shostack

The cruelty is the point.

For individual use without specially negotiated Enterprise stuff I can't afford both offer contractual assurances of not using your data for training or ads and measurement and selling your data.

But these are not the same level of technical assurance you get from say, a zero data retention provider on OpenRouter.

Right now I am finding I have to tolerate substantial friction to use Hermes for personal stuff with a ZDR provider and ChatGPT and Codex for less personal stuff because the products and models are simply so much better.

Debugging any LLM output when you also have done substantial harness engineering is a total pita and I wish there were better tools for it to isolate issues.

I spent ages tracking down start appears to be an issue with the current Deepseek v4 flash 0731 version that would cause it to output giant walls of gibberish in Hermes with reasoning turned on.

I needed this. I had no idea it existed.

I have bent over backwards trying to enforce brevity with deepseek v4 flash to the point where I think I broke some things trying to do prompt injection in my Hermes setup and was still unsuccessful.

Meanwhile Sol blows me away and I want that to be my default for everything now.

In general though I seem to have the most success with a "<=10w" requirement in my prompts.

What I don't see listed and would be a good comparison is the STS models. OpenAI's live model is an absolute joy to talk with.

Having been at smaller companies without the data, tooling, discipline, and resourcing to conduct viable experiments, and then being at a company that is actually one of the best in the world at it and building solutions for these problems at scale for marketing teams, so many problems come back to the human element in how data is tracked, how teams collaborate or fail to which can lead to pollution and dirty data, and how decisions are made as the moment trade-offs need to be considered it becomes personal and enters messy human relationship territory.

But ultimately from a pure "did this work or not" standpoint you are right. Incrementality experiments are the gold standard.

So like, some spren from Roshar?

Could memories and blocks of context be made pluggable and unpluggable in this manner to "hot swap" context?

Specifically, I want to use the ChatGPT Android app to use it like I would a desktop remote control session. It is a nice mobile interface. I do not want to have to use a mobile terminal app like Termius. The UX is awful.

I already have that effectively via pinning Fireworks on Openrouter. I am unclear how this would solve the architecture issue I raised though.

Love the post and have a similar setup. But I have questions about how you handle sensitive personal info with Codex and Claude with this setup given Hermes has no boundaries on its own with its integrations with Codex for example. It can leak memory and tool calls and context through.

At home I have Hermes on a VPS with matrix and mnemosyne and it is largely for personal use. While I was excited to try out the seamless Codex wrapper, this leakage was unacceptable compared to a ZDR OpenRouter provider like fireworks.

Where I struggle is now in architecting a clean way to have Hermes separate out personal stuff from coding stuff, and then orchestrate coding through Codex as that seems to be the only way to use it headlessly on my phone since they have no way of doing that in the ChatGPT app right now on Linux.

Does Buzz let you neatly silo that?

I'm struggling between: "I want to use Hermes with frontier models with a seamless wrapper on my $20 ChatGPT pro subscription so I can go through my Hermes setup for everything including coding"

And

"OpenAI and Anthropic are not ZDR and I am not comfortable sending sensitive personal and family data to them. But I at least want a mobile-friendly remote coding experience with them on my phone and VPS"

It is great for getting aggregate insights about the privacy and security picture of model providers.

For example if you want zero data retention and US -based hosting you can find that easily. You will not find that through Deepseek.

Am I understanding correctly that I still can't run a headless session on my VPS in a way that let's me use remote control for a better 1st party mobile experience that doesn't require my laptop to be awake?

I really don't want to have to boot into a whole desktop Ubuntu instance simply to run this.

I've gone as far as starting to build my own handoff layer with Hermes as I want to use codex purely for coding in a more isolated fashion and not have it co-mingle my memory and unrelated sensitive context and tool calls. The newly announced integration was so close to what I wanted and yet so far for anyone that wants to keep sensitive data separated.

I wish they offered a consumer zero data retention plan.

