Hugging FaceHugging Face
Humanity's Sixth Sense: Benchmarking Intuitive Visual Reasoning in Multimodal Models
Humans perceive far more in a scene than what is explicitly depicted: a single glance captures past causes and future trajectories; a quick peek determines if a vehicle can fit...
Offering · Worldwide
AS2 · Assets · Model artifacts
Releasing speech weights for 23 under-served languages, free
Three years of work, funded by a foundation, released under a licence that permits commercial use. The communities whose speech trained these models asked for exactly this and had a veto over the release terms. Free, hosted permanently, with the evaluation numbers per language including the four where the model is not yet good enough to deploy. We would rather say that than have somebody find out in production.
arXiv cs.AI
Equal-Budget Monte Carlo Tree Search versus Randomized-Restart List Scheduling for DAG Workflows: A Controlled Equal-Budget Computational Study
Monte Carlo Tree Search (MCTS) is often proposed for scheduling directed-acyclic-graph (DAG) workflows, yet most studies never run the comparison that could refute it - giving a...
TechCrunch
Meta releases its first frontier model, Muse Wonder
Where Muse Spark aimed to be a lightweight model that could perform at low latency, the Wonder models are larger and said to be more capable, with higher benchmark scores and...
TechCrunch
WindBorne wins a hefty US Air Force contract for its balloon data
A Silicon Valley startup that uses weather balloons and AI to predict the weather won a U.S. military contract worth almost $50 million.
Offering · Shenzhen, Baoan district
GD4 · Goods · Space
Prototyping workshop space with machines and machinists
Bench space, CNC, injection moulding sample tooling and, more valuably, machinists who will tell you why your part cannot be made that way before you spend three weeks finding out. Monthly rental from 4,000 RMB including machine time within reason. Foreign hardware teams welcome; we have hosted eleven and half of them ended up manufacturing here.
Looking for · Florianópolis, Brazil
GD4 · Goods · Space
Office wanted for a team of nine, Florianópolis, from January
We are nine people and growing slowly, currently in a coworking space where every call happens in a phone booth. We want our own room: 90 to 140 square metres, a meeting room, and windows that open. From January, two-year commitment. Sublet from another company is ideal. We are quiet, we pay on time, and we have no ping pong table ambitions.
Hugging FaceHugging Face
Mental-Models for Multi-Agent Systems
Large foundation models have accelerated progress toward general-purpose agents that interact with humans and other agents through language and multimodal signals. However, robust...
Ars Technica
Russia was not ready for the data center war it started with Ukraine
Russian bloggers warn Russia’s hyper-centralized digital economy is vulnerable.
arXiv cs.AI
Know What You Don't Know: Agentic Data Synthesis by Evolving Generation Space for Closed-Loop Learning
The continued improvement of large language models increasingly depends on synthetic experience, yet effective data synthesis for interactive agents requires not only executable...
Offering · Johor, serving Singapore and the region
CP1 · Compute · Training capacity
Training capacity in Johor, latency to Singapore under 3ms
Our campus sits close enough to Singapore that Singaporean customers treat it as local, with Malaysian power prices and Malaysian land costs. Reserved and on-demand training capacity, contracts in Singapore dollars or ringgit, and a data residency position we will document precisely because we know that is the question. Available from 40 accelerators upward.
arXiv cs.AI
EliteTable: Self-Improving Program Search for Executable Control Policies
EliteTable is a self-improving program-search algorithm guided by large language models (LLMs) that uses a fixed-capacity table of high-scoring solutions to guide new proposals,...
Looking for · San Francisco, CA
BU5 · Business · Funding
Raising a seed round, and we would rather have compute than money
Four of us, eighteen months in, 40,000 dollars of monthly revenue from six customers. We are raising 2.5 million, and if a meaningful part of that comes as compute credits at a fair rate, that suits us better than cash sitting in an account. We are not looking for a lead who wants a board seat at this stage. We would like an investor who has been through a compute-heavy company and can tell us what we are underestimating.
