We are cooking some pretty interesting products around post training and inference @tplr_ai .
Would love to talk to engineers and researchers to understand where you bottlenecks are .
The rewards are my eternal gratitude and some credits.
DM if you're up for it.
Crucible, Templar's pre-training platform, has completed its first production end-to-end training runs. The latest trained an 8B model on 50.53B tokens across 48 distributed A100s, at an estimated $0.1202 per million tokens of GPU rental.
The run reached 48.3% effective MFU. At AWS p4de Capacity Blocks pricing, a 48-A100 cluster operating at the literature-derived 65% compute ceiling comes to an estimated $0.1686 per million tokens. Crucible's measured $0.1202 was about 29% lower after its low-bandwidth overhead.
The comparison excludes R2 storage and operations. The full writeup shows the method and accounting.
Blog: tplr.ai/publications/blog/in…
i'm opensourcing my playbooks; up and running automation to have a local RL & inference cluster on a GB10 (DGX Sparks, ASUS Ascent GX10, etc..)
github.com/epappas/gx10-clus…
Introduction to LLM Design & Covenant 72B: This week at the Ethereum & AI Engineering Group meetup I provided a 15 minute introduction to LLM design and then dived into the design and training of Covenant 72B; the largest model trained over the Internet.
youtube.com/watch?v=TlVQbiTd…
That shape is all over this year's agent work.
Self-Harness ran one improvement across three models and got three different sets of harness changes, because each model had different weaknesses to fix. Its abstract calls effective harness design "inherently model-specific."
You cannot write this off as a naive scheduler either.
Continuum is built for multi-turn agentic serving, and its KV hit rate still fell from above 90% to about 60% once batches grew enough to create memory pressure.
Memory gets tight. Between one model turn and the next, while a tool call is out, the inference scheduler evicts the loop's KV cache. The tool returns and the server prefills the entire growing history again.
An agent loop that passes every test it has can run seven times slower the moment it shares a cluster, with nothing inside it changed.
open.substack.com/pub/hypera…
[1/9] Edinburgh Decentralised Finance Summit 2026.
Tuesday 1 September, Royal College of Physicians of Edinburgh.
A full day on institutional DeFi, tokenisation, risk management, regulation and autonomous finance.
Around 25 speakers. 300 attendees.
luma.com/scotland-kgav
Two runnable demos, red-teamed before I trusted them. One caught me faking my own leak: a partial block real vLLM never caches.
Fixed it, re-verified from scratch. Full writeup + code: hyperautomation.substack.com…
The confidentiality fix shipped a year+ ago: a per-tenant cache_salt, assigned server-side. If the caller picks it, an attacker just picks yours.
The fairness fix, reserved cache partitions, is still research. Not a stock engine flag.
One design choice causes both: keying the cache by content and sharing it across tenants who don't trust each other.
Residency becomes observable (by timing) and unprotected (from eviction) at the same time.
Same root cause, second bug.
That shared pool evicts by recency. A noisy neighbor's flood flushes your warm cache no matter how high your priority.
Priority orders the compute queue. It doesn't touch eviction.