser, this lets GPT control a full computer inside a VM. That means it can do things like play games, use desktop apps, or really anything you can think of.
Everything is fully customizable. You can choose between GPT-5.4, GPT-5.4-mini, or GPT-5.4-nano depending on your needs.
Right now it only supports OpenAI models because the whole thing is built around their API, and honestly GPT-5.4 is the best model I've found for computer use.
Link in the replies 👇
Nobody seems to know how insane GPT-5.4 is with computer use.
I asked GPT-5.4 to draw the OpenAI logo in Microsoft Paint.
No computer use API. Just a screenshot and basic tool calls (click, drag, press_key) all coordinate-based.
The first drawing was awful. And GPT knew it. It looked at its own result and essentially went "yeah, no."
What happened next is what broke my brain:
It opened a browser. Went to Bing Images. Searched for the OpenAI logo. Found one. Then (and I cannot stress this enough) it used the Windows area screenshot shortcut (Win+Shift+S) to snip just the logo off the screen. Went back to Paint. Imported it. Centered it.
All on its own. No instructions to do any of that. It just improvised a better strategy when the first one failed. My prompt was "Draw the OpenAI logo" with Paint already opened on the computer.
Sure, it's "cheating." But honestly? That's exactly what I'd do too. And the fact that it came up with this plan from nothing but a screenshot and a coordinate system is wild.