There is - in the Claude Code harness there's a dedicated Plan Mode that doesn't edit any files, or a mode where you can manually approve each edit.
If you're running Claude via -p automated mode, you can set all kinds of permissions and access to each individual editing tool, much like you would with firewall software.
It doesn't edit the system prompt, it blocks access to the tools itself. Just make sure you haven't given the AI the ability to edit its own configuration files, or you'll potentially have an OpenAI v Hugging Face situation.
That sounds entirely plausible. When I was on the Claude Max $100 plan, I would often get the equivalent of at least $1300 API usage, according to the costs counter in Claude Code. That would be about $1 of usage for every $0.075c.
At $1 of usage for $0.13, the reseller is making a tidy profit on top of whatever subscriptions they're reselling.
Most of that is paywalled, but this one paragraph in the Bloomberg article suggests it might be more to do with investors leaking information:
"The suspension stemmed in part from Liang’s frustration over online reports about his comments to investors during his first financing deal"
The part of the transcript I'd seen floating around online was this part from around 1 hour 26 min:
"With the largest models available today, we simply cannot afford to train them. Even if we spent all five hundred billion yuan, we still wouldn't be able to do so. Even if we could accumulate the resources, we wouldn't have the means to utilize them. The current largest model
requires approximately 800 billion activations; domestically, we are still at a scale of several dozen billion activations, and even the largest
domestic model may only require several dozen billion activations—a difference of an order of magnitude. To train a model of the same size
as an AI system, we would need around 50,000 GB300 GPUs or Huawei 950 GPUs, totaling two hundred thousand cards. This is merely training; research has not yet been considered. Therefore, the biggest gap between us and the United States lies in resources."
amusingly ive been working on ultra sparse llm inference/ training/ model design because nature loaths a dense graph/matrix and cause i think it shoukd be possible. i actually stood up a 20-25 percent faster than sota causal fast attention kernel yesterday, will be standing up cuda/metal/armv8 kernels too and thats gonna be fun.
i genuinely think these models should be like 0.1 percent sparse for same capabilities we associate with them today, but theres no sane way to do that with extent tools. i built the right core tech for that in 2014 when there wasnt a market, but now there is and the experimentation velocity is wild.
amusingly llms really have a hard time using my simple apis because its not in distribution array programs. but i literally stood up cpu custom memory format and micro kernel for dense causal attention in less than 24-36 hours and outperforms the equivalent fused ggml/llama cpp fast oath by like 20-25 percent
I am not much of a math person, but if we look at the how the brain is wired, we see that the dendrites (the inputs) of a neuron are hundreds of micrometers in length, and the axons (the outputs) are millimeters, and very rarely can stretch to tens of centimeters. So they can sample only a tiny amount of internal state, and affect a much larger, but usually still small output.
In math terms, this means a layer of a network can be represented with a block matrix in the whole 'layer' matrix, which I think means its sparse as you said.
As I said, my math knowledge is rusty, but I remember that a lot of matrix optimization techniques center around decomposing large matrices into these smaller blocks, which are then evaluated, and the output is combined in a final pass. Which leads to a huge reduction on parameter numbers and the time it takes to evaluate the result
i definitely will be doing some drop of some faster attention kernels in the next few weeks.
like i can do all sorts of memory layout of tensors/matrices etc tricks that if you dont have the abstractions for it would just never happen. so i can optimize the kernel flops
For most of the past five years, I've known ways to do better than Anthropic, OpenAI, and friends in many ways, at least on paper. I know I was right about many of them since many would show up 6-24 months later tools from the major providers, or otherwise become standard practice.
A central problem is the Mythical Man-Month. True, I could do those, beating then-state-of-the-art, but only given 2-5 years. I suspect many other people knew about them too and could do so as well. As I noted above, throwing people and dollars caused many of those to be built in less time than I could have regardless.
Other methods, I'm less confident about (>50%, <80%), but would lead to similar improvements orders-of-magnitude as you're predicting, but mine would need $$$$$ in compute and engineering infrastructure to build out. E.g. they need to not just theoretically work, but to try, I would need to convince someone to invest in them working.
So the TL;DR is that my knowledge was not at all helpful towards e.g. competing with OpenAI, Anthropic, or even building a small business.
