Sift.

Week 2026-34 · Aug 17–23, 2026

10 stories · 17/20 feeds live · $0.18 run

Week at a glance

Models & Research 3Tooling 3Infra 2Business 2

hi 9 · lo 4 · avg 6.0

Models & Research

7
#3

Qwen 3.8 27B matches much larger models on intelligence benchmarks

Qwen 3.8 27B, an Apache 2 licensed vision-capable model from Alibaba, scored 52 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Luna and trailing far larger models by one point. Simon Willison notes it is excellent but tends to overthink by default.

Notable capability-per-parameter result relevant to open-weight local deployment.

5
#6

Z.ai CEO on GLM-5.3 and a post-training scaling lawneeds verification

Latent.Space covers Z.ai CEO Jie Tang discussing GLM-5.3 and what he frames as a new post-training scaling law beyond parameter count. The discussion touches on shifting emphasis from model size to post-training.

Model-release and scaling-research signal, but via newsletter interview.

5
#7

Simulation and autonomous AI researchers as an emerging frontierneeds verification

Latent.Space and Import AI cover a wave of interest in simulation as a scaling approach and in autonomous AI researchers, including Simile AI's plan to build billions of human digital twins. The pieces frame simulation and recursive self-improvement as the next research frontier.

Interesting research direction but speculative and secondary-sourced.

Tooling

9
#1

Asana clears 5 years of engineering work in 2 weeks using Codex

Asana used OpenAI Codex to replace an outdated testing framework in two weeks, work it estimated would have taken five years, for roughly $12K. The Pragmatic Engineer places this in a broader trend of AI-driven migrations across engineering organizations.

Concrete enterprise coding-agent deployment with hard numbers, corroborated by independent analysis.

7
#4

Agent harnesses, coding-agent workflows, and verification practicesneeds verification

Latent.Space examines how agent harnesses are being absorbed into model weights, while Simon Willison and Matt Pocock discuss verification-over-line-review practices and planning skills for coding agents. Together they map evolving best practices for agentic software development.

Core agentic-SDLC content matching reader's stated focus; practitioner analysis.

5
#8

Mojo programming language is now open source

The Mojo language released its compiler and toolchain under an Apache 2 license, following its 1.0 release. The move fulfills an open-source promise made in 2023, though Mojo's Python-superset goals have since shifted.

Relevant systems/AI tooling release with concrete licensing detail.

Infra

7
#2

Memory prices reportedly up 500% in 12 monthsneeds verification

Latent.Space reports memory prices have risen roughly 500% over the past year, describing it as Moore's Law effectively reversed to 2007 levels. The trend has significant implications for AI datacenter and hardware costs.

Directly relevant to AI hardware economics and compute costs; needs primary verification.

5
#9

Hugging Face engineering: inference speedups, agent memory, and GPU utilization

Hugging Face posts cover up to 3.2x faster LFM2.5-DSpark inference, agent memory requirements, a 33-point GPU utilization gain from reordering work, multi-vector embeddings, and ASR benchmark optimization. The collection focuses on efficiency and evaluation for real deployments.

Practical infra/efficiency techniques with numbers, relevant to deployment costs.

Business

6
#5

AI infra dealmaking: Poolside $12B reverse-execuhire to NVIDIA, Stripe buys OpenRouterneeds verification

Latent.Space reports NVIDIA's $12B reverse-execuhire of Poolside with a 7GW neocloud buildout, and Stripe's reported $7B acquisition of OpenRouter, alongside NVIDIA pushing customers to build their own models. These reflect consolidation and vertical integration across AI compute, routing, and distribution.

Major infra/business moves reshaping the model and routing landscape, but secondary newsletter sourcing.

4
#10

Enterprises report large productivity gains from ChatGPT Work (Stampli, NVIDIA)

Stampli reported cutting launch production hours by 68% using Codex and ChatGPT Work, while NVIDIA teams described scaling workflows and reducing manual tasks. Both are OpenAI-published customer case studies.

Real enterprise adoption stories with numbers, but vendor-authored case studies.

Sources scanned · 17/20 live