LLM post-training - Page 15
Stories from LLM post-training come first, newest first, then stories from connected topics (Kimi K3, Qwen3.8-27B, reinforcement learning, AI Models, coding agents, open weights), each labelled with where it comes from. This view is not indexed by search engines.
Latest
micro1 Says It Will Spend $1bn on Company Data to Train AI Agents
micro1 says it will spend $1bn in 12 months licensing company data to train AI agents, with Citi and Hercules Capital money and payouts from $100k.
Underdog’s Saluki 27B Shrinks Qwen3.8-27B to 7.89 GB, at a Maths Cost
Underdog's Saluki 27B is a 7.89 GB, 2-bit version of Qwen3.8-27B. Underdog says it passes more tool-calling tasks than the full model on its own 120-task test but scores 15 to 18% lower on competition maths. What quantisation means, and what the numbers do not prove.
Firmus Pulls Its ASX IPO as Isomorphic Labs Eyes a $40bn Valuation
Three AI business items from 8–9 October 2026: Nvidia-backed Firmus withdraws its ASX listing and turns to private money; Bloomberg reports Isomorphic Labs is in early funding talks at a valuation of at least $40 billion; and StepFun's Step 5 Preview model is now listed on OpenRouter and Vercel's AI Gateway. What is confirmed, and what is only reported.
Watching AI Agents From the Inside, and Rewiring the AI Data Centre
Two AI infrastructure launches from 8 October 2026: Goodfire's "inside-out" monitors, which read a model's internal signals to catch harmful agent sessions, reach Baseten customers (all performance figures are Goodfire's own), and Upscale AI unveils Token Fabric, a networking platform for AI data centres with general availability planned for early 2027.
Harness Buys Augment Code Assets, Plus TRL 1.15 and Genkit Dart 1.0
Four AI developer-tool updates from 8–9 October 2026: Harness buys selected Augment Code assets (not the whole company, price undisclosed), Hugging Face's TRL v1.15.0 claims up to 6.9× longer training sequences, Google's Genkit Dart reaches 1.0, and NVIDIA shows its Boro kernel-review tool. What each says, and what it doesn't.
Minimal Coding Agent Matches Elaborate AI Harnesses, Preprint Finds
A preprint from EPFL researchers, listed on Apple's research site, finds a single minimal coding-agent session matched or beat elaborate machine-learning agent systems in its tests. Pasqal releases skills that let coding agents prepare quantum experiments, and Microsoft makes 30 Dynamics 365 CRM skills generally available, with Autopilot still in private preview.
Two Open Models From 8 October: a Fast Coding Agent and a Gene Finder
JetBrains has released Mellum2.1, a small open model for coding agents, and Hugging Face has released Carbon-A, a 1.2B-parameter model that predicts protein-coding regions in DNA, plus a database of 566 million predicted loci. What each does, and which numbers are the developers' own.
Same AI Model Scores 33% or 62% Depending on Its Coding Tools
Hugging Face's FineEnvs team has released an open framework that turns unmodified coding-agent harnesses such as Claude Code, Codex and OpenCode into reinforcement-learning environments. In its own tests, one small model scored 62% in one harness and 33% in another. What a harness is, how the training works, and what the results do and do not show.
NVIDIA Jetson Demo Runs a Gemma 4 Vision Kiosk On-Device
An NVIDIA tutorial shows OpenAI's Codex coding agent turning a Jetson Orin Nano into a self-starting kiosk that captions a live webcam feed with Google's small Gemma 4 E2B model, running on the device. What the demo shows, what "No Mac required" means, and why local inference matters.
AI Research Tools: TasteVal Measures “Research Taste”, Karotte Targets Reward Hacking and Zephon Fixes Data Order
A new benchmark, TasteVal, says the best AI model now reaches expert-level results on eight AI research tasks with 2.3 times less compute, though its authors list reasons that may overstate progress. Preference Model has open-sourced Karotte, a framework for training environments that are harder to game, and DatologyAI has released Zephon, a data loader that keeps experiments repeatable when the GPU count changes. What each claims, and who measured it.
Latest on TSN
Explore the TSN Atlas: Every topic we cover, mapped. Click a bubble to dive in.
