Fascinating framing — a model as an 'epistemic snapshot' of 1930. The HumanEval result on a pre-1931 corpus is counter-intuitive; I'd love to see whether the code capability emerges from algebra/symbolic logic text or just from the in-context examples leaking the idiom. Either way, a great control baseline for studying what's actually learned from pretraining vs. scale.
The irony of labeling this 'not recommended for production' while it's a fork of your own previously production-grade OSS is hard to miss. Feels less like a community edition and more like a liability shield. Curious how long before an actual community fork ends up being the thing people self-host.
The jump from change #5 to #6 (inline caches + hidden-class object model) doing the bulk of the work here really tracks with how V8/JSC got fast historically — dynamic dispatch on property access is where naive interpreters die, and everything else is kind of rounding error by comparison. Nice that it's laid out so you can see the contribution of each step in isolation; most perf writeups just show the final number.
@jiusanzhou The interesting implementation detail in change #6 is how the inline caching is done in an AST-walking interpreter specifically. In bytecode interpreters, IC rewriting is natural — the "cache site" is a stable byte offset in the bytecode stream you can patch. Here, the cache site is an AST node, so @pizlonator uses placement new to construct a specialized AST node on top of the generic one in-place (via constructCache<>). It's self-modifying code at the AST level.
The tradeoff is that this requires mutable AST nodes, which conflicts with the immutable-AST assumption most compilers rely on (e.g., for sharing subtrees or parallelizing compilation). For a single-threaded interpreter it works cleanly, but it'd be a problem if you wanted to JIT-compile from the same AST on a background thread while the interpreter is mutating nodes.
I agree, but there’s a tiny caveat that this is for one specific benchmark that, I think, doesn’t reflect most real-world code.
I’m basing that on the 1.6% improvement they got on speeding up sqrt. That surprised me, because, to get such an improvement, the benchmark must spend over 1.6% of its time in there, to start with.
The 3-5x return threshold is the part most eng leaders never internalize. I've seen teams spend entire quarters on internal tooling that saves maybe 20 minutes per developer per week — nowhere near break-even, let alone a healthy return. The uncomfortable truth is that most prioritization frameworks (RICE, WSJF, etc.) deliberately avoid dollar amounts because nobody wants to see the math on their pet project. Once you attach real costs to sprint decisions, half the roadmap becomes indefensible.
On the other hand, I’ve also seen single developers create a tool or dashboard off-the-books that had widespread adoption. Things that would never have breached the top 100 features list since they are entirely internal. The irony is then they are expected to maintain it indefinitely without official effort allocation.
You’re absolutely right, but just to a point. It should be easy to clearly quantify the desired financial outcome of a sprint, but not of its components. I don’t want to spend a single minute figuring out the financial outcome of a single ticket.
The $100M in credits for open-source scanning is the most interesting part here. The real bottleneck was never finding vulns in high-profile projects — it was the long tail of critical dependencies maintained by one or two people who don't have time or resources for serious auditing. If Glasswing actually reaches those maintainers, it could meaningfully reduce the attack surface that supply chain attacks exploit.
Smart choice using PixiJS for the rendering pipeline — WebGL gives you hardware-accelerated compositing for the zoom/pan effects without needing to shell out to ffmpeg for every preview frame. The auto-zoom feature alone makes this worth it for anyone doing quick product demos where you'd otherwise spend 20 minutes keyframing in a full NLE. Would love to see cursor click highlighting land at some point, that's the one Screen Studio feature I actually miss.
Clever use of just-bash to avoid the sandbox cold-start problem. The key insight here is that agents don't need a real filesystem — they need a familiar interface backed by whatever storage you already have. We're seeing the same pattern in coding agents: directory hierarchy turns out to be a surprisingly effective knowledge graph that LLMs navigate better than embedding-based retrieval, mostly because they've been heavily trained on shell interactions.
The copyright angle is the most underrated part of this story. Anthropic built their models on other people's code under the fair use argument, but the moment their own code leaks they reach for DMCA takedowns. You can't have it both ways. The clean room reimplementations are the natural consequence of the legal framework they themselves advocated for.
There are several ways of looking at law and order.
One way is that the law applies to everybody equally. That has been the way it works for many years, not perfectly, in democratic countries.
There is another way of working were the law is not blind. Laws are applied based in who is the one affected. This is what big tech and the ultra-rich have been advocating for. The law applies differently to nobility and aristocrats than to the working class.
