Can an MCP provide prompts for your model to download and use CLIs (and ensure they have the right versions of those tools) in such a way that the data flows through the client side tools?

The more I read this thread the more I'm convinced that the main value of MCP is to provide a server managed release process. This is the same advantage that SaaS has over client side apps.

However MCPs couples with a client willing to run tools locally can provide the best of both worlds

As far as I know, the only way an MCP can provide you data that doesn't go into the context is by providing URLs to the data, and then the model uses e.g. curl to access that data for data manipulation purposes. You could also return result set ids and provide accessors to such data, but ultimately you'd need to provide powerful accessors to that result set to avoid polluting context. Things like shell with all its power already provides.

It seems like there's little point in MCP in that case. Maybe more point if it was a standard mechanism for MCP to provide additional data, in a completely compatible fashion with all other tools? You could perhaps even pass such URLs to other MCPs. You could have an MCP for jq for doing stream processing. Starts to look a lot like a shell, though.

Seems like MCPs could also be extended to store auxiliary data to your memory (or filesystem..), and then an additional extension to provide that kind of data as auxiliary data in the calls to MCP.

Well, even as is, MCP still provides a standard method of using OAuth for accessing such services. And you must use MCP if you wish to add something to the ChatGPT.com web service, so it's easy to see why OpenAI folks are seeing companies going that way.

Possibly. But otoh one cannot complain that agents don't produce high quality code while at the same time not allowing them to thoroughly go through all the steps required to produce high quality code

How do you determine what's a fair price?

Sometimes it's hard but with things like gold it's pretty obvious.

Who's currently getting a below-market deal on gold from Africa?

A local director of a gold mine here in Tanzania told me that all the profits of the mine go to the owners in China. I don't know how much profit they are making, but if Tanzania as a country does not benefit from the mines then the deal is better than it should be.

The informed claim is not that the obscurity layer has no value. Quite the contrary, it has such a great value that it basically reduces the incentives to have great proper security and thus once the obscurity layer is breached the second line of defense is weaker.

The argument is that it's much easier to secure proper key material rather than design and config information that can often be leaked accidentally because it's actually directly manipulated by humans (employee onboarding, employee churn etc)

That's an interesting way to describe it. It's kind of like the turn away from requiring regular password updates. On paper, password rotation is good. But when you consider its interaction with human psychology, the policy makes security worse by causing people to make bad decisions.

Even taking into account the fact that they are billing at 75% discount it's still quite cheaper

Aren't they all billing at discount?

Anthropic's and OpenAI's costs seem to include a fairly ok margin, from the very fourth hand info I have.

In total, how many hands do you have?

Enough to reach the bottom of the rabbit hole.

Counting turtles on the way down?

If I was a betting man I'd bet that at least one of those hands is an LLM

Those aren't their hands.

> Aren't they all billing at discount?

Microsoft just announced the availability of OpenAI GPT-5.5, which they are charging 30x for it. In contrast, they charge 7.5x for Claude Opus 4.6 and 1x for OpenAI GPT-5.4

Check out the token-based pricing, and compare GPT-5.5 with all other models.

https://docs.github.com/en/copilot/reference/copilot-billing...

Actually it is $30 for GPT 5.5, $25 for both Opus 4.6 and 4.7.

If you're referring to the multipliers that are used for subscription-based usage, GPT 5.5 is not available yet (according to https://docs.github.com/en/copilot/reference/copilot-billing...) and Opus will be at 27x at the end of the month.

When I check GH Copilot right now, it looks like Opus 4.7 multiplier was increased to 15x (I think it was 6x just a few days ago) but 4.6 is still at 3x. But these relatively cheap multipliers exist only until the end of the month.

Aren't OCI caches content addressed?

I was thinking of people were to use an image…:$my_tag on the host cluster and some roughe pod on the child cluster (but same underlying physical nodes) somehow overwriting the local cached :my_tag, you could do something on the parent cluster.

But I don’t fully understand what you meant with content adressed :)

Maybe one has to ensure in the host cluster that the image pull policy is set to Always or all references to images have to be based on the shasum rather than Tags.

Sparc (not a VLIW ISA) also had rotating register windows. But ia64 had a twist on it: the register window size was dynamic and "allocated" by the callee with an alloc instruction

The only other ISA I know of that did something similar was the Am29000

The Am29000 modeled it in an interesting way though:

The register file consisted of 128 global registers but the instruction encoding allowed to specify an "indirect register index" mode where the operand register was computed from the content of gr1 plus an offset. Thus gr1 acted as a "register window stack pointer". I _think_ such a computed register index would then be used to index into a separate register file for locals (and arguments etc) but I'm not sure.

Anybody here is familiar with this quite old ISA?

(I'm really interested in the richness of the CPU design space, the history of which is fascinating)

Is there evidence that frontier models at anthropic, openai or google or whatnot are not using comparable optimizations to draw down their coats and that their markup is just higher because they can?

The auth between your app and the proxy can be scoped more easily.

For example if the proxy runs in localhost you can trust the localhost workload.

Or you can use some other kind of workload identity proof (like cloud based metadata servers). If you leak such a key no other VM can use it, because it's scoped to your VM

That's not true. Both AWS' as well as GCP's workload identity tokens are not bound to the VM. If you leak the credentials they're valid until they expire. on AWS the expiry is 6 hours (non-configurable). Even if your IAM role has a shorter expiration, the credentials assumed by the VM will always be valid for 6 hours.

That entirely depends on the location of the proxy and the extra conditions you can express. E.g. you could bind it to a source IP and have the proxy check that, or use some overlay network (like tailscale does)

My point was that you don't literally have to run the proxy on localhost in order to scope the request.

in exchange for service that presumably a) costs something to amazon to operate (so not pure 100B profit) and b) anthropic would have to spend anyway to operate their business.

so basically ...

you could view this as a kind of discount, but instead of paying less later, you get some cash now and then pay full later.