Bijan Bowen runs Mistral Large 4 through a browser OS, games and a watch site. His reported bill: about $47. The results are mixed. The launch-day benchmark charts did not show that part.

Meanwhile, a tiny decision pipeline finds a Kubernetes fault without a chatty agent. A 98-star app keeps Claude working after the laptop shuts. One developer notices his arms hurt less.

Big models make headlines. Small changes make a working day.

🐈 $47 to meet le Chonk

Bijan Bowen tests a browser OS and several games through Mistral Vibe. The OS can install little apps inside itself; the subway shooter initially fails to fire. His transcript ends with a roughly $47 total and no single result that blew him away. This is his test, not ours.

Mistral Large 4 is a multimodal preview API: Mistral says 1T parameters, 49B active. Weights are promised by the end of October, not available for self-hosting today. Its coding and security benchmark results are vendor claims.

Mistral announcement  ·  Bowen test + transcript  ·  HN discussion

🔧 AI designs the thing running AI

The openTPU author shows LFM2.5-230M chatting on a Kintex-7 PCIe card with a utilization monitor beside it. The README reports 82.1 tokens/s wall-clock for one 4-bit configuration and says the card matches its simulator token for token. Those are the author's measurements, not an independent hardware test.

openTPU puts SystemVerilog RTL, an ISA simulator, compiler and host software in one Apache-2.0 repo, created September 24. The claim is AI-assisted accelerator design running on an FPGA. It is not a newly fabricated chip.

README + measurements  ·  HN discussion

🧲 One search box, four senses

Google's AI Edge team ships two demos: search local media and find a moment in a video. Instead of captioning every frame and transcribing every clip first, the model embeds the different inputs in a shared space. These are vendor demos; we have not run them.

EmbeddingGemma 2 is Apache-2.0 and modular: Google's developer guide lists 270M parameters for text/code and 740M for all modalities, with 768-dimensional embeddings. It is a retrieval model, not a text generator.

Announcement  ·  Developer guide  ·  HN discussion

🚨 The pod says 16Mi. YAML says 256Mi.

Yiming Su and coauthors investigate a SREGym-Lite fault: nginx-thrift pods keep running out of memory. The Deployment says 256Mi; the live Pod has 16Mi. Their collector spots the mismatch and a matching admission webhook. Jev picks the component and supporting evidence. No LLM agent writes commands.

The authors' Jev-driven diagnosis pipeline passes 80 of 105 attempts across 21 faults, with a 14.6-second median. All five attempts fail on each of five faults. A GPT-6-Astra rubric judges the diagnoses; this is a lab cohort, not production incident resolution.

Study + failure analysis  ·  HN thread

⚓ Close the laptop. Keep the agent.

Sean Brydon's Berth demo starts a task by choosing a project, a box and Claude Code. It creates a worktree on that box and shows the agent's edits as diffs. The agent runs there, so closing the laptop need not kill the task. This is the project's demo, not our live run.

Berth is MIT-licensed: a Mac app, a CLI and a daemon for your own Linux or Mac boxes. The install docs require git, tmux and normally a Tailscale address. An open issue asks for Linear integration; do not assume your team workflow is already covered.

Project  ·  Repo + issues  ·  Install + side effects

🎹 Less code. Less wrist gymnastics.

Vaughan Hilts sits down at the piano and notices little of his old arm pain. Mechanical refactors and feature work increasingly become prompts plus manual refinement. He types more English and fewer modifier-key contortions. He also has more meetings and a more senior role, so he does not claim the model caused it.

A first-person account of changing the physical shape of coding work. Not a clinical study, not a cure, and not evidence to ignore pain. The useful detail is which keyboard motions disappeared.

Hilts' account  ·  HN discussion

A 300ms benchmark. matklad sizes microbenchmarks to about 300ms: long enough to see, short enough to repeat. His rule of thumb, not a universal measurement standard. Read →

Backprop is not replaced. Dust explores activation perturbations for transformer pretraining. Its authors explicitly do not claim enough compute efficiency to replace backprop today. Paper →