Agentic AI & LLM Weekly
2026-W38 — 11 September – 18 September 2026
Editor’s note, 15 October 2026: this issue went out before the newsletter’s current sourcing rules were enforced, and some of its links wouldn’t pass them today. It stays as published; every issue from 25 September 2026 on passes them.
Models were pointed at the infrastructure beneath them this week — serving stacks, query planners, compilers.
Editor’s Picks
The through-line this week is inward: models are being aimed at the systems that run them rather than at the users in front of them. Z.ai says GLM-5.3 did most of the work building the inference stack that now serves GLM-5.3-Flash on more than 100,000 Chinese-made accelerators, tripling throughput in under a fortnight. A solo researcher spent $1,200 training a 4B model that produces Postgres query plans 81% faster than Postgres’s own optimiser. And Bend 2 inverts the usual arrangement between developer and coding agent: you write the laws, the compiler demands a proof, and the agent’s mistakes simply don’t compile.
Community Pulse
What the AI community is talking about this week
A Language Where the AI’s Mistakes Don’t Compile — and 295 Comments Asking Whether That’s Easier
[Community]
Bend 2 asks developers to formalise their application’s rules in a LAWS.bend file, then has the compiler demand a mathematical proof on every edit, so code that violates those laws is rejected outright. It type-checks in under a second where Lean or Rocq take minutes, which is the difference between a proof system an agent can run after every change and one it can’t. The thread’s sharpest objection is the obvious one: is writing a correct formal specification actually easier than writing the code, and what happens when the laws are underspecified and the agent finds a solution that satisfies them literally?
[Source: Hacker News]
Research Highlights
Papers and findings worth your time
Let the Search History Become the Simulator: Recursive Self-Improvement Without the Bill
[Research]
Dream-RSI treats an agent’s accumulated discovery tree as a replay simulator over the space it has already searched, then “dreams” inside that simulator to evaluate and refine its exploration policy without paying for fresh online evaluations. The improved policy goes back online, expands the tree, and enlarges the simulator — a self-improving loop whose cost curve bends the right way. Across algorithm engineering, mathematical optimisation and GPU kernel work the team reports competitive or better discovery quality at substantially lower cost, which matters because expensive online evaluation is what makes most self-improvement schemes academic.
[Source: arXiv]
What If the Weights Weren’t Frozen After Training?
[Research]
Scaling laws and MoE both assume a model finishes learning at the end of pretraining, but a deployed model meets most of its useful data in the live interaction it’s currently handling — and can’t touch it, because the weights are fixed. This paper asks what an architecture would look like if it generated and adapted weights from that live data instead, writing the interaction into the parameters rather than stuffing it into a context window. It’s a direct challenge to the assumption that long context is the answer to personalisation and adaptation.
[Source: arXiv]
Engineering & Technical Blogs
What builders are shipping and writing
GLM-5.3 Built the Inference Stack That Now Serves GLM-5.3-Flash
[Tool]
Z.ai published a detailed account of standing up production inference for GLM-5.3-Flash on a cluster of more than 100,000 Chinese-made accelerators — limited on-chip memory and bandwidth, incomplete kernel support, a 1M-token context window and multimodal traffic, on hardware nobody had operated at that scale. Much of the work was done by an Infra Agent powered by GLM-5.3 itself: first successful run to production readiness in under two weeks, with end-to-end throughput tripling against the initial baseline and per-token cost reaching parity with mainstream NVIDIA GPUs. Z.ai is careful to say this isn’t recursive self-improvement — choosing objectives, setting boundaries and assessing risk stayed human — but “the model optimises the system; the system runs the model” is a fair description of what happened.
[Source: Z.ai]
$1,200 and a 4B Model Beat the Postgres Query Planner by 81%
[Tool]
QORL fine-tunes a 4B model to emit pg_hint_plan hints inside a custom agent harness, steering Postgres’s execution strategy rather than replacing the optimiser. On 113 join-heavy queries from the Join Order Benchmark it reports a 1.81× geometric-mean speedup and a 44.7% cut in total latency, trained with supervised fine-tuning on frontier-model trajectories plus an RL variant, for about $1,200 all in. The interesting part isn’t the number but the loop: propose a plan, run it for real, learn from the measured latency — now cheap enough for one person over a few weeks.