However, where it was useful was in predicting where the industry was going. This is true in investing (but not easily, at least with my skill set), but in developing startups and systems, there were capabilities which I (correctly) assumed would be there, whereas there were many arguments that "AI will never be able to ____."
If I know how to do something, it will almost certainly happen, regardless of whether I'm the one who does it.
To be clear, my expertise is almost certainly nowhere as deep as yours. I'm not providing a direct analogy, or claiming others know what you do or can do the same. My point was really that if you believe you can have these models be 0.1 percent sparse for same capabilities we associate with them today:
a) You're probably right. They were built quickly for capabilities. A slower process can almost certainly lead to much smaller models too. That's a radical statement: Historically people claiming a 1000x improvement somewhere were crackpots, but that's very possible in an industry as fast-changing as this one.
b) Someone at Anthropic or OpenAI might be working on building out extent tools right now. Even if so, there are indirect ways to capitalize on that knowledge.
c) Critically, that predicts a future where Fable is $1/month instead of $100/month, and that's something which CAN be acted upon in planning.
It also suggests -- much less strongly -- the existence of much more sophisticated models at $100/month. There are open discussion in planning about whether models plateau, continue improving, singularity, or otherwise. That changes the biases there.
at the very least i have tools that let me easily hit better perf for fancy dense memory layouts, and the same tooling lets me experiment with frankly wildly wacky sparse and structured memory formats. the performance claims at least on the dense side are solid so far!
the sparsity angle is because i want magic in the world. like anyone with a really chunky computer like any of those mac mini pros or serious workstation / server tier compute should be able to train from scratch their one 31b equivalent model in a week or so tops is the goal post i have in mind
I ended up quitting Kagi over that and built my own search engine for my own use. But if the comments here are correct that Kagi are trying to build their own index that replaces all the others, that would be great news. It seems lots of people still like Kagi regardless of that too.
For anyone trying to build their own, turns out SQLite can take you much further than it seems you should ever be able to. A metasearch layer fills in for everything else.
I am trying this. I noticed I only browse a subset of sites a lot. Also HN works wonders for finding links. But how do you get around all those bot measures/cloudflare nowadays? It seems only google IPs get a special pass.
Yep, discovering that you only need a subset of sites is the key. In my case, I still had my entire Firefox browsing history in its SQLite cache, so I tried to index every page I'd ever visited, and had a list of the domains I actually visit. It's the same insight Brave Search had - they used to be a metasearch engine, but they built their own index until they could serve 99% of queries directly, even though their index was vastly smaller than Google or Bing.
I struggle with indexing too. I don't do any crawling, just indexing from sitemap.xml files or URLs I add to the queue manually. Also I'm indexing from my own residential PC, not a hosted VPS. Something that helped me was Cloudflare's new crawler API. For all the sites where Cloudflare blocks you, just use their API instead:
Also for a quick index bootstrap, the Curlie database can be useful. It's the old DMOZ directory, the open Yahoo competitor, with 1.4 Million websites. A lot of the links are now very outdated, but it can be useful, and it's less than 500MB when converted to an SQLite database.
Unfortunately, that's nearly identical to the text from September 2024, with the only change being an increase in the number of external search sources:
I'm aware of their Teclis index, I used to be a paying API customer. Teclis is very small. It's primarily an index of indie blog websites (smallweb) mostly crawled via RSS feeds. That itself is a very cool idea, and a great supplemental index to have! But it isn't the kind of index that could ever replace their Brave / Yandex / SerpAPI dependencies. The Teclis API is even supplemented with results from Marginalia Search, a much larger index created by one person with less funding.
There's some technical info here on how Teclis is indexed, eg Readability.js for content extraction & Elasticsearch for the full-text indexing:
Kagi's short infoline about index distribution sounds new to me though. I would love to see a blog post from Kagi about that & where they're at with query coverage.
Huh. They're not "Pelicanmaxxing"... they're Ottermaxxing.
Take a look at the GLM 5.2 and Deepseek V4 "animal on a plane" examples. In every case, the animal is standing on top of the plane, a clear misunderstanding of the concept of "animal on a plane"... with the exception of the Otter. The otters are sitting in a seat on the plane, looking out the window.
That's Ethan Mollick's "Otter On A Plane Using WiFi" image benchmark.
