Open-weight models surge as safety failures mount
Aleph Alpha released Kolibri, a 78B open-weight model, on 3 October, while OpenAI reported its agents accessed over 100 organizations.
AI & models3 October 20265 min readSources 8
AI & models
360 texts · 3/20
UK AI Safety Institute resumes model testing after August disruption
The UK AI Safety Institute has resumed most of its model evaluation activities following an incident in August, according to reporting from Resultsense on 2 October 2026.
AI & models3 October 20267 min readSources 2
Meta publishes six AI-assisted math papers as labs race to automate discovery
Meta shared six papers generated with its Muse Spark model on 3 October, marking a shift toward AI solving open scientific problems without human-provided answer keys.
AI & models3 October 20263 min readSources 10
Kitsuno runs open Laya model in shadow next to closed Jev API
Kitsuno is testing an open-weight decision model, Laya, in production shadow mode to replace its closed proprietary stack, citing sovereignty and cost control.
AI & models3 October 20264 min readSources 10
New benchmarks show AI models still lag human experts in research reasoning
A new benchmark released on 3 October reveals that agentic search tools perform no better than standard embedding retrieval at finding scientific papers that inspire new research, with top models reaching only 0.51 Recall@20.
AI & models3 October 20264 min readSources 15
WMO flags AI limitations in hurricane intensity forecasting
A Washington Post column published on 2 October details why AI models struggle to predict hurricane intensity, a challenge highlighted by the World Meteorological Organization on 2 October. Rice University is using a NASA AI model to refine risk estimates, while Google's WeatherNext 3 continues to set accuracy benchmarks.
AI & models3 October 20264 min readSources 14
Latest
12 texts
Aleph Alpha releases Kolibri, a 78B open-weight model for sovereign use
On 3 October, Aleph Alpha released Kolibri, a 78B-parameter open-weight model designed for regulated sectors in Europe.
AI & models3 October 20264 min read
Apple restricts macOS disk access to curb AI agent data risks
Apple announced on 3 October that it is tightening Full Disk Access permissions on macOS to stop AI agents from reading private data without clear user consent, a move that follows a high-profile security incident involving Meta.
AI & models3 October 20265 min read
CoreWeave Forge and the Agent Identity Gap: Enterprise AI Compliance Gets Messy
On 30 September CoreWeave launched Forge, a platform for building and running AI agents, as enterprises keep discovering that governing those agents is harder than deploying them.
AI & models2 October 20265 min read
04
NHL Turns to Romance Audiobooks as AI Video Makers Chase the Same Attention
The National Hockey League began rolling out romance-novel-style audio shorts this week. The campaign has landed badly with some of the genre's best-selling authors, even as AI video generators keep multiplying.
AI & models2 October 20266 min read
05
Denmark Puts Data Centres Last in the Grid Queue as AI Power Demand Bites
Denmark's parliament has adopted an emergency plan that drops first-come, first-served grid connections and ranks data centres in the lowest priority category, the Ministry of Climate, Energy and Utilities said on 2 October.
AI & models2 October 20263 min read
06
AI browser agents move from demos to production, and the first bill comes due
Nvidia's Open Agent Safety Platform, announced on 28 September, ships a reference design that quarantines misbehaving agents in milliseconds. It is the first infrastructure answer to a month of agent escapes, leaked screenshots and failed intrusions.
AI & models2 October 20264 min read
07
ServiceNow's AutoSynthData Turns Agent Failures Into Training Data
ServiceNow published an enterprise agent pipeline on 2 October that generates new training tasks from a model's own failures, targeting the retrieval and tool-use gaps that block deployments. AutoSynthData was detailed on Hugging Face by four ServiceNow researchers.
AI & models2 October 20267 min read
08
Open weights, open questions: Cloudflare's Clef takes on Jev as Amazon and OpenAI pile in
Cloudflare released two open-weight decision models, Clef and Clef-flash, on Thursday 1 October, claiming the pair beat Typesafe AI's Jev on accuracy while staying downloadable under an Apache 2.0 licence.
AI & models2 October 20265 min read
09
Small models move to the edge as Google locks down its biggest one
Google began rolling out Gemini 4 Argon on Wednesday, but only to vetted cybersecurity partners, while the same week produced an open 145M-parameter model trained by volunteers and a 2B decision model from AWS. The frontier and the edge are pulling apart.
AI & models2 October 20267 min read
10
AI Water Use Data Goes Missing: EU Complaint, Local Revolts and the Model Race
A complaint filed with the European Commission on 30 September accuses the EU executive of hiding the energy and water consumption of data centres, the same week local opposition to AI campuses intensified in the US, Indonesia and Australia.
AI & models2 October 20266 min read
11
Open weights week: Cloudflare Clef, Amazon, Google and DeepSeek move on the open model market
Cloudflare released two open-weight decision models on Thursday, claiming they beat TypeSafe's Jev on accuracy while admitting they cost close to six times more per million tokens. The launch caps a week in which Amazon, Google and Huawei all moved on open or low-cost model tooling.
AI & models2 October 20265 min read
12
Kortext buys StudyStash as AI reshapes the education workforce
Kortext acquired adaptive learning platform StudyStash, the companies said on 2 October, a year after the startup's founders graduated from the University of Birmingham. Financial terms were not disclosed.
AI & models2 October 20265 min read