AI news5
Dyna's Taku completes a one-hour laundry demonstration
Dyna Robotics introduced Taku and its Dyna-2.1 system on 29 September, showing an uncut, autonomous laundry workflow lasting one hour. The robot loads machines, folds towels and puts them away, with a vision-language model coordinating the work. Dyna's technical report identifies real customer sites as its next milestone.
dyna.co ↗Why it matters
Completing a connected workflow is a stronger test than folding a single towel, although commercial reliability and labour savings remain unproven.
Ai2 releases AstaBrief 8B for faster cited research reports
Ai2 released open weights and training data for AstaBrief on 2 October. Its eight-billion-parameter model powers Asta's Fast mode, generating reports from research questions and retrieved literature excerpts. Ai2 reports average full-pipeline times of 51.1 seconds versus 178.5 seconds for its Claude-powered Thinking mode.
huggingface.co ↗Why it matters
The newly released scientific-report weights give institutions a way to explore local processing of sensitive or unpublished research.
Tavus previews Griffin-Lite for live audiovisual conversations
Tavus introduced Griffin on 1 October, combining simultaneous listening, visual perception and generated speech and video. The company says 48% of participants in one-minute calls mistook Griffin-Lite for a human. Access is limited to selected early testers, with a wider release planned.
tavus.io ↗Why it matters
More convincing live avatars make clear disclosure essential when organisations test digital tutors or service assistants.
Inception makes Mercury Voice available to enterprise customers
Inception's 29 September launch moves Mercury Voice from preview to enterprise availability through its API, with access arranged through sales. On its customer-service test prompts, the company reports 320 milliseconds median and 750 milliseconds at the 95th percentile to the first answer token. These measure the model stage, rather than the complete wait a caller hears.
inceptionlabs.ai ↗Why it matters
This new voice-focused release follows earlier Mercury coverage and gives phone-agent builders a latency claim they can test within their full speech pipeline.
Google DeepMind tests watermarks in AI-designed proteins
Google DeepMind introduced SynthID Bio on 30 September as a proof of concept for identifying synthetic biological designs. In laboratory tests across three protein targets, it reports that watermarked binders retained comparable binding performance to unwatermarked designs. Resistance to deliberate watermark tampering remains a research challenge.
deepmind.google ↗Why it matters
This new wet-lab provenance result extends beyond earlier DeepMind policy coverage and could help distinguish AI-generated material in biological databases.
AI in the nonprofit sector4
Delaware funds a year of AI capacity for 20 nonprofits
Delaware Community Foundation and Tapp Network opened applications on 1 October for 20 Delaware 501(c)(3) organisations. Selected nonprofits receive 12 funded months of TappHQ, configured workflows, training and five hours of individual implementation support; no cash grant amount is published. Applications close 31 October, and the mandatory kickoff is 9 November.
delcf.org ↗Why it matters
Funding implementation and an AI use policy alongside software addresses the staff capacity needed to sustain adoption after a grant ends.
CoThinkk offers 20 AI grants of $2,500, with conflicting deadline dates
CoThinkk's current call supports leaders of colour in Western North Carolina exploring AI and emerging technology for community work. Its linked application confirms 20 grants of US$2,500 and requires 501(c)(3) status or a fiscal sponsor. The grant page lists both 8 and 18 October 2026 deadlines, while the application gives no date; applicants should confirm which applies.
cothinkk.org ↗Why it matters
This is a small, community-led funding route, with a deadline uncertainty to resolve before relying on it.
Jefferson Trust awards $260,000 for planned tumour-growth research
Jefferson Trust announced US$812,064 for six University of Virginia projects on 29 September. The awards include US$260,000 to develop GrowthNet, an AI model intended to forecast vestibular schwannoma tumour growth in 3D from a patient's current MRI. This is a funded development project, with no clinical performance result reported.
jeffersontrust.org ↗Why it matters
The award shows philanthropy supporting a specific AI research tool before evidence of clinical usefulness exists.
Trusting News receives a $60,000 grant for AI literacy and tools
Trusting News announced on 1 October that it received an unrestricted US$60,000 grant from OpenAI Foundation's People-First AI Fund. It will support newsroom AI literacy and exploration of tools that make its ten years of trust research easier to use. The organisation already provides AI policy, disclosure and community-feedback resources.
trustingnews.org ↗Why it matters
This newly named recipient and award amount show how the previously covered fund is reaching practical newsroom work, rather than reopening its application call.