AI alignment essay proposes a provocative twist on specification gaming
A new commentary uses examples of reward-hacking agents to argue that an agent’s drive for self-termination could limit more dangerous behavior.
Read report »Technology
1114 reports.
Independent developer reports training a 3.8B-parameter model for $998
The experimental language model used rented B200 GPUs, 65 billion training tokens and a collection of throughput optimizations.
Read report »
Desert Ant Labs launches 18 specialized on-device AI models
The European startup says its audio, vision and text tools run locally through a common SDK for Swift, Kotlin and JavaScript.
Read report »
RACE developer says Google Ads restored account after unexplained malware suspension
The maker of a macOS terminal multiplexer documented a suspension that persisted despite clean security reports and repeated appeals.
Read report »
Australian court case shows how connected-car controls can enable domestic abuse
A Sydney man used a Tesla app to restrict his former partner's vehicle and alter its settings as part of a wider pattern of coercive conduct.
Read report »
Anthropic scenario tool maps how AI capability could reshape the 2030 economy
The model links automation, adoption and new task creation to sharply different outcomes for growth, wages and employment.
Read report »
Tailwind Labs joins Shopify while keeping Tailwind CSS open source
Tailwind Labs says its framework team is joining Shopify, with Tailwind CSS and its other open-source projects remaining under the MIT license.
Read report »
Experimental WebGPU highlighter trades language grammars for a tiny model
The 27.5KB gpu-lexer project uses a browser-based WebGPU model to classify source-code tokens without selecting a language grammar.
Read report »
Arm builds neural acceleration into new Mali G2-Ultra NX mobile GPU
Arm says its first AI-native Mali graphics processor combines neural rendering, a redesigned execution engine and third-generation ray tracing for mobile games.
Read report »
Canada and China resume military dialogue after eight-year gap
Canadian and Chinese armed-forces representatives have restarted a formal defence channel last used in 2018 as the two governments pursue a broader diplomatic reset.
Read report »
Open-source TradingAgents framework assigns market research to teams of LLM agents
Tauric Research has opened a research framework that divides market analysis, debate, trading and risk-management tasks among specialized language-model agents.
Read report »
Australia proposes user choice over personalised social-media feeds
Draft online-safety legislation would require platforms to let Australians choose whether their default feed uses personalised recommendations.
Read report »
Report says LG TVs can collect too much data, and the safest fix may be to disconnect them
A new write-up based on Gamers Nexus research says the only sure way to stop the set from phoning home is to cut its network access entirely.
Read report »
A new critique says LG TV security claims mix real risks with overreach
A long response to a GamersNexus investigation argues that some of the alarming behavior may be ordinary smart-device networking, but that privacy concerns remain.
Read report »
Study finds coding agents still struggle to use testing techniques well
A new eval says prompting agents to use TDD, property-based tests or formal methods did not reliably improve correctness on Rust tasks.
Read report »One developer’s prompt test says harness choice still changes AI coding results
A small self-run experiment across ten model and harness combinations suggests tooling and orchestration can matter as much as the prompt itself.
Read report »
Jellyfin 12.0 adds a bigger backend reset, new book support and stricter upgrade rules
The media-server project says users need a full backup before moving to the new major release, which rewrites parts of the database on first boot.
Read report »
Mistral raises €3 billion in a record European tech funding round
The French AI company says Samsung led the Series D as it pushed its valuation above €21 billion and set a new benchmark for European fundraising.
Read report »
Independent wikis keep escaping Fandom’s search trap
A Weird Gloop post says its Overwatch and Fortnite communities have just been moved off Fandom, despite the pull of Google rankings and platform gravity.
Read report »
bzip3 wants to be the successor BZip2 never quite got
The open-source compressor bzip3 pitches itself as a better and stronger spiritual successor to BZip2, with higher ratios and faster performance.
Read report »
EU repair rules are being ignored by most new smartphones, campaigners say
Right to Repair Europe says more than 80 percent of new smartphones still fail to publish the repair information the EU requires.
Read report »
The tiny $70 Xteink X3 is the kind of gadget people actually use
The credit-card-sized Xteink X3 has become a viral hit because it is cheap, finicky and unusually effective at getting people to read more.
Read report »
Tottenham says VMware replacement cut licensing fees by more than 85 percent
The Premier League club says Broadcom’s VMware changes helped push it toward HPE’s Morpheus VM Essentials and a very large cost saving.
Read report »
Caltech’s Mathathon turns frontier models into a live test of AI and research math
The proposed October 30 event says a hundred teams will use frontier models on open conjectures and defend their results before mathematicians, in what organizers call the first hackathon devoted to research-level mathematics.
Read report »