mmaunder

Seems that “oracle routing” is a term the authors invented. Also sounds like they’re sending requests to all routes but assuming the readers request will route to the “best” model. The closest thing to what they’re describing is semantic routing using a NN search in a vector DB to make a routing decision, but the efficacy of this approach isn’t a slam dunk.

Re local LLMs: “You avoid the uncanny valley where the model appears to be intelligent before doing something stupid; the stupidity is self-evident and this helps calibrate your interactions.”

Making the authors stance on LLMs clear.

Agreed re reasoning. I’ve seen this play out with 5x reasoning negating cost savings.

Historically these have been quickly shut down without much of an explanation.

Or an allegedly even more dangerous model that they refuse to release. What a joke Anthropic has become.

25 gig is still expensive for consumers to configure end to end.

The new normal for the next decade: You must protect the public from us and all others, and we are your closest ally so we make your rules for us and our competition. This wasn’t a lucky outcome. They laid the groundwork years ago with the AI “ethics” movement and this was the play all along.

Not quite vaporware, but few commits, PRs, history, actual examples etc. It's pretty thin.

A Wasm decoder takes encoded bytes and returns an iterator of Arrow Buffers. In case you were wondering.

And poached eggs.

You lost me at forced scrolling. Ugh!

You’re under 50 and haven’t seen the results of a stroke. There are states worse than death.

Fish, tree nuts, avocados and olive oil. Avoid butter, red meat and cheese. It’ll save your life.

Unusable thanks to the top animation pushing the rest of the site down repeatedly as you’re trying to read.

Political content on HN? That’s a first. Please flag this. Let’s keep HN free of rage bait and focused on innovation.

Amazing how the dev community is suffering from a similar inability to approach the subject of real world AI efficiencies and business benefits. I don’t think it’s helpful to accuse the other side of psychosis. It disqualifies any data or experience they bring to the conversation.

Metrics would help others who may want to rescue the project consider the options. Eg user base would make it clear if there’s an immediate opportunity to work with the author to launch a paid backup service around the project, funding continued work on it.

BM stills camera coming soon. It would replicate their video model with their software driving their hardware sales.

This comment deserves it's own HN post. Thanks!

24/7 coverage here: https://www.youtube.com/watch?v=m3kR2KK8TEs

If creating OSS is this low effort, the right question is: What high effort assets, that are valuable to other builders, should open communities be working on? And I think the answer is open source models with open training and open training data.

The generous take is that this is someone's pet project that marketing got too excited about, and that the leadership haven't applied their minds to. GPL provides a moat for the community, who are contributing their time and energy into a project. It ensures that, even if a commercial company grabs your software, extends it, and commercializes it, that you can fold those improvements back into your original distribution. While the commercial entity benefits from your free labor, you benefit back from theirs.

Re-implementing WordPress (their words, not mine) as MIT licensed, while legally questionable, breaks that virtuous cycle and removes the community's moat. They've taken WordPress's roles and menus and borrowed its Gutenberg code (which is GPL), and launched it as an MIT licensed product, which breaks that virtuous cycle. It means e.g. a hosting company can take the product closed source if they want to, and never have to contribute any of what they build on top of the community's work, back to the community.

https://github.com/emdash-cms/emdash/tree/main/packages/core... says "The core EmDash CMS package - an Astro-native, agent-portable reimplementation of WordPress."

Emdash uses WP's RBAC roles. Also uses their menus. Also depends on @wordpress/block-serialization-default-parser which I think might (??) be able to be used by an MIT project even though WordPress is GPL.

They used Claude Code it seems because the first commit has a CLAUDE.md file which became an AGENTS.md.

Who's we, and how do you know this?

In the source there is an outbound-only Remote Control session that can forward recent transcript history and ongoing user/assistant/local-command events to a claude.ai session, likely for cross-device/session sync, remote viewing, internal dogfooding, or telemetry/ops experiments. It’s separate from the normal explicit /remote-control flow. But in the actual production binary I checked, the mirror helpers are compiled down to hard false, so it does not appear enabled in the shipped distribution build.

Same story for the anti_distillation: ['fake_tools'] path: I could find it in source, but the prod binary I checked does not contain the anti_distillation / fake_tools strings at all.

Come on guys. Yet another article distilling the HN discussion in the original post, in the same order the comments appear in that discussion? Here's another since y'all love this stuff: https://venturebeat.com/technology/claude-codes-source-code-...

Claude Code has a server-side anti-distillation opt-in called fake_tools, but the local code does not show the actual mechanism.

The client sometimes sends anti_distillation: ['fake_tools'] in the request body at services/api/claude.ts:301

The client still sends its normal real tools: allTools at services/api/claude.ts:1711

If the model emits a tool name the client does not actually have, the client turns that into No such tool available errors at services/tools/StreamingToolExecutor.ts:77 and services/tools/toolExecution.ts:369

If Anthropic were literally appending extra normal tool definitions to the live tool set, and Claude used them, that would be user-visible breakage.

That leaves a few more plausible possibilities:

Fake_tools is just the name of the server-side experiment, but the implementation is subtler than “append fake tools to the real tool list.”

or

The server may inject tool-looking text into hidden prompt context, with separate hidden instructions not to call it.

or

The server may use decoys only in an internal representation that is useful for poisoning traces/training data but not exposed as real executable tools.

Great theory. I'll dig deeper.

The only sensible response is to immediately open source it.

Agreed. This is a big deal.

Bet you’re pissed.