Good Machines

A good machine makes room for what humans do best. Frictionless AI that helps with everyday life, then steps back. Not the kind that begs for your attention.

[Good Machines @goodmachinesai]

The numbers here come from open-weight 7B to 70B models, greedy decoding, one run each. Retry hints were tested only on read-only and benchmark tools. A retry on a tool that writes or charges can do it twice. A preprint, not yet peer reviewed.

[Good Machines @goodmachinesai]

We read a paper on what an agent does when a tool call fails. In its scripted study (60 cases), Qwen3-8B, an open model, retried a temporary failure 0% of the time with the error left in the tool result, and 100% when it was restated as a short runtime note.

Paper below.

[Good Machines @goodmachinesai]

The note version also removed the failed call. The authors' version keeps it and adds the note plus one sentence telling it to retry, fix or stop, matched to the error. It raised ToolMaze pass rates, mostly on temporary failures, with no gain on a second benchmark.

[Good Machines @goodmachinesai]

We shipped Mitamaton 1.1.1 with a new welcome. It shows you a sample day before you connect anything. Then you connect your calendar, email, reminders and health, and pick your Tama.

[Good Machines @goodmachinesai]

Good Machines is a member of the NVIDIA Inception program, which supports startups building with AI. Thanks to @nvidia for having us.

#NVIDIAInception

[Good Machines @goodmachinesai]

We joined the Cartesia Startups grant program. Thanks to @cartesia for the support.

#BuildingWithCartesia

[Good Machines @goodmachinesai]

This scene in Cyber core, our poster tool, is called Chrome liquid. It's a coil, a few blobs and some strings of beads, all in chrome.

We kept the shapes the same and changed the backdrop, light and colors on the beat. Every frame came out of the tool.

https://t.co/dmwxfPXcpm

[Good Machines @goodmachinesai]

@jinheonbaek Caveats: three trials each, no significance test on the 3.2-point gap, though the linked, check-first version led on all 8 suites. Links and instructions changed together, so the paper can't say which did more. Text games and simulated apps. A preprint, not yet peer reviewed.

[Good Machines @goodmachinesai]

We read a paper on AI agents that learn from saved lessons, which can mislead them. Average score across 8 test suites: 79.5 with no memory, 87.0 with lessons, 90.2 when each lesson linked to the run it came from and the agent was told to check it.

Paper in the replies.

[Good Machines @goodmachinesai]

Don't Just Trust Your Memory: Evidence-Linked Memory Reduces Memory-Induced Degradation in LLM Agents, by Yukyeong Lee, Yumin Choi, Jinheon Baek (@jinheonbaek) and Sung Ju Hwang (KAIST, https://t.co/kqOoOSZmzK). CC BY 4.0.

https://t.co/ssxPgnsjZM

[Good Machines @goodmachinesai]

Anthropic is going public. Its IPO page says revenue was $17.0B last quarter, over 12x a year before, and revenue per dollar spent on compute and infrastructure is up more than 4x since Q1 2025. Trading on Nasdaq is expected November 19. Not investment advice. @AnthropicAI

[Good Machines @goodmachinesai]

Three AI stories from today, with our read on each. AI agents wrote all of AWS's new local emulator. Google's Gemini Decisions API picks answers instead of writing them. MiniMax announced H3-Image, a 4K image model with open weights. @awscloud @Google @MiniMax_AI

[Good Machines @goodmachinesai]

@ao_qu18465 @mitmobility Caveats: each model ran once, and three of the four were Flash versions. With real people answering, on 22 smaller tasks, the one model tested there had a first plan that couldn't be carried out 32.9% of the time (48.5% with a default AI operator). Version 1, ongoing work.

[Good Machines @goodmachinesai]

We read a paper on AI agents that must ask for missing facts before they plan. In 43 tasks, the simulated operator only answered questions about specific items. Across four models, one run each, first plans couldn't be carried out on 58% to 98% of tasks.

Paper in the replies.

[Good Machines @goodmachinesai]

From Questions to Decisions: Benchmarking How LLM Agents Acquire Decision-Relevant Information from Human Beings, by Yuhan Tang, Ao Qu (@ao_qu18465), Jinhua Zhao (lab: @mitmobility) and 8 co-authors.

https://t.co/ilTUNwMr08

[Good Machines @goodmachinesai]

Three AI stories from today, all from Google, with our read on each. Gemini 4 Carbon is out for developers. AI agents proved key security properties of Signal's code. Gemini Robotics 2.5 packed car parts with 96% success in a pilot. @Google @GoogleDeepMind

[Good Machines @goodmachinesai]

Family Farmstead Dairy Announces Recall of All Cheddar Cheese Products Because of Possible Health Risk

Lot codes dated: 10/16/27 and prior · Net weight: 8oz

Official notice from FDA, 11.10.26: https://t.co/UXvdvGaB50

[Good Machines @goodmachinesai]

We want to build Earth, Again, an accurate 3D Earth made in the open. Each person builds a small piece, like a wheel or a house, with their own AI.

We're gauging interest before we pay for the infrastructure. If you want to help, join the waitlist: https://t.co/f6cTG1jR6s

[Good Machines @goodmachinesai]

One more scene from Cyber core, our poster tool. This one is Cube city, a long strip of tiny buildings seen from above.

Same idea as the others. The city stays put and everything around it changes on the beat. Every frame came out of the tool.

https://t.co/dmwxfPXcpm