I’m not that deep in the JS ecosystem or runtimes, but I’m a little surprised by some of the claims here about being smaller, having fast starts, sandboxing, performance-competitive, etc.
Does anyone have a sense of what insights, design choices, big bets, etc, unlock all these advantages against already mature and highly optimised JS stacks?
Apple has accused OpenAI and two of its employees of stealing top-secret information in a lawsuit filed in a California federal court on Friday, as the relationship between two of the biggest names in Silicon Valley unravels.
Apple claimed that the ChatGPT maker has used former and current Apple employees to steal hardware designs as the start-up prepares to launch its own AI-focused devices. It alleged that this was part of a pattern of misconduct normalised by OpenAI’s top leadership.
In general I agree with you, especially about how things should work, and find the current trend of claiming LLM-washing code hugely problematic.
I do wonder what the analysis would be of these prompts were vibe-coded, though, since in general LLM output can’t be copyrighted without significant human authorship, and Anthropic is pretty noisy about minimal human supervision.
What about work in units of median annual household disposable income, which are at least somewhat responsive to the distribution of money?
What % do you think a reasonable voter should accept a person donating to a political campaign before it causes concern about the donor's influence vs the median household's voice?
Off the top of my head, I'd guess 500k USD is about 1000% / 10x median annual household disposable income in SE, which I think would give the median voter pause.
For what it's worth (my own view): I think about 10% (~5k USD) is obviously acceptable, and I expect most anyone would agree that donations at that level are fine. I think your proposed 1000% is obviously unacceptable, and I expect most people would agree with me on that as well.
I'm not sure exactly where the level is that opinion would flip, but I feel pretty confident about those boundaries.
For most people, the concern is the money, not the voting. People don't want wealthy people reshaping politics to fit their interests through their wealth. They can vote for whomever they want.
I agree with your common sense take of how it should play out, but Google has and will argue Section 230 protection for AI overviews, eg in Wolf River Electric v Google.
Attacking the cult of progress is a major through-line:
> 12: Today, the human desire for fullness of life is at risk of being misled by deceitful goals, such as the prospect of a technology that promises to free us from all weakness, and models of wellbeing that leave behind entire populations. All too often, we place our hope in unlimited 'upgrades,' in forms of progress that exacerbate inequalities, and in immediate solutions incapable of healing people's wounds.
> 94: The danger of humanity becoming a victim of its own achievements was already clearly recognized by Saint Paul VI, who warned that 'the most extraordinary scientific progress, the most astounding technical feats and the most amazing economic growth, unless accompanied by authentic moral and social progress, will in the long run go against man.' For this reason, technological progress — valuable in itself — requires careful discernment of the anthropological vision that guides it and the ends it pursues. If technological development advances without a corresponding ethical and social progress, the result may be an increase in means without a growth in humanity: 'having more' without 'being more.' In such a scenario, there is a risk that individuals will be evaluated principally according to the outcomes they produce.
> 112: More gravely, the pervasive technocratic paradigm in which we are immersed, and that is amplified by the digital revolution and AI, threatens to normalize an anti-human vision. In that vision, the fullness of life is equated with having more, reducing weakness, eliminating uncertainty and exerting total control. When efficiency becomes the ultimate measure of value, human beings are tempted to see themselves as a project to be optimized rather than as persons called to relationship and communion.
I have a lot of sympathy and affection for this project. When I started working in the US Congress in 2009, I was shocked there wasn't a more concerted effort to put US code in some kind of version control system.
Over time (~15 years in law and public policy in US and EU) I've come to view that kind of project as something worthwhile, and interesting, but not something that will meaningfully change how laws are created and understood.
In the US, as a mixed but substantially common law system, the text of the US Code, which in the LawVM model is the tree, can remain fixed, but the interpretation of that text will still vary over time. (Same syntax, different semantics, which feels like an AI-ism as I write it.)
A couple examples:
I have worked in competition policy. Section 2 of the Sherman Antitrust Act has read the same since 1890 but it's operated across structuralist and consumer-welfare regimes, and the outcomes in Standard Oil, AT&T, MSFT, and more recent cases have all been shaped by those regimes.
US constitutional law may be an even better example of the issue: Article I, sec. 8, cl. 3 (the commerce clause) has not been amended since 1789, but it has expanded and contracted and been reinterpreted over time (most recently with NFIB v Sebelius).
Another example of the problem with a LawVM like system making "the law in force [...] knowable" is US overturning of Chevron deference on 28 June 2024.
Basically, the day before, courts were expected to defer to expert agencies, constrained by the Administrative Procedures Act, in their interpretation of legislation. The day after, courts were expected to make an independent assessment. All sorts of legislative texts were unchanged, but their meaning shifted dramatically: Clean Air Act, Communications Act, etc.
There are a bunch of other issues in figuring out "what the law is", like agency rulemakings in the CFR, and then instruments which sit below those, including sub-regulatory guidance, prosecutorial discretion, etc.
Things which fall into this bucket:
- DOJ pot enforcement
- SEC no-action letters (letters which say in effect "we won't try to enforce if you do X")
Altogether, my thinking of this shifted to believing that the real thing that needs to be modeled is not the US Code as a tree serialization format, but the interpretative algorithm (the things lawyers and courts do) that translates that serialized tree into decisions. That interpretive algorithm has a lot of inputs, one which is the serialized tree, but also a bunch of other stuff.
As a final note, I will add that the ambiguity in legislative text is often a feature, not a bug. I have worked on legislative text that was intentionally crafted to provide different plausible interpretations to different coalition members both in the immediate context and in the long term.
In summary: projects like this are good and admirable; they're likely to have more direct utility in jurisdictions on one extreme of the civil law spectrum (i.e. not the US); and in all jurisdictions there likely needs to be an interpretive mechanism that sits above the representation of legal text which has more inputs if the goal is to model what the law is at any given time.
> That's not an incremental improvement. That's a category change in how Claude Code navigates your code.
I don’t know anything about the human(s) behind this project, assuming there are any, and intend no malice towards them, but when I encounter language like this, it just kills my enthusiasm for a project.
I wonder if that reaction is now, or will become, a majority one, and that AI flavoured language and products will face some audience headwinds if there aren’t indications of some level of human authorship / editing.
The article isn’t describing someone who learned the concept of sortable IDs and then wrote their own implementation.
It describes copying and pasting actual code from one project into a prompt so a language model can reproduce it in another project.
It’s a mechanical transformation of someone else’s copyrighted expression (their code) laundered through a statistical model instead of a human copyist.
> One trick I use constantly: for well-contained features where I’ve seen a good implementation in an open source repo, I’ll share that code as a reference alongside the plan request. If I want to add sortable IDs, I paste the ID generation code from a project that does it well and say “this is how they do sortable IDs, write a plan.md explaining how we can adopt a similar approach.” Claude works dramatically better when it has a concrete reference implementation to work from rather than designing from scratch.
Licensing apparently means nothing.
Ripped off in the training data, ripped off in the prompt.
> (a) all human beings are born with identical talents and inclinations.
> (b) human beings may be born with different talents and inclinations, but these talents and inclinations are distributed identically across all living populations.
Or: (c) The inter-group variation in the talents and inclinations of human beings is completely dominated by the intra-group variation.
In other words, "living populations" (= ethnic groups) don't matter. You'll have (e.g.) smart and dumb people in every group, and everything else is noise.
I'm not sure how a reasonable person could choose hypothesis (b) over (c), given the long history of hypothesis (b) proponents trying (and failing) to make the math work out for them.
Does anyone have a sense of what insights, design choices, big bets, etc, unlock all these advantages against already mature and highly optimised JS stacks?