Most programming best practices were created to solve for the same problem LLMs face: limited context. The principles around well-crafted code strove to make it modifiable by future agents who are not the author. Maintainable code stabilizes around the invariants of the domain it address while providing a spectrum of modifiability to future users, running from UI > configs > cleanly abstracted code > core.
I do wonder if there will be a convergence between sodium-ion battery architectures and cheaper, renewable-powered desalination. Could industrial seawater mining be competitive as a sodium feedstock source?
I think the limitation to team "let-people-do-what-they-want-even-if-it's-bad-for-them" is that it's rooted in outdated assumptions about what free will and coercion look like. Broadly speaking, libertarian notions of negative freedom assume a model of coercion with a very small footprint, primarily the threat of immanent physical violence. Similarly, it sees the exercise of free will as a binary action, rather than a spectrum of ability.
We entered this era a long time ago, where psychological insight and computational or chemical power can combine to override people's ability to make free choices. But I'm still not sure there is a widely shared philosophical or ideological framework that has fully digested and responded to these new realities yet.
I've been running Claude Pro at home, supplemented with Deepseek configured in Claude Code. I've had decent luck throwing Opus (and briefly Fable, RIP) at architecture / product problems and producing plans to hand off to Deepseek (I personally find v4 to operate somewhere between Opus and Sonnet in capability).
Lately I've been able to cut down on token usage with context-mode and codebase-memory to wring more out of my subscription, as well as doing things like make sure all terminal operations run in quiet mode. I've found codebase-memory particularly effective: it creates an index of your codebase that the agent can query for code tracing without reading all of the associated files, and I've also found it more accurate at analysis
This is the answer for me as well. My partner and I try to be intentional about giving each other at least an evening a week to go do something social, but it doesn't work out nearly that often because everyone else is juggling their own kids' schedules. So even if I didn't personally have kids, I would still be facing the same issue.
Yeah, I like the "get out of jail free" card approach. The thing I always used to hate before this era was getting stuck in a hole on something that would take days or worse to grind through. It's nice to drop a little plank bridge across those now
I have been building an iOS app that I had kicking around in my head for years but never had time to build. I have been a frontend UX engineer for the better part of a decade and went through a handful of tutorials on Swift. The project definitely sits in this uncanny valley for me. I have test suites for every aspect of the app and have the agent using TDD to avoid cheating - this has gotten me pretty far without having to look too close at the output other than general structure. As I'm reaching a more mature stage of the project though, I'm finding that I want to tweak a lot by hand in the code to get the details right without burning tokens.
This is appalling and I agree the technology is creepy. However, human verification is already a big problem that seems like it will only grow from here.
It does seem to me that this should be solvable at the device level by having a biometric scan produce a signed key on your device that can be used to issue a token of authenticity, similar to the way payment systems or certificate authorities work.
Then again, this only intensifies a different, growing problem where access to a smartphone or computer becomes a basic requirement for participation in society. No easy answers.
"But the economy" is an out-of-date framing. The cost of renewables has been plummeting for well over a decade. New renewables are now cheaper than new fossil fuel plants in most of the world, and in many regions they're already competitive with or cheaper than simply running existing fossil fuel infrastructure. As modern wars in Ukraine and now Iran are increasingly demonstrating, they are not only cost effective but rapidly a matter of energy sovereignty and national security.
That's not to say we won't need treaties and supranational entities for some aspects of decarbonization. Methane emissions outside of agriculture are notably a problem of enforcement.
We're badly in need of a collective update to our priors regarding renewables. In the US, a hostile policy toward renewables is not only shooting ourselves in the foot environmentally, we are now actively impoverishing ourselves due to entrenched economic interests across the fossil fuel industry and the cultural inertia they actively worked to develop.
Maybe this is a naive question, but why wouldn't there be market for this even for frontier models? If Anthropic wanted to burn Opus 4.6 into a chip, wouldn't there theoretically be a price point where this would lower inference costs for them?
The demo was so fast it highlighted a UX component of LLMs I hadn’t considered before: there’s such a thing as too fast, at least in the chatbot context. The demo answered with a page of text so fast I had to scroll up every time to see where it started. It completely broke the illusion of conversation where I can usually interrupt if we’re headed in the wrong direction. At least in some contexts, it may become useful to artificially slow down the delivery of output or somehow tune it to the reader’s speed based on how quickly they reply. TTS probably does this naturally, but for text based interactions, still a thing to think about.
I tend to agree, this has been my experience with LLM-powered coding, especially more recently with the advent of new harnesses around context management and planning. I’ve been building software for over ten years so I feel comfortable looking under the hood, but it’s been less of that lately and more talking with users and trying to understand and effectively shape the experience, which I guess means I’m being pushed toward product work.
I spent some very enjoyable time browsing courses and tutorials in the Santa Fe Institute’s complexity explorer![1]
I wish I had encountered complexity science earlier in life. It touches on so many of the questions that have sparked my imagination over the years, I’m so pleased to find such an accessible introduction.
The ending of the article left me feeling he had more of an axe to grind here. The mostly unspoken ideological background is that classical art is often appropriated by proponents of Western chauvinism to demonstrate their supposed innate cultural superiority. Poorly painted reconstructions undermine that image, but it does not mean this was done intentionally. I agree that a more neutral observer would have been interested in learning the thought process of those researchers.
Interesting. Like many people here, I've thought a great deal about what it means for LLMs to be trained on the whole available corpus of written text, but real world conversation is a kind of dark matter of language as far as LLMs are concerned, isn't it? I imagine there is plenty of transcription in training data, but the total amount of language use in real conversational surely far exceeds any available written output and is qualitatively different in character.
This also makes me curious to what degree this phenomenon manifests when interacting with LLMs in languages other than English? Which languages have less tendency toward sycophantic confidence? More? Or does it exist at a layer abstracted from the particular language?