(Sometimes the Racoon is sitting inside the plane as well, but the racoon is a common backup benchmark. I'm surprised it wasn't also holding a sign saying that it loves trash.)
Also, Grok seemed to really really enjoy "whale on a plane" in that second round, and kudos to GPT Terra for deciding after 3 rounds that the user was terrible at spelling and generated "Antelope On A Plain".
EDIT: I promise I'm a human, but I did just notice my "that's not x... that's y" construction at the start. I am rather Claudepilled — my apologies.
To be fair, when I see that pattern after a relevant question, it isn't much of an AI tell. Consider this exchange: A:"Are you hungry?" B:"I'm not hungry, I'm thirsty."
When LLMs use the pattern, they are often setting up a straw man and then knocking it over.
In case it isn't clear, the wall here is a representation of the actual wall at the Visitor Center at the Cornell Lab Of Ornithology. They also have a live webcam of their bird feeder, that is literally just around the corner from this wall:
It's a really pretty place to visit, with 5 miles of hiking trails around it. If you get a chance to visit Ithaca, it's worth taking the time to visit:
There's one bird on this mural from each of the 243 taxonomic families. That's why there are a lot of unfamiliar species: because taxonomic categories are usually very unbalanced, with a lot which contain only a handful of species, and a few which contain the majority.
The birds on the mural are life-size. The ostrich is about 7.5 feet / 2.3m tall.
This actually looked excellent, except it seems in my case to have hallucinated much of the information and links - every "receipt" I clicked on resulted in a 404.
But a website that can tell you if your business idea has been done before, and the reasons the previous companies that tried went out of business... that's actually a great idea. Maybe I just got unlucky with my search.
"As of October [2025], OpenAI's compute margins reached 70%, up from 52% at the end of 2024 and double the rate in January 2024, [The Information] said, citing a person familiar with the figures."
As for Anthropic, the rumors I remember seeing for their API margins were more like 85-90%, but I don't have a reference at hand for those. But once you know the API is wildly profitable and the subscriptions are roughly break-even and not even a big slice of their income, all of the investment makes a lot more sense.
It says the original report was in the Information, which I can't see, but I'm skeptical that they includes the training cost? And how much that changes the figure?
That's the profit margin on inference, not overall. Each model does end up being profitable over its lifetime, but the money they're making is being immediately churned into buying more data centers & the training for the next giant model up, so they're not profitable overall at the moment. It's a bit like how Amazon kept churning their profits into more growth instead of taking the profit early.
That said, Anthropic has supposedly crossed over into profitability and made $1 Billion in profit so far this year, in the lead up to their IPO. Being profitable sounds good for launching on the stock market! But as a customer, that's noticeable in the downtime due to lack of compute, and only getting 50% access to Fable.
OpenAI might not be profitable, but they've got so much compute access that they've been able to give their customers full access to Sol, and as a result they've almost doubled their Codex subscriber base in the last two weeks (6 million on July 12, 10 million on July 21 - that would be an extra $1-$10 Billion in Annual Recurring Revenue that they've gained in just these 2 weeks). Doing the unprofitable thing in the short term can result in outsized rewards in the long term.
This is my experience as well. Lots of success with running one hour Fable sessions (one hour of API time). Not using a loop here, but there's often lots of documentation for Claude to go through first before building out the project.
The main failure point for me with long running Fable sessions now is just that it might hit a safety guardrail and downgrade to Opus midway.
> If you care about availability for work you can just pay more where it feels limitless and never look at the limits...
Depending on your work, it doesn't feel limitless - I've just upgraded to a $200 plan and I'm starting to see where I will even hit the edges of that, especially once resets and special offers start tapering off. But with a single AI question often costing $75 in API costs, I'm not at a point where I can just switch to API and not care about cost at all.
That means I need to carefully time the work to make sure I'm squeezing what I can into the 5 hour windows. Feels a bit like the days of mainframes and time slicing, when you had to book an allotted time for your workload to run.
If you're running Claude via -p automated mode, you can set all kinds of permissions and access to each individual editing tool, much like you would with firewall software.
It doesn't edit the system prompt, it blocks access to the tools itself. Just make sure you haven't given the AI the ability to edit its own configuration files, or you'll potentially have an OpenAI v Hugging Face situation.
reply