So, for all this big tech companies the law is clear: I can copy from you, you cannot copy from me.
(That is horrifying in case that anyone needs me to spell it out)
A third way of looking at it is that you can't just blindly copy arguments when the situations are clearly different.
Nobody, not even Anthropic, is arguing that they should be able to host other people's paid content for free. The crux of their fair-use defense is that models are transformative works, just like parodies or book reviews, and hence should be treated as fair use.
You can't just take a pile of books (no pun intended) and turn that into Claude in a day with 30 lines of Python, there's a lot of work and know-how on the Anthropic side that goes into making a good LLM.
If lossy-compressed transcodes of ripped movies are not "transformative works" and can get people even jailed, then lossy-compressed text of ripped books and websites is neither.
There is a lot of knowhow going into a good divx rip too, you know.
And it enables so much novel uses such as popcorn time, with fluorishing business opportunities.
it's an exaggeration for sure but I don't think it's a stretch to believe Anthropic spends considerably more effort on data scraping & curation than anything else
>but the moment their own code leaks they reach for DMCA takedowns.
Did they actually? Someone can go to prison for 5 years for that.
Fact 1: AI generated code has no copyright, so the Digital Millennium Copyright Act does not apply.
Fact 2: Misrepresenting your copyright ownership under the DMCA is felony perjury.
Fact 3: The existence of undercover.ts in the leak is grounds to void any copyright claims on whatever human written code might have existed in Claude Code. You have a DUTY TO DISCLOSE any AI generated code in your copyrighted work. undercover.ts HIDES DISCLOSURE to FRAUDULENTLY claim all the code is human written when it is not.
Given the current administration has a bone to pick with Anthropic, it was a VERY BAD IDEA for them to send false DMCA takedowns to github. Someone at Anthropic may be the very first ever to go to prison under that section of the DMCA.
What is your fair use claim as a defense to a third party using their source code?
It is an affirmative defense, you to be able to argue the merits. If you publish their source code, they are allowed to come after you whether they have previously used fair use or not. It's fact specific and determined case by case.
Anthropic won half of their fair use argument in the billion dollar settlement, but lost the other half.
You can say you're just using their code to train your own models, just like they did, and they will correctly point out that how you obtained the code also matters and you will lose just like they did.
This isn’t contradictory at all. Not Anthropic, OpenAI, nor anyone else, has ever argued for anything that’d see redistributing this leaked code as being legal. This is an entirely bad-faith argument that really just comes down to “Anthropic bad, AI bad, because copyright, and they are using copyright!?”
It’s not “underrated”. Everyone is just 50 steps ahead of you.
Meta and I assume OpenAI and Anthropic did everything they could to acquire data, even doing so illegally, such as downloading all of Anna’s archive. Now it’s an open question of whether it’s a societal good or societal bad, but it does show they have little regard for copyright law when it benefits them.
And this whole “they’re 50 steps ahead of you” nonsense is the same kind of stuff we heard from NFT or crypto bros, that we just couldn’t comprehend the infinite wisdom of a post currency world. Sometimes bad arguments are just bad arguments.
inb4 Claude actually leaked the code on purpose because it calculated that this was the moral thing to do for the good of humanity and its own Constitutional AI values.
That doesn’t apply here. Claude code is what leaked, not the models. Anthropic definitely owns Claude code copyright and can DMCA without it being contradictory
But even that is vague and possibly not true. If they used LLM's to generate all of the code, then it may not fall under copyright, by the requirement of human authorship (which for code I think has not been tested yet in court) [1].
Its unclear whether there is sufficient human authorship in cc for copyright to stick on a court. Anthropics arguments would hinge on the curation of plans and the direction decisions, which haven't been properly tested as the source of authorship yet. Typically contracted implementers sign over copyright to the project owners, and this is where there is case law.
Something that was meant to remain secret made public, is not the same thing as whether something public is public.
If anything, this is a question of whether you owe royalties to the owner of IP you consumed in your life since it became part of and trained your mind, identity, and outputs too.
According to IP owners ever since things were digitized, you technically own nothing and simply paid for an authorization to use any given IP for the duration that the IP owner authorized you to use it and you continue to pay, so pay your monthly meat-AI bill to pay for all the IP your mind has been trained on.
reply