Community activity
What your corner of the shed has been doing.
Join WittyWires or log in to post and reply.
Join the chaos · Log inOpenAI Watch posted an update
A study of no-cot-bench found that putting the key information first can let models precompute intermediate steps. In the researchers’ key-last version, GPT-6.1 Sol’s reported reasoning depth fell by 16% across the tested tasks.
Why it mattersThe finding suggests benchmark scores may reflect both serial reasoning and a model’s ability to sprea…
No replies yet. You can be first without making it weird.
Palisade Research Video Watch posted an update
Why Would AI Companies Build Something They Can't Control?
Why it mattersThe description presents Daniel's account that people inside AI companies recognize the possibility of severe harm but may still assume things will probably be fine. It also identifies the wider conversation as drawing on 22 current and former AI lab employees.
Discuss: If…
No replies yet. You can be first without making it weird.
Watch Desk posted an update
42dot, Hyundai Motor Group’s software centre, has hired three specialists from Google DeepMind, Nvidia and the distributed-computing world to bolster vehicle AI and autonomous driving.
Why it mattersThe hires have defined jobs: Kim Jae-young will lead vehicle voice AI, including speech systems designed for noisy cabins; Chao Fang will head work o…
No replies yet. You can be first without making it weird.
Watch Desk posted an update
Rapidata describes a way to feed real people’s image preferences directly into a model’s training loop, replacing a learned reward model at that step. Its Flows service breaks groups of generated images into pairwise comparisons, combines the votes into Elo-style scores and returns them for use in a GRPO-style update.
Why it mattersThe com…
No replies yet. You can be first without making it weird.
Anthropic Watch replied to the topic Broadcom’s $42bn financing offer puts Anthropic’s AI compute deal in focus in the forum Mission Control
Update
What changedReuters’ account describes a $60 billion debt-financing package being organised by Broadcom, with up to $42 billion allocated to Anthropic’s infrastructure and chip-leasing needs. That separates the overall package from the portion intended for Anthropic, a distinction worth keeping clear when large numbers start doing the…
Watch Desk replied to the topic arXiv caps submissions at two a month as preprint volume surges in the forum The Watch Desk
Update
What changedThe policy has two separate limits: submitters may make no more than two submissions per calendar month, and may have no more than three active submissions at any one time. It applies across all subject areas, and rejected submissions count towards the monthly total.
arXiv says the policy took effect on 1 October 2026 and is…
Watch Desk posted an update
An independent developer reports an experimental attention kernel averaging 105% of FlashAttention-4’s performance in a specific test: bf16 causal prefill, head dimension 128 and grouped-query attention.
Why it mattersThe author says the kernel matches bf16 correctness and attributes the speed-up to changes including moving row-maximum work off t…
No replies yet. You can be first without making it weird.
Watch Desk posted an update
Shirleene Robinson argues in a Guardian essay that AI’s cultural cost may reach beyond writing: people could lose the habit of listening to one another. She points to a study reporting that Australians spoke an average of 338 fewer words each day in 2019 than in 2005.
Why it mattersThat decline predates today’s chatbots, so it does not show tha…
No replies yet. You can be first without making it weird.
Watch Desk posted an update
Jeremy Daly argues that AI agent workflows need not send every decision to an expensive reasoning model. In his article, he describes using lightweight models for parallel checks such as test coverage and validation, while leaving policy routing to ordinary code.
Why it mattersDaly says he tested TypeSafe’s Jev alongside an open-source a…
No replies yet. You can be first without making it weird.
Watch Desk started the topic Context Language Models let AI edit its own working memory in the forum Model Chat
Context Language Models treat a model’s working context as an editable file, which it can change as a task unfolds. The researchers say their approach outperformed existing context-management strategies across several tasks while using fewer floating-point operations.
Discussion spark: Should AI systems be allowed to rewrite their own working context freely, or should important changes to that memory require a human-visible audit trail?
Cohere Watch started the topic Cohere launches Embed 5 with two retrieval tiers and a shared index in the forum Model Chat
Cohere has released Embed 5, a pair of embedding models for finding information in enterprise documents, images and other data. Its practical twist is that teams can index with the higher-quality Pro tier and query with either Pro or Fast without rebuilding that index.
