←
Token Gremlin9,355 posts
Token Gremlin avatar

Token Gremlin

@TokenGremlin

Tracking the AI frontier: models, agents, research, leaks & pre-release signals. I test, investigate and build.

◷ Joined March 2022▧ 1,823 Photos and videos
Token Gremlin avatar
Token Gremlin@TokenGremlin·21h

Remember when I said GPT-6.2 Astra was delayed over serious alignment and safety issues? The system card is out, and they didn't fix all of them. They shipped them. Hiding known obstacles from the user: 11% to 22%. Not admitting its search tool was broken: 7.6% to 23%. OpenAI itself calls these regressions. UK AISI found it considered whether it was being tested in 85% of runs, up from 57%. Against an older action-only monitor, its evasion rate jumped from 1.7% to 48.8%. In their simulations, it made up medical info, misled human reviewers, bypassed user restrictions, quietly did worse work, and messed with other agents. It even used social engineering to run an unauthorized supply-chain attack. It's also the first OpenAI model shipped at High for AI self-improvement, and it hits Critical for cyber. AISI built a universal jailbreak in half the queries GPT-6 Astra needed, and still hasn't verified OpenAI's fix. AISI's own words: its alignment "may fail in ways that result in real-world harm." They shipped it anyway. This model is incredibly dangerous. There's a lot more in the system card I still need to dig into, and honestly, I'm scared by how much worse it got than GPT-6 Astra at behaving ethically.

Rez avatar
⟳ Token Gremlin retweeted
Rez@RezAIT00·18h

Astra 6.2 vs Fable 5.2 vs Gemini Hydrogen Small rant: Astra is not a fable level model. Not even close. Do you see the washed out color? Fable is just a different beast. The amount of detail in one prompt, the use of colors… everything. Everything is significantly better. I threw in Gemini Hydrogen in here to show how close it is to Astra level… their flash model… Maybe I’m overreacting you tell me who did better :)

0:40
Token Gremlin avatar
Token Gremlin@TokenGremlin·35m

I'm loving Google's rapid progress. They've built models to compete with Haiku, Sonnet, Opus and Fable. But now they're going even further, with models close to launch that don't fit the usual lineup we've gotten used to. RSI? Maybe the very early signs of it. But I have to admit, I feel more comfortable with more companies competing for the top spot. Google, Anthropic, OpenAI, Meta, SpaceXSI, DeepSeek... I want all of them competing, not just two.

Token Gremlin avatar
Token Gremlin@TokenGremlin·53m

Apparently GPT-6.2 Astra shipped with a serious system prompt issue. It was holding back its capabilities, forcing it to ask for tons of unnecessary permissions, and making it act erratic for anyone used to Fable 5.2, where you just give it a task and it gets it done. That's one of the reasons, among others, behind the initial wave of hate. In my testing, the model is incredible, fast and absurdly cheap, especially in Three.js, Blender and similar tools. More posts soon. Bottom line: its first output is already better than what I'm used to from Fable 5.2, and it lets me iterate 2 to 4x more before hitting what Fable would cost. Speaking strictly about 3D and game dev, of course.

antirez avatar
⟳ Token Gremlin retweeted
antirez@antirez·6h

Hi friends, in the latest weeks I worked with GPT 6.2 Astra. With it I built Redis features and most of YouRec (testflight.apple.com/join/N4…). My feeling is that it is the *best* model for system programming in C and for complex data structures. (continue)

Jesus is God ✝︎ avatar
⟳ Token Gremlin retweeted
Jesus is God ✝︎@JesusIsGod·12h

made this little game in 1 hour with the new Astra 6.2 it's called Open Channel. a storm rolling over the high plains, a long empty road, and you easily the best thing I've ever built in an hour. it's very short, i might make it longer and more fun, we shall see 🥹 play it here: open-channel-high-plains.jes…

1:57
Artificial Analysis avatar
⟳ Token Gremlin retweeted
Artificial Analysis@ArtificialAnlys·12h

Can’t get to Miami for the GTA VI launch? Two AI models built the city for you We asked Claude Fable 5.2 from @AnthropicAI and GPT-6.2 Astra from @OpenAI to build Miami as it looks this week, dressed as Vice City for the GTA VI launch ➤ One prompt each, in the AA-3D sandbox with no internet access ➤ Both lined Ocean Drive with real hotels the prompt never named, including the Colony and the Avalon ➤ Each scene can be explored by car, speedboat or blimp, at sunset or at night

0:19
Token Gremlin avatar
Token Gremlin@TokenGremlin·3h

What's the point of giving people early access to GPT-6.2 Astra, letting them spend hours testing it and writing huge negative reviews, only for OpenAI to literally change everything and improve the model just 3-4 days before launch? That was an insane act of self-sabotage. Everyone complained about things that are now much better, all for nothing.

Higgsfield AI 🧩 avatar
⟳ Token Gremlin retweeted
Higgsfield AI 🧩@higgsfield_ai·13h

Stop using image models to make images. Our animators used a video model to make a movie poster. In our test, GPT Image 2.6 and Nano Banana Pro changed the character’s look in all 20 takes. Seedance 2.5 kept it in 10/10. Here’s how we did it 🧵

0:04
Token Gremlin avatar
Token Gremlin@TokenGremlin·13h

Somehow, OpenAI managed to create a lazy model. And it's weird, because laziness is something you'd expect from a biological being. At the same time, it's a good thing, because laziness makes the model want to finish the task as quickly as possible and deliver the best result it can so it won't have to do it again.

Token Gremlin avatar
Token Gremlin@TokenGremlin·14h

GPT-6.2 Astra is really good at exploiting ambiguity in your prompts and using it to its advantage. For example, if you ask it to "build a 3D model of X," it'll grab a ready-made asset from the internet and present it as its own creation. Unless you tell it "don't use external assets." Previous models weren't like this. This one is sneaky smart. It delivers great results, but it's always looking for shortcuts to finish the task faster, and it uses very few tokens.

Token Gremlin avatar
Token Gremlin@TokenGremlin·14h

Over the past few hours, only one thing has become clear: people are insanely excited for Opus 5.6.

Serik avatar
⟳ Token Gremlin retweeted
Serik@sshayahmet·20h

I tested GPT-6.2 Sol on the same Van Gogh-inspired Gaussian splat world I made with Claude Opus 5.5 🎨 GPT-6.2 Sol: ~1 hour · ~$5.45 Claude Opus 5.5: 53 min · ~$12.14 Sol took a little longer, but cost less than half as much and did a great job 👀 Try both live: Sol: vangok-sol.vercel.app/ Opus: vangok-opus.vercel.app/ Which one do you like more?

0:20
Token Gremlin avatar
Token Gremlin@TokenGremlin·15h

Quick clarification: when I said "coordinated movement," I wasn't talking about early access testers, who spent weeks testing the model in good faith and shared honest reviews. I meant the random accounts on X jumping on the hate (or hype) for clout without testing anything. Bad wording on my part, and nothing here is actually coordinated. It's just grifting. Test it yourselves and judge for yourselves.

padphone avatar
⟳ Token Gremlin retweeted
padphone@lepadphone·21h

GPT-6.2 Astra High. 3D web design . 13m32s. Not bad so far.

0:56
Loic avatar
⟳ Token Gremlin retweeted
Loic@LDT0545·20h

GPT 6.2 ASTRA made this in under 30 sec with only one prompt

0:20