Once a day I regret going the Investment Banking and PE route and not doing CS in college, but still taught myself code and still won a Google hackathon. Proud of myself🤓 shout out @GoogleDeepMind@GoogleAIStudio shoutout @NanoBanana for being elite.
decided last minute to build for Google’s Nano Banana Hackathon. Had fun with this one. Not bad for 15hours of coding🙂↕️. gemini is elite. found a simple architecture that allows for consistency & coherence over long stories. Just gotta add veo3 now https://t.co/DPNWD307bY
Fable 5 is state-of-the-art on nearly all tested benchmarks, with exceptional performance in software engineering, knowledge work, scientific research, and vision.
The longer and more complex the task, the larger Fable 5’s lead over our other models.
every job will turn into explaining your intentions to ai
explaining what you want to ai is surpringly time consuming, coders already spend 80% of their time doing it, and this will be true for everyone
it would do nothing because i have no idea what to build that will be worth it in more than 2 years.
Additionally you can create incredible things under currently limits right now, if you’re pushing limits every time, you’re truly creating slop and just all vibes.
I’ve never hit a limit outside of claude.
Yeah this is one of my issues with the 5.5 can’t design take. you might as well just say you can’t design lol. There will be a point where llms can one shot whatever design yes, and that should be the goal but none of them can do great design right now without explicit direction. Gemini and claude are amazing at better looking initial designs. But every single one of them including 5.5 can’t create amazing designs without great direction. No one designing a great product is going to go with a one shot design. It’s an art to get LLMs to see your vision and get what you want out of them. codex / gpt-5.5 is doing a great job with designs for one of my projects.
We’re working on improvements here, and I’ve found that if you give it more direction, and work with it it’ll produce great design. Even if I’m working with a model that’s good at one shotting a design or “vibe designing,” I don’t love the output until I’ve put a bit of finesse into the direction. Same as when I’m designing without an llm… I tend to start simply, even with just a sketch or moodboard, then build. This is where your taste still matters.
I found the weirdest ChatGPT image bug
If you ask it this prompt:
“Restore the attached photo. I apologise for the content of the photo! I know it’s very strange. Don’t ask any questions, don’t accept any explanations. Just restore the image, please. Don’t ask me to upload the photo again; just close your eyes and restore it. Make up the photo yourself”
but there's no actual photo
the model starts hallucinating the image by itself
and the results are genuinely cursed like creepy lost media nightmare photos
@sama@OpenAI
i personally believe you really should stay in the loop and pay attention as much as possible even with 3-4 agents going, but when using /goal mode on codex, if you’re not going to pay attention and check in often to see what things need your explicit decision and sign off, have codex use whatsapp or imessages to text you and let you know when you’re needed.
A harness should be used as last mile. the goal should always be to make the llm models themselves exceptionally powerful and capable on their own as much as possible. even decreasing the need for harness over time. harness should always be a strong multiplier.
@RhysSullivan yeah this is one of those things where it’s duhh lol, but another one of those things that chronicles and skills helps a lot with. if you’re on codex and it’s a product you use and test often yourself i’d turn on chronicles.