Having flown in VATSIM towered air space and airports it is some of the best VR flight sim time I ever had.

Had someone on ground and tower at KPAO and several of us pilots were just flying the pattern and it was a blast. Not to mention one was a licensed pilot based out of there who gave me IRL tips for my lesson that week about the field which were really helpful.

Such a remarkable community.

Except California where it is expensive due to pg&e and corruption.

Are there any Android remote terminal apps that don't suck on mobile? Termius is painful.

I felt that way initially. It is also a headache when you invest a lot in that and need to continually test ripping that stuff out as new models come out. Or in the case of opus 5, Anthropic says ditch it entirely and trust it.

But there is another aspect which I do enjoy which is closer to the feeling of dialing in key bindings in vim or getting a really good rhythm going with your vscode extensions or zhs plugins. It is that level of "I want my system to do exactly this thing in exactly this way" customization that a lot of technical people crave.

And you can do it with harness and context engineering in many cases. In other cases it introduces friction because it will be like "cool, I will only output 15 words max unless told otherwise" and then in the next turn completely disregards it with an "oops, you did tell me to do that didn't you."

And that frustration compounds when older model versions may have done a better job of that but new models are like "thank you for your suggestion, your opinion, while appreciated, is irrelevant. Now let me get back to overspending on your token budget. "

Because AI enables a new layer more akin to an actual OS. But also because I imagine they all want all the things running through them and an OS is the widest net they can cast in that direction.

> * When an action requires approval, the agent does not need to stop and wait for it. A Gatekeeper will simulate the outcome, allowing the agent to keep running and queue up more work. You can then approve everything in a batch at the end. Hopefully, this means you no longer feel the need to turn on auto-approve! (But you still can if you want.)

Can you say more about how you simulate outcomes of approvals? For example, if I need an approval to be granted permission to read a document, how could it simulate the document contents?

An example of this in action is the utter inability to get deepseek v4 flash (even the new version) to stay concise. I have jumped through all sorts of hoops with deterministic checks, pre-message injection hooks, memory framework, etc and when it fails still and I ask why it essentially says "I forgot."

This makes it unreliable and preferences are things I need to assume are treated as exactly that, preferences, not hard settings.

It is an area where it is more like working with an unreliable human than I would prefer.

The closest I get to that is when I am getting a massage actually. If I get a 90 minute one, about halfway through my brain stops racing and I enter this, not quite hypnogogic, state where things are quiet and fluid and peaceful. The closest thing is waking up after a good night's sleep with an empty brain before the weight of the world comes crashing in.

Am I parsing the providers correctly that if I want a US-based zdr provider that offers the best price and performance that is currently Cloudflare and no longer Fireworks (which is that it was with the previous version)?

What am I missing?

Introduce a larger one that you have to sign up for digitally to wait in line and be notified when it is available and charge for the privilege. Hell, make it auction based so if you want to jump the line you can if you pay for it.

Is fireworks not private enough?

Or for anyone on Windows 10 Home

I have built out something similar that let's me use my phone's hardware buttons to open an input stream with the mic for my Hermes agent over my matrix gateway and then has it play back with a local TTS model on my Pixel 10 Pro.

But the Deepseek v4 flash model I am using through OpenRouter is killing me on latency. Any suggestions to improve that?

I work with someone who did a lot of work with this to improve our ability to generate awesome visualizations with little thinking. It's a very powerful language but needs guardrails and guidance, particularly if you want end users to be able to produce consistent and standardized visualizations without knowing anything about it.

Affiliates trying to get you to buy something have been using fake reviews since the beginning of time. AI simply helps do it at scale.

Generally when people have a house cleaner they know who it is and have some form of relationship with them that establishes a baseline of trust. That would not exist here.

The world would be a smarter place if every study and resulting news article were required by law to include something like this at the top.

If you're using Matrix, consider Hermes as a harness if you haven't already. Native gateway support. I've been primarily using mine through Element and it has largely been great.