colinhb

Thanks - so much easier to understand what Elk is (and why it is) than Ant.

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.

My independent judgement and experience is along your lines, but I’m trying to also take Anthropic seriously:

“I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.”

“Can confirm Claude Code is 100% written by Claude Code”

Both from head of Claude Code. Taken either literally or seriously they point to Anthropic shipping code with limited human authorship.

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.

https://www.startribune.com/google-ai-overview-lawsuit-defam...

Previous cases in this space (eg Meta v KGM, Walters v OpenAI) have not turned on Section 230 specifically.

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.

There's much more along these and related lines.

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.

Thanks - appreciate it!

Maybe something fossil-adjacent:

- amber.dev

- quarry.sh

You’ll probably need to play with gTLDs to find something that works.

Can also echo “scm” from fossil’s domain:

- amberscm.dev

Along with useX.com, Xhq.com, etc., patterns.

Of the two you have listed, I’d choose fossilforge, but would vote for an alternative TLD since .io has an expected meaning coming from GitHub.

Tangent - does anyone immediately recognise how this was typeset? I’m guessing it’s some kind of pandoc output?

I read the original chapters online but appreciate this format.

Wild! Fun to see 9p filesystem protocol continue to have a life in this form.

From the intro section:

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.

Quoting the article:

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.

Did not replicate for me w/ Opus 4.6: https://imgur.com/a/4FckOCL

Very much appreciated the internet's #1 dog-rating account covering this:

https://www.instagram.com/reels/DUlye8NETR3/

Skipped the metaverse, slotted between the two

Can it self-drive a Tesla?

Neat.

When working in a linguistics lab as an undergraduate long ago, we looked at spectrograms to identify sounds (specifically places of articulation) as much as listened to recordings.

So it makes some sense to build a model on them rather than some other representation of the sound.

is a polynomial inequality in 26 variables, and the set of prime numbers is identical to the set of positive values taken on by the left-hand side as the variables a, b, …, z range over the nonnegative integers.

I hadn’t heard of this result, and my exposure to Diophantine equations is limited to precisely one seminar from undergrad, but this feels like taking von Neumann’s famous quip to its most fantastical extreme:

With four parameters I can fit an elephant, and with five I can make him wiggle his trunk.

It's difficult to draw meaningful conclusions from country-level differences like this, especially when the countries are as dissimilar as the U.S. (pop. 340M) and Singapore (pop. 4M).

Looking specifically at PISA, it's not usually administered at the state level in the US, but when it has, individual states have outperformed national scores, as one should expect.

For example, Massachusetts has scored similarly to Singapore in Reading and Science (zero or small statistical difference between) and not far in Math.[1] It would be a reasonable hypothesis that a PISA score for the Greater Boston schools (pop. 5M) would even further outperform the U.S.

Sweeping country comparisons tend to amplify noise rather than reveal a clear signal, and are often about regression to the mean as much as anything else. It's not impossible to make sound inferences, but it's difficult to avoid motivated reasoning.

1. https://www.doe.mass.edu/news/news.aspx?id=24050

Ah got it - pinned notes float to the top

FWIW agree with your call on not building too many organisational tools in

Thanks again for sharing; look forward to trying it out

Thanks for the clarification, still haven’t really played with it, but take it these are the mechanics (not the workflow)…

* You have a big chronologically ordered list of notes

* By default, all notes are in view

* You can make a title (instructions in demo), but significance of title is only internal to note (not for ordering/management)

* Big list of notes is union of two disjoint sets: pinned and unpinned

* Can view either all notes, pinned notes, or un-pinned notes

But not sure if I have that right. FWIW appreciate some of the design decisions I’m seeing, just haven’t had time to poke around to understand.

Looks great! One thing I’d suggest (which still isn’t clear to me but interested enough to investigate later): make the note taking workflow clear…

* Is this a bunch of titled markdown docs organized (conceptually) into folders/hierarchically?

* Is this a bunch of untitled/title optional “cards” organized by tag?

* Is this a long, markdown document, which you append to?

These are similar to different mainstream and less so note taking systems, and would appreciate understanding what workflow you’ve designed and optimized around.

Saying you started as a Google Keep user is helpful, buy I’ve only used other systems (Homebrew textfiles, Simplenote, obsidian, etc), and have some concepts around what Evernote and OneNote are like, so giving a couple more signposts on usage would be helpful.

Archive link: https://archive.is/zwJbj

This piece is responding to an Economist op-ed by Bret Taylor and Larry Summers representing the OpenAI board, and comes to many of the same conclusions I did.

- Economist: https://www.economist.com/by-invitation/2024/05/30/openai-bo...

- Archive link for Economist: https://archive.is/rwRju

IMO key paragraphs...

The review’s findings rejected the idea that any kind of ai safety concern necessitated Mr Altman’s replacement. In fact, WilmerHale found that “the prior board’s decision did not arise out of concerns regarding product safety or security, the pace of development, Openai’s finances, or its statements to investors, customers, or business partners.”

Comment: I'm surprised that they don't refute any of the concerns about the CEO, and if the investigation was so redemptive, they should release the findings. (It must be that it wasn't, so they won't.)

Ms Toner has continued to make claims in the press. Although perhaps difficult to remember now, Openai released Chatgpt in November 2022 as a research project to learn more about how useful its models are in conversational settings. It was built on gpt-3.5, an existing ai model which had already been available for more than eight months at the time.

Comment: As someone who had access to OpenAI models prior to the release of ChatGPT, it's disingenuous to say that GPT-3.5 was "available". Yes, available to enrolled researchers willing to suffer through the tools to interact with a model not fine-tuned on conversation.

Archive link (for paywall): https://archive.is/rwRju

Key paragraphs...

The review’s findings rejected the idea that any kind of ai safety concern necessitated Mr Altman’s replacement. In fact, WilmerHale found that “the prior board’s decision did not arise out of concerns regarding product safety or security, the pace of development, Openai’s finances, or its statements to investors, customers, or business partners.”

Comment: I'm surprised that they don't refute any of the concerns about the CEO, and if the investigation was so redemptive, they should release the findings. (It must be that it wasn't, so they won't.)

Ms Toner has continued to make claims in the press. Although perhaps difficult to remember now, Openai released Chatgpt in November 2022 as a research project to learn more about how useful its models are in conversational settings. It was built on gpt-3.5, an existing ai model which had already been available for more than eight months at the time.

Comment: As someone who had access to OpenAI models prior to the release of ChatGPT, it's disingenuous to say that GPT-3.5 was "available". Yes, available to enrolled researchers willing to suffer through the tools used to interact with it. And beyond that, ChatGPT as a product, used by humans, is a world apart from a trained model: https://openai.com/index/chatgpt/

They shouldn't have written this pieces. Makes them look worse.

Agree in part but also I think Terry is such an outlier (though also generous and humble) that it’s hard to extrapolate from this example to a more general case