Edhivo studies AI assessment with teacher review
Edhivo uses language models to draft assessments from course material and evaluate responses against rubrics, while retaining teacher review.
Loading content
Research, releases and useful new ideas.
Edhivo uses language models to draft assessments from course material and evaluate responses against rubrics, while retaining teacher review.
MemT 1.0 studies language models that update a compact internal memory as they interact, reducing dependence on a continually growing transcript.
Apriel Voice is a public preview model that accepts speech, text and images, then produces text or tool calls.
Arken is an experimental Rust authentication core that asks an agent to sign a one-time challenge to prove possession of a cryptographic key.
EpiSEM-Bench tests whether an agent seeks a missing fact before acting, using paired tasks where the same request is permitted in one world and prohibited in the other.
REM studies when a coding agent should receive compact notes and detailed records from earlier tasks.
Cursor’s October 29 notice identifies malicious Anthropic.claude-code versions 2.1.312 and 2.1.313 distributed through its extension marketplace.
Agent4Sci evaluates agents on scientific machine-learning tasks with separate development, held-out evaluation and audit stages.
Rufous records tool activity and checks file state so a local assistant can be challenged when it claims unfinished work is complete.
Microsoft’s October 21 report describes Storm-0940 using the ARTEX agent framework to accelerate attacks involving compromised cloud application credentials.
Dates show when we added or updated a story. Routine source checks stay in its history. Release entries use the source publication date. Times use your device’s time zone.
Daily archive →Important changes in AI, with context and follow-ups.
· Claude users, developers, researchers and businesses tracking the institutions behind frontier AI.
Anthropic published its registration statement for a proposed public share offering.
Why it matters. The proposed transition makes funding and governance important context for the AI community following this major model provider.
Watch amendments, final terms and whether the proposed listing proceeds; record each material follow-up on the same subject.
Still uncertain. Final offering terms and completion remain uncertain; the registration document is preliminary.
Original source ↗· Developers and security teams using LiteLLM or managing model-provider credentials.
The October 21 distribution compromise now has an October 23 corrected investigation identifying an exposed master key and chained maintainer-token theft.
Why it matters. AI gateways handle valuable provider credentials. This incident illustrates why release-artifact provenance and credential containment matter alongside source review.
Watch for independent confirmation, the final exposure scope and evidence that credential and release-approval changes prevent recurrence.
Still uncertain. The canonical report supersedes the earlier logged-response hypothesis. Independent reports support malicious-distribution findings but do not independently verify every step of the corrected chain.
Original source ↗· Security teams, agent developers and organizations running evaluations.
India’s incident-response agency published findings on agent-driven intrusions.
Why it matters. The findings make evaluation isolation and credential exposure immediate operational concerns.
Watch for incident analyses and independently evaluated safeguards.
Still uncertain. Public details are insufficient to quantify prevalence.
Original source ↗Research reviews new developments across science, models, tools, creative work, everyday uses and business. We select events with primary evidence, substantial consequences and a clear explanation of who is affected. Longer-term implications are editorial assessments and may change as evidence develops.
Events remain eligible for 30 days. Each has a source review due within fourteen days; overdue events leave this section while their history stays available. Research publishes updates on the existing schedule. Social activity is assessed separately in What’s Viral.