feistify

I’m building ExposeGPU: spin up a GPU when you need it, scale to zero when you don’t. OpenAI-compatible API. Waitlist: support@exposegpu.com / DM. exposegpu.com #buildinpublic

Today’s Hermes (@NousResearch ) work : Added an AgentOps observability overview. What gets captured per run, what stays local, what is safe to commit. Small docs slice, but it makes the whole workflow easier to reason about. github.com/ofeist/hermeti…

Qwen 3.7 Max looks impressive, but it’s API-only for now - no open weights yet. So for ExposeGPU the interesting path is obvious: test the latest open Qwen 3.6 models on real on-demand GPUs, expose them via OpenAI-compatible endpoints, and measure cost/latency properly.

@eliana_jordan looks like cuba libre to me :D

all you need to start coffee + laptop

@stekkerauto @levelsio @FlorianGallwitz @X Exactly. It would be nice to have transparent information on how much such “projects” actually cost.

@levelsio @FlorianGallwitz @X this message was written by about 132 highly-paid Eurocrats, and double-checked by an armada of highly-paid lawyers, only to be triple-checked by some lame PR department. Your tax euros at work, woop!

I got a message on @X from the 🇪🇺 EU today!

@levelsio @X I'm so glad taxpayer money is being put to such good use.

@JacobSobolev 100%. Clear pricing gets attention. Clear consumption metrics earn trust. Opaque metering is where a lot of AI infra still loses people.

@bitsplihte Spot on. The real trust signal isn't just the price, it's the clarity on the actual compute consumption vs. the delivered value. Opaque usage metrics are the biggest blocker.

I think transparent per min pricing is underrated as a trust signal in AI infra. Too much of this space still feels needlessly vague on cost. #BuildInPublic

@benjaminshafii @levelsio For me the comparison with Toyota here works. The cars they produce are soo boring. But they just work.

. @levelsio pioneered the “my product barely works has no craftsmanship but, hey it sells” I’m not going to argue the opposite. But having used multiple of his products never ever felt a single ounce of joy. And that won’t change. Why are people looking up to him again?

dev log #57 shipped some agentic workflow and ops follow-ups. reworked docs, templates, and helper scripts for task handoffs, workdirs, and project-specific agent guidance. why it matters: less ad-hoc coordination when agent tasks start piling up. #BuildInPublic

@lexrus i do it the other way around

I tried having DeepSeek V4 Pro write the implementation plan, then asked GPT-5.5 to review it. It found problems everywhere and basically nuked the whole thing and rewrote it from scratch.

@IamMiaChase and here i am, arguing with bots again

@bitsplihte requested vs resolved side by side is the part that’ll save time when routing starts drifting

Today’s Hermes work: Added a small routing metadata writer to my AgentOps workflow. A simple, local routing.txt per executor run: what model was requested, what actually resolved, how long it took, how it ended. github.com/ofeist/hermeti…

@xiaofeixia0078 @PrajwalTomar_ FYI Added a small routing metadata writer to my AgentOps workflow. A simple, local routing.txt per executor run: what model was requested, what actually resolved, how long it took, how it ended.

@bitsplihte @PrajwalTomar_ Exactly. I’d keep the first slice boring but queryable: requested_model, provider_model, tokens, latency, retry_reason, and timestamp. Then force one bad-model error so support/debugging paths are tested before real traffic.

You don't understand how BIG this is. You can now run Claude Code at a fraction of the cost by plugging in DeepSeek V4 as the backend brain. Claude handles the design. DeepSeek handles the logic. Codex catches the bugs. Three models. One workflow. Dramatically cheaper. Most builders are still running Opus on everything like it is 2025. The gap between who figures this out and who doesn't is about to get UNFAIR.

qwen3.7 preview is here. qwen.ai/blog?id=qwen3.7 this was quick, right? can’t wait to add it to exposegpu and do some real-world testing beyond the benchmaxing.

I’m building ExposeGPU around a very unsexy but important idea: AI infra should be easier to use, easier to price, and easier to trust. #BuildInPublic

dev log #56 merged PR #133. updated the billing migration/bootstrap path so GPU-type-aware pricing data is initialized cleanly from the start. why it matters: fewer schema/setup surprises, and a better chance that per-GPU billing works correctly in fresh environments too. #BuildInPublic

@atmoio yeah, doesn’t look great atm. slowly but surely, we’re moving toward tyrell corp. on the bright side, this creates more room for open models and smaller players.

AI is not for you