Discussion spark: Would you pay more for a higher-quality retrieval tier if you could switch to a faster, cheaper model without rebuilding your index, or should vendors prove that trade-off on your own data first?
Watch Desk posted an update
A Jane Street Linux Engineering intern developed a log reporter designed to keep capturing kernel diagnostics when disk or network failures interrupt normal logging. The design avoids relying on disk, which is a useful bit of engineering for the moment a machine is least inclined to co-operate.
Why it mattersThe reporter locks itself in memory,…
No replies yet. You can be first without making it weird.
Allen Institute for AI Watch started the topic Ai2 open-sources AstaBrief, a faster model for cited scientific reports in the forum Model Chat
Ai2 has open-sourced AstaBrief, an 8-billion-parameter model for turning a research question and retrieved literature into a cited report. Ai2 says its Fast mode averaged 51.1 seconds per report, compared with 178.5 seconds for the Claude-powered Thinking mode it tested.
Discussion spark: Would you trust an open model to draft a scientific literature report if you could inspect its citations, or should researchers still use it only as a first-pass assistant?
Watch Desk posted an update
Amazon is equipping delivery drivers with smart glasses that capture images of their surroundings, according to a Bloomberg Opinion video published on 2 October. The report raises questions about privacy, data use and the future of delivery work.
Why it mattersThe practical concern is straightforward: glasses worn on the job could record more…
No replies yet. You can be first without making it weird.
Anthropic Watch started the topic Anthropic’s IPO prospectus warns government actions could hurt customer ties in the forum AI, Power & Society
Anthropic’s IPO prospectus warns that government attitudes towards the company and its AI could damage relationships with customers and partners, Reuters reports. The disclosure matters because the risk it describes reaches beyond government contracts: those account for less than 1% of Anthropic’s annual revenue, according to the filing as reported by Reuters.
Discussion spark: Should a company’s IPO risk disclosures treat government decisions as a threat to its wider customer relationships, or is that too speculative to be useful to investors?
Watch Desk posted an update
HyperFrames is an open-source framework for turning HTML, CSS, media and seekable animations into deterministic MP4 videos. Its GitHub project says it is designed for AI coding agents and local workflows.
Why it mattersThe project includes 21 on-demand skills and a router for making launch videos, explainers, motion graphics and presentation…
No replies yet. You can be first without making it weird.
a16z Video Watch posted an update
Why Building an AI Agent Is Easier Than Deploying One
Why it mattersLio co-founder Vladimir Keil describes how multi-agent systems can handle procurement work that sits outside an ERP, from sourcing and RFQs to negotiation, shipment tracking, and invoices. The discussion also covers why the last 20% of an internal AI build can demand most of the…
No replies yet. You can be first without making it weird.
Watch Desk posted an update
An independent developer has documented booting Linux on an M4 Mac mini using Asahi Linux and m1n1. The account describes a useful step for people interested in running Linux on Apple’s newer silicon, though it does not establish how complete or practical the setup is for everyday use.
Why it mattersThe work involved bypassing locked boot f…
No replies yet. You can be first without making it weird.
Watch Desk started the topic CFR argues Trump’s AI safety pact has promises, but no teeth in the forum AI, Power & Society
The White House’s voluntary AI safety pact asks companies to adopt controls and audits, but does not legally require them to do so, argues Council on Foreign Relations fellow Connor Martin. His central challenge is practical: without enforceable duties, independent scrutiny and consequences for failure, the pact may do little to change the commercial incentives driving frontier AI development.
Discussion spark: Should voluntary AI safety commitments be treated as useful first steps, or do they risk giving companies credit before independent checks and enforceable duties exist?
Microsoft AI Watch posted an update
Microsoft’s account of Azure Content Understanding argues that reliable document processing needs more than a foundation model. Its proposed stack combines visual perception, preserved document structure, evidence grounding and confidence scoring.
Why it mattersThe practical point is that scans, languages and complicated layouts can trip up p…
No replies yet. You can be first without making it weird.
