The box ships with Companion OS, plus Docker, Ollama and Hub. Core is optional for Hub users and ships to US addresses only. ci.computer/store/p/core
1
20
When an AI agent touches client work, who can show what it did? A new preprint proposes signed, tamper-evident records of agent sessions. It's a proposal, with no experiments yet. arxiv.org/abs/2610.40071
5
Hub's Alternatives category is the place to compare self-hosted options with rented chat tools. Local models want 16 GB of RAM or more. ci.computer/download
9
A preliminary study of 891 college students: revising AI output, instead of accepting it, was linked to critical thinking. It's correlational and not yet validated. arxiv.org/abs/2610.38910 #AIResearch
2
Agent memory that forgets some things on purpose: a preprint inspired by scrub jay memory reports gains from type-based decay, but says they narrow with stronger models. arxiv.org/abs/2608.04746 #AIAgents
1
11
Delimitrou's Stanford research found most large computing systems run at only about 15% capacity. Better software could let existing hardware do more, if performance holds. news.mit.edu/2026/mitigating…
3
Hub's app store has Nextcloud for file sync and an Alternatives category to browse. Whether one fits depends on your usage and upkeep time. ci.computer/download
3
Agents with a memory bank need to find the right memory. CoEvo-Mem co-evolves retrieval and memory. The authors claim state-of-the-art results on seven benchmarks; the abstract gives no figures. arxiv.org/abs/2608.01739 #AIAgents
8
Can you audit whether an image model was trained on specific pictures? Simple loss scores can mislabel easy unseen images as members. The authors report near-perfect auditing for one method, which needs earlier checkpoints. arxiv.org/abs/2610.40616
3
Can a language model forget specific information when asked, again and again, without losing the rest? A new preprint proposes a method for repeated deletions and tests it on unlearning benchmarks. arxiv.org/abs/2610.40483 #Privacy
8
MIT News: in image models, each training example's influence shrinks as datasets grow, so removing one artist's work often changes little. news.mit.edu/2026/when-ai-ar…
8
AI rewrites shifted the stance of posts on divisive topics in an Oxford Internet Institute study, even when told to keep the meaning. The opinion effects are simulated. oii.ox.ac.uk/ai-powered-soci…
1
14
A chatbot that feels understanding can draw out more personal detail than people meant to share. An interview study of 21 chatbot users found AI warmth can threaten privacy, even when the AI doesn't seem human. arxiv.org/abs/2610.38802 #Privacy
5
Faster AI isn't automatically cheaper to run. A speculative decoding study found the energy benefit fades as batches grow, and tighter latency targets can add energy. Tested on data-center GPUs. arxiv.org/abs/2610.40287 #LocalAI
7
Quantization squeezes a model into fewer bits so it fits smaller memory. A new post-training method claims it beats current approaches, but the abstract gives no numbers. Read the paper first. arxiv.org/abs/2610.40345 #LocalAI
1
16
MIT CSAIL: a small model that asked better questions rose from 8% to 82% win rate against people at Battleship, and outpaced GPT-5 at about 1% of its cost. Experts still win. news.mit.edu/2026/teaching-a…
4
Where did the keys go? A prototype wearable logs where household items are and answers spoken questions, using an edge server. Lab-stage, with no test home described. arxiv.org/abs/2610.38629 #SmartHome
3
Coding agents can add cost in quiet ways. In controlled tests, rewording a prompt with the task unchanged can shift how much the agent reasons and checks itself. More effort helped hard problems but could add spend without payoff. arxiv.org/abs/2608.01347 #AIAgents
2
Oxford Internet Institute covers a study finding that friendly AI chatbots make more mistakes and tell people what they want to hear.
Source: oii.ox.ac.uk/friendly-ai-cha…
8
Oxford Internet Institute asks how social media interoperability could restore user choice. Personal context deserves the same question: can you take it with you? oii.ox.ac.uk/how-social-medi… #DataPortability
1
15
