Gradient Sky @gradient_sky

@simonw Has anyone tried to frame other bad bad bad forbidden things as a code and then break the guardrails by asking to fix the code?

It could be worse. Some parts could be classified as top secret. Here’s the story from a book about skunkworks, when they worked on early prototype of F-117 “Because of the picture of the airplane’s nose, security classified the mugs as top secret.” The case was that Keith Beswick’s flight-test crew had coffee mugs made with a logo showing part of Have Blue. Since the nose shape revealed classified stealth geometry, the mugs themselves had to be locked in a safe between coffee breaks. Whole they are not locked up, they generally must be under continuous control by cleared, authorized people.

@__tinygrad__ Where’s Apple? They always own s large chunk of the most advanced node

@nuvolore @jullerino Pro tip: ask to use haiku and sonnet for sub-agents

@nuvolore @jullerino I burned full max x20 5h limit in 15 minutes on max… it spawned 300 deep research agents

@dexhorthy 1.5: people that have some idea what they want, but because they never aligned with the agent, something else got built in the loops

@thdxr Study Japan

@maxinomics Only 7% of the company valuation is space and communication. 93% is based on a AI promise that is not built yet and IDK if there’s a clear path to resolve chips bottleneck to get there (orbital inference).

@thorstenball Dostoevsky - Notes from underground?

@mitsuhiko Turn 90 degrees, take a picture… then google internet who is a creator of Pi…

@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?

@housecor It’s a diff viewer, but not in the shape of what we’ve been using last few decades

@GergelyOrosz @Spotify why?

@tom_13z @mitsuhiko It’s not simple and not random, let’s not deny the presence of the structure. But yes, it doesn’t live, it does not exists in time.

@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.

@badlogicgames Serious business… more fashion, more Balenciaga

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: youtu.be/Pls7lt9Fz-E (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).