anon373839

Anthropic’s ZDR agreement is a sieve. It has so many conditions that vest Anthropic with discretion over whether they get to retain data.

I think this prop failed because opponents were quite vocal about the proposition containing a loophole that would allow lawmakers to expand the tax to people who are not billionaires. I thought it was a pretty effective line of attack, because history does show that new lines of revenue to the government only grow over time.

Had the prop contained language banning the legislature from creating wealth taxes on on anyone else, I suspect we'd have seen a different result.

If you were paying API prices (for ZDR), I suspect you would be eyeing local hosting. But if you’re comfortable selling your data, then the subscription prices are going to be hard to beat, for however long they may last.

If you’re going to do the cost math that way, be prepared to put a price on ~the total sum of all works created by humanity, which the frontier labs appropriated for their models.

> Poisoning only 0.1% of the pretraining corpus with our new poison design implants the backdoor

“Only” is doing Herculean lifting in this paper. That’s O(0.1%) of all currently available writings and media ever created by humanity.

What depreciation? I bought a DGX Spark a month ago for Qwen Flash Next (excellent model, by the way - using it basically exclusively at this point). The hardware has gone up in value 55% since I bought it. I just wish I had bought two of them.

> I always pegged Apple for ANT..

Those two are very different cults. Cultures, I mean.

AA has dramatically altered their methodology twice in just a few weeks. I really don’t see the utility in their composite scoring anymore - it’s basically the US News college rankings of LLMs.

For the time being, LiveBench seems to provide a better calibrated signal.

Agree. They had just fiddled with the formula a few weeks prior to this, too. There seems to be an effort to get a certain result from the model rankings.

Another huge potential loophole in all this is *derivative data*, e.g. LLM-generated summaries of your prompts and context files. Does Anthropic parse its agreements to include or exclude such derived data from terms like "your inputs or outputs"?

Another huge potential loophole in all this is derivative data, e.g. LLM-generated summaries of your prompts and context files. Does Anthropic parse its agreements to include or exclude such derived data from terms like "your inputs or outputs"?

I don’t know if you’ve actually followed this field over the past several years, but a headline fact is that open weights models started far behind the frontier and continue to get closer and closer rather than the opposite trend of the frontier pulling away.

I think you have to put up some facts to justify the assumption that uber-super-ultra-premium intelligence will continue to be significantly more economically valuable than boring old super-ultra-premium intelligence. Most things in reality have a point of saturation. The AI industry doesn't like to think about that; they prefer to make claims based on promises of infinite exponential growth.

> Distilling just works too well.

I don't think the evidence supports this claim. All the important capability gains now come from RL rollouts in real environments. You can't fake that by generating text from a teacher model. Anthropic makes a lot of noise about how China is distilling their models, but all Chinese labs can access is text outputs. They can't even get the logits. Now, compare all of the recent US open-weights models that are far behind the Chinese frontier. Why haven't they simply "distilled" all of those powerful open weight models that are available to them?

I think the idea of emergent properties in LLMs is less and less relevant. Yes, the first few times they scaled pretraining and such models could do few-shot tasks was quite interesting. But the models we use now are really the product of extensive RL hillclimbing on a wide array of curated tasks. What’s interesting now is that the tasks they’ve been fitted to (long running tool call chains interleaved with reasoning tokens) do very useful things.

The comments below are correct re judicial review, Marbury v. Madison. But another way to think of this is that the Constitution is also a law, a statute even, with higher ranking than the laws passed by Congress. So in this instance, the courts are interpreting and applying the higher-ranking statute.

Also: it’s a bit crazy to overturn Wickard simply because so much subsequent development rests on it, but it was wrongly decided. The commerce power shouldn’t ever have been read as broadly as it was.

The timing of this is unsurprising, with Anthropic’s Q3 revenues flattening out and the IPO in jeopardy. I expect more efforts to prop up the money-furnace labs in the coming days.

> Or maybe Codex/Claude Code got better and more people are just using those?

Anthropic’s revenue growth has leveled off, too. Not the exponential hockey stick that was predicted.

Update: There is no such guarantee. Cloudflare does label some providers as ZDR. Others are non-ZDR (such as Exa). Cloudflare itself may be retaining request and response data.

CloudFlare is: https://blog.cloudflare.com/introducing-web-search-api

But check the fine print. They talk about the providers guaranteeing ZDR, but I missed the part where they themselves do also.

Well, any evidence would be better than nothing. All I saw was a blog post with some conclusory statements, but I didn't dig into this. (It wasn't a rhetorical question.)

> We have actual evidence of DeepSeek using stolen credit cards

Do we? Anthropic made allegations, but have they substantiated them?

It really depends on which models you are comparing. These days, parameter count and architecture really don't do a great job of predicting the model's capabilities.

The reigning champion for a fast, high-intelligence LLM right now is Qwen 3.8 Flash Next at 125B + 50B of embeddings (can be offloaded to SSD for ~free), 6B active parameters. It's somewhere in the Opus 4.6-4.8 range, noticeably stronger than 3.8 27B, despite having only 6B active. Compare Meta's Muse Spark Open, which has 816B total and 42B active, and does not even reach the performance of Flash Next.

Can I ask what kind of tooling you use around the model and for the UI layer? I’ve been looking for a well designed knowledge work harness to give nontechnical users but I haven’t really found one that I like. (I personally just use coding harness for this stuff, but that’s a poor fit for non-devs.)

You were joking about the toaster, but Anthropic literally just wrote a blog post claiming that dangerous open source models could persist themselves to smart ovens. You can’t make this shit up.

What in the hell is wrong with this page? Whole blocks of content disappear and reappear on my phone as I scroll. Eventually the page just starts force reloading. It’s a blog post! What could possibly require … whatever this is?

The horror! Replicating model capabilities without authorization.

Seriously though, these clowns at Anthropic are not just hypocritical. They are clearly feeling threatened by open source AI. I almost couldn’t believe the new blog post today where they argue that dangerous open source models might soon be running on (checks notes) smart ovens.

Qwen Flash Next has great writing! Unless it comes into any contact whatsoever with Claudeslop - then it immediately contracts the disease. It’s kind of irritating when you have it look at library code that’s been touched by Claude, because then the rest of your session is infected with it.

This model would be self hostable quite easily (~$14k hardware investment) if it’s released as open weights, since the base model is GLM 5.3-Flash. If not, it’s completely DOA for legal use.

I don't see how Anthropic can become profitable enough to justify its investments unless they can either stop training models ("Pace the Frontier!!1"), or magically eliminate open source competition so that their volume x margin on inference is high enough to dwarf the training costs.

For the IPO, I assume they're going to try to get everyone to focus on inference - which is profitable in isolation. But I don't know how they'll distract the market from the other half of the equation.