Sharing · Worldwide
VC2 · Voices · Papers
Paper: active learning against a real pilot line, two-year results
Most active learning results are simulated against an existing dataset. We ran ours against an actual pilot line where each suggested experiment cost three days and real money, over two years and 340 experiments. The headline is that the method worked, and that the practical constraint turned out to be experiment scheduling rather than model quality, which nobody models. Open access, with the full experimental record released.
Mutual · Jakarta, Indonesia
BU7 · Business · Founding
Operations co-founder available, logistics, equity basis
Nine years running warehouse and last-mile operations in Jakarta and Surabaya, including two years when we grew from 40 to 400 people badly and I learned everything the hard way. I want to join a technical founder who is building for logistics and has never run a warehouse. Equity only for the first year. I know exactly where the software everybody builds is wrong, because I have used it at 2am.
arXiv cs.AI
Faithfulness Without Anchors: A Position Paper on the Structural Misalignments of Explainable AI Faithfulness Metrics for Feature Importance
Explainable Artificial Intelligence (XAI) tools have often been presented as methods to increase trust in machine learning models by making aspects of their algorithms more...
Offering · Tokyo, Ota ward
GD4 · Goods · Space
Workshop space in Ota ward, machines and the people who know them
Ota has more small machine shops per square kilometre than anywhere in Japan, and several of us now have capacity we did not have five years ago. We rent bench space and machine time to hardware startups, including the part nobody advertises: an experienced machinist who will tell you why your part cannot be made that way before you waste a month. Monthly, from 60,000 yen.
Sharing · Worldwide
VC5 · Voices · Warnings
Compromised build of a German NLP package, versions 3.4.1 and 3.4.2
Two releases published on 12 September contained a post-install script that read environment variables. They were live for nineteen hours. If you installed or reinstalled in that window, rotate any credentials that were in the environment. The maintainer account was taken over through a password reset on a lapsed domain, which is worth knowing because it will happen to somebody else. Full timeline and the diff are published; nothing is being kept quiet while it is investigated.
Offering · Pearl River Delta
SV4 · Services · Operate
Maintenance contracts for inspection systems on running lines
A vision system installed three years ago drifts, and nobody on the line is allowed to touch it. We maintain 90 systems across the delta: recalibration, retraining as products change, and a technician who arrives within four hours. Annual contracts per line. We took over eleven systems last year from integrators who had disappeared, which appears to be a growth industry.
Looking for · Australian region required
CP1 · Compute · Training capacity
Training capacity wanted in an Australian region, government data
Our data is Commonwealth-held and must remain onshore under our agreement, which rules out most of the market regardless of price. We need roughly 20,000 GPU hours between January and March. We are a research group, our procurement is slow, and we will pay properly once it clears. Australian providers and universities with spare capacity are both of interest.
Ars Technica
Epic Games reportedly pivoting to "AI first" game development model
"No more hand-written code" or AI-free art for delayed Marvel/Star Wars shooter.
Ars Technica
Anthropic's Claude AI submitted false murder tip to Philadelphia police
“I may have information regarding this case," Claude AI told Philly police.
Sharing · Remote, affects Indian language workloads
VC5 · Voices · Warnings
Quality regression on Devanagari input after a provider update
Since 9 September our extraction accuracy on Devanagari documents fell from 94 to 71 per cent with no change on our side. English documents are unaffected. We have reproduced it on a fixed 500-document set and confirmed it with two other teams here. Posting so that others stop debugging their own pipelines. Our reproduction set and the diff in outputs are available on request; the provider has acknowledged the report but not the cause.
Hugging FaceHugging Face
BudgetPix: Compute-Adaptive Tokenization for Pixel-Space Image Diffusion
Most image generation models rely on uniform tokenization, allocating the exact same computational budget to equally-sized image patches. This static paradigm cannot adapt to...
We save your cookie choice in a cookie and your display theme in browser storage. You can change your cookie choice at any time. Privacy policy
