DailyWeights & Takes

3 Nov 2026 · own the stack, cut the bill

Since the previous edition, 2 Nov 10:57 ET to 3 Nov 2026 13:09 ET

Edition 15 · 5 news · no releases · 4 repos · 5 papers

At a glance

News 5

  1. Trending Topics

    Revolut moves about 2,000 engineers onto self-hosted GLM for coding

    In a talk reported on 3 Nov, Revolut's head of AI infrastructure described running Z.ai's open-weight GLM (GLM-5.3, per Analytics India Magazine) on about 56 Nvidia B200s for roughly 30 cents per million tokens, and putting the bill at about a tenth of what it paid Anthropic and OpenAI. Revolut also reports time to first token falling from over 10 seconds to 1.6, with context capped at 256K and image work still sent to Gemini. The productivity figures (60% faster coding, 70% shorter lead time) are its own. A regulated bank has made an open-weight model its default coding agent, so the self-hosting option now has a public cost figure to compare against.

    174 ptsopen weightsreporteddiscussion

  2. OpenAI

    Scaling Git at OpenAI: Pando, a Git backend built for agent traffic

    OpenAI says Codex agents pushed commits to its monorepo so fast that it had to replace the system behind Git. Engineers and agents still use the same Git commands, but OpenAI reports about 80% lower median push latency and close to a petabyte of Git data served a day. It also sent several fixes upstream to Git. Pando isn't a product. If agents write most of your commits, your source control becomes a scaling problem long before your model bill does.

    257 ptsagent infrastructurediscussion

  3. Hacker News

    Subvocal: glasses that turn silently mouthed words into text

    Antennas in the frame send radio signals through your face, and a model trained on many wearers decodes the mouth movements into text after about 20 calibration sentences. The founders dropped EMG electrodes as impractical. Decoding runs in the cloud for now, at under 500 ms by their own figures. Preorders are open for 500 units shipping in Q1 2027, with an SDK promised later. Silent dictation would let you talk to a coding agent in an open-plan office, though nothing ships until next year.

    797 pts373 commentshardware

  4. Momentic

    Momentic's Mo, an agent that QA-tests your app, is out of beta

    Mo explores web, iOS and Android apps in cloud sessions, reports bugs with steps to reproduce them, and turns its runs into regression tests that run on each pull request. You can start a session from a coding agent, a PR comment or a qa CLI. There's a free start; pricing beyond that is through sales. Coding agents write code faster than anyone reviews it, so tools that check the result from the user's side are filling that gap.

    158 ptstesting agentsdiscussion

  5. Sean Goedecke

    LLMs are not superhuman at reading and writing neural network weights

    A rebuttal to a recent paper that credited GPT-6 Astra with reading and editing model weights better than humans. Goedecke shows that both headline tasks fall to standard methods: a norm over first-layer weights reveals which game a small network learned, and stopping a 0.8B Qwen model from looping is ordinary steering-style pruning. Treat "the model edits its own weights" claims as an impressive trick, not evidence of a new research capability.

    129 ptsanalysisdiscussion

No model releases in this window: nothing from the labs on our watch list shipped weights or an API model since the last edition.

GitHub repos 4

  1. GitHub

    Balanov/skein

    Runs graphs of local commands, such as a swarm of agent jobs, with bounded concurrency and an emergency stop that actually holds: queued work is invalidated and late results from the stopped run are rejected rather than merged. JSON manifests, no implicit shell, a report and event log per run. Not a sandbox. Go, MIT, version 0.2.

    259 ★new, created 2 Nov

  2. GitHub

    phone-harness/mcp

    An MCP server that gives an agent a phone to drive: screenshots, on-screen elements, taps, swipes, typing and app launches. The server is MIT, but the phones are a paid cloud service (Android emulators by the minute, dedicated iPhones by the month), so it's a client for that service rather than something you run on your own devices.

    +443 ★ in window≈449/day · was ~0/day (breakout)

  3. GitHub

    Muvon/octomind

    An open-source coding-agent harness in one Rust binary, launched on HN on 3 Nov. It compresses long sessions, has a supervisor model condense tool output, and enforces per-request and per-session budget caps. Its own benchmark puts it at about half opencode's cost on the same model. Apache-2.0.