[Source: Rohan Bansal]
Product & Company News
Model releases, funding, and notable moves
A Ternary 27B That Keeps 98% of Its Full-Precision Scores in 5.9 GB
[Industry]
PrismML released Bonsai 2 27B, built on Qwen3.8 27B and shipped in a ternary form that cuts the memory footprint more than ninefold to 5.9 GB. The generational claim is about the loss, not the size: the previous Ternary Bonsai 27B retained 95% of its full-precision counterpart’s aggregate benchmark performance, and this one retains over 98%. Apache 2.0, downloadable from 17 September — which puts a genuinely capable reasoning, coding, vision and agentic model inside the memory budget of consumer hardware.
[Source: PrismML]
Mozilla Picks Mistral to Power Firefox’s Browsing Assistant
[Industry]
Firefox’s Smart Window beta is now powered by Mistral models, with Mistral Small 4 arriving for users in the US and Canada and the beta expanding to France — its first market beyond North America — before the UK and Germany later this year. Both companies frame multilingual and multicultural tuning as a model feature rather than a localisation layer bolted onto an English-first assistant. For anyone weighing a browser-resident agent, this is the clearest signal yet that the privacy-and-choice pitch is being made at the model layer, not just the UI.
[Source: Mistral AI]
Open Source on the Rise
AI projects gaining stars, downloads and users this week
WeKnora — Tencent’s Document-to-RAG-to-Agent Pipeline
[Tool]
Tencent’s open-source knowledge platform turns raw documents into a queryable RAG corpus and then puts an autonomous reasoning agent on top of it, in Go. It added roughly 3,980 stars this week to reach about 26,900 (GitHub Trending, weekly view, 18 September). Worth a look if you’ve been assembling that pipeline yourself out of a parser, a vector store and a framework.
[Source: GitHub]
context-mode — Sandboxing Tool Output to Reclaim the Context Window
[Tool]
A TypeScript project aimed squarely at the failure mode every coding-agent user knows: a single verbose tool call eating the window. It sandboxes tool output — claiming a 98% reduction — and persists session memory across runs, adding about 1,480 stars this week for roughly 23,500 total (GitHub Trending, weekly view, 18 September). The HarnessTax numbers below suggest why this kind of plumbing is where the cost differences actually live.
[Source: GitHub]
worktrunk — Git Worktrees Built for Running Several Agents at Once
[Tool]
A Rust CLI for managing git worktrees when the thing checking out branches is a fleet of agents rather than a person. About 1,060 stars this week takes it to roughly 8,000 (GitHub Trending, weekly view, 18 September) — small in absolute terms, but it’s the kind of tool whose existence tells you how people are really working now.
[Source: GitHub]
Hardware & Macro Watch
Chips, compute, and the infrastructure layer
NVIDIA Brings CUDA to Rust, on Two Separate Tracks
[Industry]
NVIDIA announced native GPU kernel authoring in Rust, compiling to PTX through two routes: cuda-oxide, a custom rustc codegen backend that takes #[kernel] functions through Rust MIR and LLVM IR, and cutile-rs, a tile-based system for safe idiomatic kernels. The stated motivation is that more and more of the systems layer of AI is written in Rust, and that layer has been paying a foreign-function tax to reach the GPU. NVIDIA says the three frontends — Rust, C++ and Python — will interoperate, so picking one shouldn’t lock you out of the others.
[Source: NVIDIA]
Model Evaluations & Transparency
How models are being measured, compared, and held accountable
Same Model, Same Score, Five Times the Bill: the Harness Is the Variable Nobody Benchmarks
[Eval]
Arena ran 21 model–harness pairs over 30 sampled tasks each from SWE-bench Lite and Terminal-Bench 2.0, three attempts per task, each harness in its native high-effort configuration and capped at 100 turns. Success rates barely move with the harness — Claude Fable 5 solves 97.8% in Claude Code, 96.7% in Codex CLI and 96.7% in Pi — but cost moves a great deal, with Claude Code running about twice Pi’s price on the same model and spreads reaching 5× elsewhere. If you’re choosing an agent stack on leaderboard scores alone, you’re optimising the variable that turns out not to matter much and ignoring the one that does.
[Source: Arena]
Quick Links
Worth a bookmark — no summary needed
- Astra for Law — 664 comments, the week’s largest thread, on AI in legal practice
- How to Write with an LLM — a long argument about drafting with models rather than delegating to them
- Breaking the 1.58-bit Barrier for Ternary LLMs — the quantisation research sitting underneath this week’s ternary releases
Curated by Claude Code