@s8mb 100% all of this. Create IFSC’s for AI across EU and mid powers - special rules apply in these zones
Companies in mid power countries will be fearful of US models - create the env in these zones where labs have diff reg env and corps can purchase with confidence from them
Hey @bcherny - can we have a unified way to access skills across Chat / Code / Cowork. Skills that live in folders are v useful. If we could add a folder to Claude Chat (like cowork) + make coworker better at picking up skills I think that would unify it?
@solo1y Because your input does not update the model weights … yet. Likely will by year end - then no more internet slop. Cake is a trivial example - long term health may be a batter example - the models will extrapolate not just on one convo, but entire history and learn from mistakes
@solo1y Probable domain = alpha go. The models can learn without any human data. They experiment and find answers humans did not.
If you tell AI your cake preferences, happy to bet it will suggest a cake you never knew you liked before the year is out
@solo1y But this year, there will be a feedback loop from you - you will tell it which cake you liked / hated, and it will learn your prefs. This is stylised, but the general idea holds and I think applies to art also. Your art preferences become the training data, not internet slop
@solo1y Yea … but no. For maths / engineering etc models are now trained on synthetic data - easy in provable domains. For things like art, this year you will see self improving AI’s - ref David Silver - ask a model for you fav cake recipe you will get the recipe liked by most …
claude code + opus 4.5 injected the immaculate hacker vibes back into ai that we haven’t had since gpt-4.
everything is new + fun + weird again.
you can feel it.
another oom of new ideas & latent economic value is waiting to be unlocked.
and building has never been this fun.
LLMs represent the "library" phase of AI. The next phase will be the "scientist" phase. A library contains answers, but a scientist knows how to find answers that don't exist yet.
@DeryaTR_@breedenlaw The framework you need is free, Clause Code, the model you need is Opus 4.5. Put the unfilled forms in one folder, case files in a second folder. Run Claude code so it can see both. Ask questions of you case, ask Claude code to generate your forms. This will work, ~$5 / form
@jeremyphoward Totally agree. Daily cause of frustration. KISS. There will be a class of company that fails / wastes tones of money through over complexing their AI stack.
This 💯! Play Wolfenstein 3d in a Blackwell GPU and be like - computing has hit a wall. Hard to demo - this thing will work reliably for 20 mins / understand 3k lines of code and not be confused - in a way that is engaging. Agentic framework is becoming important!
We've entered a new phase where progress in chatbots is starting to top out but progress in automating AI research is steadily improving. It's a mistake the confuse the two.
This is exactly my experience. GPT5 is an undeniably great model for developing more than trivial apps. The agentic frameworks to achieve this are now becoming far more important
Seeing a lot of conflicting takes about gpt5.
If you have bad ideas, taste, or bad comms, the model is too autistic to meaningfully help you out.
For those that think architecturally, gpt5 is a game changer. If you spend some time using your brain, coming up with a spec, articulate it in full to the model, it will quietly execute without deviation or handholding.
Tldr: Gpt5 is effective for complex tasks and opinionated engineers.
@deredleritt3r@inductionheads Your right tho’. YouTube, massive cash pile, massive revenue stream, massive data stream (YouTube, gmail, years of web data), years of experience post AlphaZero, it’s obvs google will likely be first to ASI …. Did I just fall for the ragebate? 😭