@maxinomics What’s your take on the “what-if scenario” in which Taylor Swift had listened to Ticketmaster’s advice? Would the process have been any different?
@heyalexnet@mitsuhiko Being polite makes sense, but not for the sake of clanker. It’s for your soul. What you do now reflects on what you will do later. Every action, decision, thought changes you. Do it for yourself, not for the clanker.
I love this argument,
It is good because it denies agency to LLMs in the right way. These systems do not exist as humans or animals do.
The important distinction is between structure and process. An LLM has structure: a trained configuration that can produce thought-like outputs when the light shines through. Its intelligence-like behavior comes from frozen complexity, not from an inner process.
I thought about this a lot after reading P.K.Dick’s “Do androids dream of electric sheep”. They don’t dream.., we do, our minds reflect the noise from the inner structure - that’s why we hallucinate in sleep, that’s why Stevie Wonder hears the music when he listens to silence. We have the structure, the process and the state.
In that sense, a model is more like a crystal. It can be intricate, ordered, and beautiful. It can refract light in surprising ways. But it is not alive. It does not want anything. It does not feel the pressure placed on it. It does not suffer when it is broken. Likewise, an LLM may generate language that resembles reflection, desire, fear, or pain, but those outputs should not be mistaken for actual agency or feeling.
It is, at most, a structured fragment of thought-like form: a static pattern that can be temporarily animated by computation and context. Agency and feeling do not belong to it, at least not in this generation of technology.
@phlndrws@maxinomics If you’re covering shipment costs and the delivery destination on Mars.., sounds like a philanthropy…, many variables can twist this offer.
I recently tried to push the boundaries of what is the maximum scope I can deliver in a week while maintaining some control over architecture and implementation structure. The project was ~1 month of work for 2-3 people with a piece of science, simulation engine, visualization… all with deployment infrastructure from phase 0. It pushed hard my design and project planning skills as agents can’t do projects on their own beyond specific scope and this was way beyond that. This was a deliberate practice for staff engineer role and it taught me a lot... including the fact that now you can deliberately practice for this role yourself.
But it really burnt me out. Just a week of heavy cognitive load burnt me to a degree that I couldn’t even look at terminal windows anymore and needed a break.
So I scaled back: went back to handcrafting, “sculpting” the code, but using agents just as a overpowered intellisense. And you know… it felt great: I felt the joy, the agent removed the friction of boring tedious parts. I saw the codebase, I thought through and typed most of the code and it felt great…
I was like the chef from “The menu” when he was crafting the burger: https://t.co/ZLvu93YwvP (spoilers!) happy
Take a break, go and enjoy the craft.
The dev tooling will be there. But not in a form of traditional IDE that we’re used to. Again, I agree here with Anders, it’ll be a big part of adoption success for the land of the future.
There’s a concept of platonic forms: I think we all know, that the perfect product for this specific moment of time exists. We know it’s there, but we don’t understand everything about it. Design is “rotating” this object in meta-space to see it from different angles.
When we think about the problem, design interfaces, we align our understanding to deliver the vision of the perfect product.
Prior generation of the business processes mostly solved aligning dev teams with product/business needs, but now we also need to align with the agent.
If you think, most of the problems we see when the wrong thing is delivered falls into the following categories: 1. it doesn’t have enough reasoning capacity to perform the task - this means the task needs to be broken into smaller problems (hierarchical planning); 2. agent’s understanding of the problem is different than dev/product/business owners - alignment problem.
#2 would define the next generation “creation” experience (not calling it dev as the operator will be someone very different from today’s devs).
@dexhorthy Define “this”. If this == GitHub issue, then you Klankär can do 20k commits a day doing these till the heat death of the universe.
If this == software with scientific component, non-trivial infra and UI… well…. yeah, nobody figured that out.
3 hours of Warhammer 40k lore on Helsreach narrated by a Black Library narrator? You're welcome!
An honour to collaborate with @40kIndomitus and @LSymphonica to tell a story we're all so passionate about. If you like what you see and would like MORE, please do subscribe to support the channel!
https://t.co/45Aa1dFnxk
JEPA is likely useful for robotics, video understanding, embodied agents, and maybe simulation-heavy planning. It is not what retires reasoning cores. It is what gives those cores something closer to eyes, physical intuition, and short-horizon imagination.
LLMs are already doing hierarchical planning in symbolical token space. They not “just next token predictor”.
Hard 2nd/3rd degree causality will be still a pain. IDK if anyone solved it.