Builder, operator and investor. Infra, AI, and product nerd. Opinions are my own. Checkout @lunaroute_ai (https://xitter.gq/t.co/YlZrgrIUKq) to stop caring about token limits
Joined April 2007
- Tweets15.7K
- Following4.4K
- Followers1K
- Likes14.6K
Why summarize an AI agent's conversation when the entire history is on disk?
I built a Pi extension that starts fresh when context fills up, while keeping the full history searchable.
Cleaner context, no lossy summaries.
eran.sandler.co.il/post/2026…
Eran Sandler retweeted
If you're in need of dedicated inference that will cost you a fraction of what you're paying OAI or A/ talk to @erans and @lunaroute_ai . You'd be amazed how big of unlock it'll be for your teams
We will see more and more companies doing this. Not paying by the tokens enables Continuous AI for doing all sorts of things.
If you can't afford 53 B200 nodes talk to us at @lunaroute_ai - we can get you private fixed cost inference and more than just GLM 5.3 for much less!
Not being token scarce is a HUGE unlock for everyone. Combine that with the fact you can now run Continuous AI for sweeping tickets/stories, adversarial code reviews or anything else you never end up doing because of time or cost.
Now you can - with @lunaroute_ai
#continuousai
We will see more and more companies doing this. Not paying by the tokens enables Continuous AI for doing all sorts of things.
If you can't afford 53 B200 nodes talk to us at @lunaroute_ai - we can get you private fixed cost inference and more than just GLM 5.3 for much less!
BREAKING: Revolut moved ~2,000 engineers onto self-hosted GLM for coding and pushed Claude Code and Codex to the exception list.
Per David Abitbol, who leads AI ops there: 56 Nvidia B200s, a blended 30 cents per million tokens, roughly a tenth of what they paid Anthropic and OpenAI. Engineering lead time improved 70%.
Here's the part that matters: the gains showed up when tokens stopped being rationed. Engineers quit saving AI for the hard tickets and pointed it at everything.
Frontier models become the exception, open weights become the default. That's a pricing problem for every lab heading to IPO.
h/t @OliverMolander
Would you move your team off frontier APIs if the bill dropped 90%?
I keep thinking about all the useful work we don't do simply because it's not urgent.
Slow queries, security checks, research, customer feedback, ideas we never test.
That's the work I want agents doing in the background, and a big reason we're building @lunaroute_ai the way we are.
Important doesn't always mean urgent.
Performance, security, research, data analysis, untested ideas. There's always useful work we never get to.
Continuous AI lets agents tackle it in the background. LunaRoute's fixed-cost inference and background scheduling make that practical.
lunaroute.com/blog/continuou…
Checkout @opencomputerHQ - they have been cooking some interesting things lately!
Introducing @computer.
Claude Code & Codex in the cloud, as fast as your local machine. Sometimes even faster!
Try it today via Slack, byo sub supported.
opencomputer.dev
If an AI model refuses to help you investigate a vulnerability in your own code, that's a problem.
Security teams need to explore, test, fail, and try again. We're working on giving them models that can actually help, with verified access and without worrying about the token meter.
Curious to see how far we can push this.
Security teams shouldn't hit AI refusal walls when trying to protect their systems.
Introducing LunaRoute Cybersecurity & Cyberdefense.
🔍 Vulnerability research & testing
🛡️ Threat analysis & defense
Reduced-refusal models. Verified access. No token meter.
lunaroute.com/blog/introduci…
Look at the throughput that Vera Rubin can do even on a large model as Kimi K3, it's insane:
xitter.gq/SemiAnalysis_/status/2…
Remember what I said, even Vera Rubins have the same amount of HBM memory (at the current version) as B300 - so bigger models (think Kimi K3.1, Qwen 3.8 Max and future models) that are close or over 3T params are no longer capable of running on a single node, unless quantization and other tricks are deployed.
BREAKING NEWS🚨: NVIDIA just submitted its first Rubin NVL72 Kimi K3 submission, and it can reach over 850 tok/s/user, even at 2.8 Trillion parameters, with over 3x higher throughput on @vllm_project!
Great work by the NVIDIA team, and excited for Rubin Software 1.0 GA coming soon!
Sometimes you just want to scratch your own itch.
I wanted a way to read Markdown in the terminal, including Mermaid diagrams, tables, and images, without having to open a browser or an editor.
Having practically unlimited tokens through LunaRoute makes these little experiments much easier to justify. You can just build what you need, iterate until it works the way you want, and share it with others.
Blog post: eran.sandler.co.il/post/2026…
Eran Sandler retweeted
1/ I've been having lots of fun lately, letting my curiosity lead me down rabbit holes. This is one experiment I've been working on - trying to answer if AI models have distinctive design taste.
We're launching the largest AI design experiment today - uipalooza.ai where you can compare the design capabilities of different AI models side-by-side.
Eran Sandler retweeted
5/ Credit where credit is due - @erans' project @lunaroute_ai flipped the bit for me on using AI without worrying about token cost thanks to their concurrency based billing.
1/ Everyone's framing 3T-param open models as a compute milestone. They're not. They're a memory bill and that bill is why frontier open weights live in the datacenter, not on your desk.
A thread on what it actually takes to serve Kimi K3.1 Pro: 🧵
9/ What would actually move the line: cheaper HBM, credible 2-bit quant, Engram-style memory in a flagship model, or CXL memory pooling across nodes.
Any one of those changes the math. None of them ship this quarter.