AI news5
Ai2 releases ToyBlocks with 85,336 robot demonstrations
Ai2's 6 November ToyBlocks release provides 85,336 long-horizon demonstrations for training robots in a controlled block world. Its hierarchical Molmo 2 and ManiFlow system completed 55.5% of 200 simulated episodes and 50% of 30 real-world episodes. The small real-world sample needs further trials, and integrated reinforcement-learning training is planned for a later release.
huggingface.co ↗Why it matters
A shared simulation, dataset and evaluation setup makes robot-learning results easier to reproduce, while exposing how much still fails outside simulation.
Humans& brings a group-conversation model to Vercel
Vercel added Humans&'s Ampersand to AI Gateway on 6 November, free through 13 November. Built on GLM 5.3, it reads group conversations and chooses whether to contribute or stay quiet. Applications execute its send-message and pass actions.
vercel.com ↗Why it matters
Useful group assistants need to manage timing and shared context, giving teams a concrete behaviour to test beyond one-person chat.
Google ADK 2.14 adds stable modules and Microsoft Teams
Google published ADK 2.14.0 on 7 November, although its changelog is dated 4 November. The Python agent framework promotes 26 modules and 16 toolset and integration features to stable, and adds a Microsoft Teams runner. Google's 30-node, 200-turn workflow test ran 23% faster with 9% lower peak memory.
github.com ↗Why it matters
Stable interfaces and a Teams connection reduce integration work, while the performance figures give developers a specific workload to compare against.
Nous releases Arete-27B with a political-framing evaluation
Nous Research announced Arete-27B on 6 November, post-training Qwen3.8-27B on roughly 53 million tokens. On its own 803-prompt evaluation, the lab's defined propaganda metric fell from 22% to 0.9% with thinking disabled. This is a vendor-designed evaluation, rather than independent proof that the model is unbiased.
nousresearch.com ↗Why it matters
Fine-tuning changes whose values a model reflects, so independent testing still matters before deploying it as a fact-finding tool.
Claude Code adds clearer failure behaviour for unattended runs
Anthropic's 6 November Claude Code 2.1.319 release adds a setting that denies tool calls instead of waiting for permission prompts. It also fixes non-interactive auto-mode runs that reported a successful exit after actions were refused because no safety check was available. Those text and JSON runs now exit with code 1.
github.com ↗Why it matters
Beyond background tool execution, these permission changes let teams distinguish refused automation from a successful run.
AI in the nonprofit sector4
Rethink Ireland's AI accelerator closes today
Rethink Ireland's AI for Good accelerator closes applications on 9 November for up to 30 not-for-profit organisations, with grants of up to €15,000 each. The 12-month programme supports responsible AI adoption in 2027. Applicants must be based on the island of Ireland, make their main impact in the Republic, and commit at least two staff including a senior leader.
rethinkireland.ie ↗Why it matters
Today's deadline, rather than the earlier announcement, makes this an immediate application check for eligible organisations. The page gives no cutoff time or timezone.
Hacker Initiative's US$5,000–25,000 grants reach deadline day
Hacker Initiative's Signal/Noise round supports work including AI-generated-content detection, provenance and media literacy, with grants of US$5,000–25,000. Applications close at 6:59 pm AEDT on Tuesday 10 November, equivalent to 9 November at 11:59 pm PST. Individuals, student groups and small organisations can apply subject to legal eligibility, but salaries and staff training are excluded.
hackerinitiative.org ↗Why it matters
The US deadline falls on today's date, turning the previously announced round into a final submission reminder for eligible projects.
Tarbell offers paid AI-journalism fellowships closing 22 November
Tarbell Center's 2027–28 fellowship accepts applications until 22 November for a full-time June-to-May programme combining three months of training with a nine-month newsroom placement. Standard salaries range from US$75,000 to US$120,000. It primarily targets applicants with work rights in the US, UK or Hong Kong, and visa sponsorship is uncommon.
tarbellcenter.org ↗Why it matters
This individual fellowship is separate from Tarbell's previously covered institutional grants, adding a paid route into specialist AI reporting.
AIU offers up to KES250,000 for AI and climate research
Africa International University's internal research call includes AI, data science and remote sensing for climate resilience among its eligible themes. Four awards of up to KES250,000 are expected, with a deadline of 7:59 am AEDT on 16 November, equivalent to 15 November at 11:59 pm EAT. Each team needs two or three members and a full-time AIU academic or research staff member as principal investigator.
aiu.ac.ke ↗Why it matters
The AI-and-climate theme creates a specific funding route, but outside organisations need an eligible university-led collaboration rather than applying independently.