    369 ★launch post 3 Nov · 42 ptsdiscussion

  4. GitHub

    kgoedecke/doop

    A self-hostable, multiplayer design canvas where designs are HTML frames and agents edit them live through a built-in MCP endpoint, alongside people. Connects to Claude Code or Codex with one command. AGPL-3.0.

    +106 ★ in window≈95/day · was ~21/day (breakout)

Four this time. The window's biggest breakouts had already run here in the past week, and another is a front end for a paid model key, so we skipped it.

Papers 5

  1. arXiv

    MEMENTO: an 8B agent trained to manage its own context

    Qwen3-8B is taught to compact its history into notes, first from synthetic examples and then with RL. On question answering over books of up to 2.5M tokens, it goes from 36.9% to 67.8% accuracy while never holding more than 20K tokens. That beats a GPT-5.5 agent in the same harness (13.1%) and full-book context (56.7%) at under 1/50 the cost, and nearly matches a GLM-5.3 agent about 100 times its size. Context management is a skill you can train into a small model, not just a prompt.

    new on arXivlong context

  2. Hugging Face

    CUA-Master: a 27B computer-use agent at 80.2 on OSWorld Verified

    The training data comes from real workflows mined from instructional videos, turned into reusable pieces for generating new, checkable tasks, with a rollback-and-retry loop for collecting runs in real environments. Each round targets the model's current weaknesses. The 27B model also scores 36.9 on OSWorld 2.0. The gain comes from where the training data comes from, not from a bigger model.

    90 upvotescomputer use

  3. Hugging Face

    CrossSWE: coding tasks that span several repositories

    171 real tasks across 165 repositories and 13 languages, each checked so that changing only one repo can't pass. The best of five models, GPT-6.1 Sol, passes 58.48% on the first try, and the most common failure (44.2%) is leaving the repos inconsistent with each other. If your change crosses a service boundary, check the other side yourself.

    19 upvotesbenchmark

  4. arXiv

    Running 4-bit MXFP4 models as plain INT8 math on Apple's M5

    Almost every row of real MXFP4 weights (97.9-99.97% across seven models) can be rewritten exactly as one INT8 vector and one scale, so a layer becomes a single INT8 matrix multiply. On an M5 the authors report 1.10-1.74x faster prefill than MLX with outputs bit-identical to the quantized reference, and the files stay standard MXFP4. Faster prompt processing for local models on Macs, with no new file format.

    new on arXivlocal inference

  5. arXiv

    Hacking or just failing? Telling reward hacking from honest mistakes

    In a sandbox where an agent can game its score, small local monitors couldn't separate exploits from honest failures, a stronger model mostly could, and both missed an exploit no prompt named. A two-rule deterministic check caught every known exploit. Small study, pre-registered, with the environment and logs released. Before you put an LLM judge on your agent's score, check whether a plain rule does the job.

    new on arXivagent evals

Things to try

Put a spending cap on a coding agent. Octomind is a single-binary agent harness that compresses long sessions and stops at per-request and per-session budgets you set in TOML, with any provider key (OpenRouter included).

First step: install it (brew install muvon/tap/octomind, cargo install octomind, or the install script in the README), export a provider key such as OPENROUTER_API_KEY, then in a repo run octomind run developer:general. Its cost comparison with opencode is the project's own, so check it against a task you already know the bill for.

The take

The teams spending most on coding agents are cutting the bill by owning the layer under them, not by buying better models.

Revolut moved about 2,000 engineers onto self-hosted GLM at roughly 30 cents per million tokens and reports a tenth of its old bill. OpenAI rebuilt its Git backend because Codex pushes outgrew it. MEMENTO's 8B agent beats a GPT-5.5 agent on book-length questions at under 1/50 the cost by managing its own notes, and Octomind claims half opencode's spend on the same model. The frontier model is becoming one part you swap, and the savings come from everything around it. Prediction: by 30 Jun 2027, another company with more than 1,000 engineers will publicly say a self-hosted open-weight model is its default for coding agents.