Sam Hogan 🇺🇸

@TreSmithers Hey man, been trying to get a hold of you to chat about your boy marc. He did a vid on my project last week and we want to sponsor. shoot me a DM. instagram.com/reel/DXzLnoAOS…

Today is about to be a good day!

Today we're releasing Agora: the first ever pretraining stack that allows non-collocated consumer GPUs to be competitive with centralized clusters Agora is 15x faster than Megatron-LM in this setting and is only 1.5x less efficient in terms of tokens per unit compute than TorchTitan on H100s, despite running on devices that have no NVLink or InfiniBand support.

@vivek_2332 Very cool

releasing /synthetic-self-improve-rl. claude code (teacher) skill that designs/writes the synthetic data, env and rewards to post-train a smaller model (student). it post-trains the student on a real dataset, reads its failure traces, then writes the synthetic data, the verifiers env and the reward function to patch the gaps. re-trains. loops. loop: -> baseline on real data -> analyze low-reward rollouts -> generate ~500-1000 row synthetic dataset -> write a verifiers env + rubric around it -> resume from the post-trained checkpoint -> eval on the real test split -> keep what helps, iterate on what doesn't 1. result: qwen3-0.6B-base on gsm8k. 700 synth rows bumped it from 0.7854 -> 0.8158 on the full test set. 2. run it for any wall-clock budget or iteration cap you set. the loop keeps running until the budget expires. 3. built on @willccbb verifiers and @PrimeIntellect for training. works on any env that has a train and eval dataset. p.s. still figuring out what to call this. feels adjacent to @karpathy autoresearch or synthetic envs?

i’ve been in sf for a month and i’ve heard “agentic” more times than i can count but haven’t heard “free cash flow” once

The best production model is the one trained for the job. Gravity Ads replaced a 70B model on Cerebras with a specialized 1B model trained for their actual workload. Same quality, much faster and cheaper inference: - p50: 152ms - p99: 5.7x lower - cost: ~10x lower - model: 70x smaller Great working with @trygravityai on this. Case study: inference.net/case-study/gra…

@hwchase17 it would be really cool to see LangSmith Engine open sourced like halo github.com/context-labs/H…

great deep dive into how we built LangSmith Engine lots of fun learnings and tips and tricks

@Inner_Axiom Nice

big day of building today We’re now doing RL training on the runtime of our new agent framework The implementation is a loop: run the native agent runtime through real and ambitious economic tasks, trace every step, score behavior against verifier-backed outcomes, analyze failures with HALO, patch the harness/runtime, then rerun held-out tasks to verify improvement The tasks are intentionally long-range and block-heavy. Missing credentials, accounts, tools, buyer channels, verifiers, or capabilities should not end the run. The runtime is being trained to convert those blocks into next actions: research, setup requests, adapter work, alternate routes, or smaller proof steps So the training target is not just task completion. It’s persistence under real constraints, with evidence-backed rewards and measurable improvement across repeated runs

we cast spells on the sand and made it think for us

Good primer on what Agent Optimization means in practice and how to get started

@ejc3 No

@samhogan OpenClaw

Why hasn’t OpenAI made a native Slack integration where me and my team can collab on Chat and Codex threads? This feels like an easy win. What am I missing?

@sierracatalina I think the network effects of being in a company slack would out weigh the platform risk and they could eventually migrate folks off to a different multi player experience. once multi player coding is learned as a foundational input it’s going to be very sticky

@samhogan slack owns anything that happens inside their platform.

@hyojun_at Yeah exactly. ideally from slack

@samhogan this seems a good idea. like syncing Slack thread <-> Codex thread 1:1 and allowing multi players in a codex thread?

Not having to manually write code means I have more time to strategize on what code is worth writing Strategizing means working with team mates, discussing ideas, & sharing context, which are fundamentally collaborative Multi player codex could be so good for this

Yes I know I could hack something together but I want native multi player codex threads.

@netcapgirl heck ya

situation update: i moved to sf

My favorite detail about 'free' local inference is the depreciation math. If you amortize a $4k Mac over 5 years, running a 31B model costs $1.50 per million tokens. The API is 3x cheaper. Local compute is officially a luxury good and I respect it ✨

@signulll @SinaHartung

@SinaHartung 69% of the time, it works everytime.

@signulll lemme use you, babe

@willccbb could have been a typescript file

this skill could’ve been a specialized subagent