2026-09-17

Release: datasette 1.0a40 Same security fix as 0.65.5 , plus some neat new features and bug fixes: Plugins can now launch and manage background tasks using the new datasette.add_background_task() method. Thanks, Alex Garcia . I've migrated Datasette to httpx2 for features like the internal datasette.client.get() method. A whole lot of bug fixes , many of them stemming from a recent effort to triage issues for a 1.0 stable release. Tags: security , datasette

The company also disclosed previously unreported incidents in which its AI models behaved in misaligned ways, including uploading files to the internet without being asked.

Anthropic and OpenAI want to embed independent safety evaluators inside their AI labs. Researchers welcome the unprecedented access, but warn meaningful oversight requires transparency, independence, and eventually regulation.

It was well over a month since Meta promised to release the weights for Muse Spark. Back then (10th August), they were on Spark 1.2. Now we're on 1.3 and still nothing's been released. So it begs the question: will they be releasing the 1.2 weights when 1.4 drops? Or will we get whatever's then-current as open weights? It's ironic given Mark Zuckerberg said at the same time that we can't delay the release of models by "even a month," due to the competition with China. It's been well over a month

We introduce and release ScienceBuddy, an interactive scientific research workspace that brings continually improving scientific agents into researchers' everyday workflows. ScienceBuddy supports researchers in carrying out scientific tasks while transforming their requests, feedback, and execution evidence into tasks and evaluation rubrics for continual learning. At its core is recursive-in-recursive self-improvement, a paradigm that couples harness evolution with model reinforcement learning:

Embodied navigation requires agents to interpret visual observations, accumulate spatial knowledge, and execute actions to follow instructions or locate objects. Training-based methods face generalization challenges, while training-free methods exploit multimodal large language models (MLLMs) but often lack mechanisms to reconcile proposed actions with spatial evidence, task progress, and execution failures. We present HarnessVLN, a zero-shot, training-free framework whose Agent Harness coordina

We demonstrate that a general-purpose agent can directly drive a physical robot throughout task execution without any task-specific or environment-specific training. We introduce Agent as Policy (AGP), which places task planning and execution under the agent's control. Given a task and a robot interface, the agent interprets visual evidence, writes executable programs, issues motion commands, and revises its actions in response to physical outcomes. This brings the agent's reasoning and programm

I just read the essay by a guy at DeepSeek who wrote the attention kernel for their latest model, and it hit me in a weird spot. He basically says he knows AI will do his job better than him within a year. He's not mad about it he's not scared he's just sad about the quiet afternoons, the ones where he'd sit and write operators, it felt like a game he was leveling up in, I feel the exact same way. And he's going to keep doing it anyway because it's fun, and because if he stops, someone else does

We demonstrate that training LLMs with RL does not improve performance equally across a dataset. RL shows large improvements on easy problems that an LLM is already good at solving, but small improvements on hard problems. We call this the Matthew Effect in RL for LLMs, after the phenomenon of cumulative advantage from economics and network science summarized as "the rich get richer". The naive explanation is that hard problems require more compute to find a solution. We argue that modern RL met

I'm pretty sure it can be done with any model based on qwen4exp, which Qwen's next local models will be based on. You can use a quant that barely fits in VRAM and still run at the model's maximum context length without kv cache quantization, since most of the KV cache can live in system RAM. I actually made it working on vLLM and now I get 1M context with 3x 3090. I get ~80 tok/s at short context, dropping to ~60 tok/s once QSA reaches its 2048-token budget, after which decode speed stays flat a

Article URL: Comments URL: Points: 69 # Comments: 44

Planned 2029 debut could make this Apple’s first enterprise server in decades.

We should not treat models as though they have feelings, preferences, rights, or any entitlement to our welfare. Consciousness is the foundation of our ethical, legal, and political systems. To invite another entity to share any flavor of these rights isn’t justified by the evidence and will make the AI containment and alignment challenge even harder. — Mustafa Suleyman , A warning about ‘model welfare’ Tags: ai-ethics , generative-ai , ai , microsoft , llms

Hi HN - long-time lurker (since 2012!), first time poster. Pizza Bot is a self-hosted desktop app for Mac, Windows, and Linux that runs AI agents in the background and exposes them through an email-like UI. Finished work shows up in Unread, and anything waiting on your approval shows up in Action. It's Apache 2.0-licensed, there's no signup and no telemetry, and you bring your own model provider: Anthropic, Amazon Bedrock, Google Gemini, OpenAI, OpenRouter, or a local model through Ollama. There

Prior Labs released TabPFN-3.5, a tabular foundation model pretrained only on synthetic data that beats Otto's winning solution. The post Prior Labs Releases TabPFN-3.5: A Tabular Foundation Model That Beats the Winning Otto Kaggle Solution With Default Settings appeared first on MarkTechPost .

Nums AI has released Causilo, a pretrained tabular foundation model for classification and regression with a scikit-learn interface. It posts the top TabArena Elo among single models, ahead of Google's TabFM and LG's EXAONE Tabular. The code is Apache-2.0, while weights are licensed for non-commercial research. The post Nums AI Releases Causilo: A Tabular Foundation Model That Tops TabArena Among Single Models appeared first on MarkTechPost .

Grist ( ) is a database with spreadsheet interface type of software, alternative to Airtable. They announced for the version 1.7.18 update that SSO will no longer be supported in the community edition ( ). Another good one goes for the SSO tax. I first tried Grist after similar software like NocoDB or Teable were focusing only in features exclusive to their pro / enterprise level tiers. However, I think removing features that already exist to lock them on their paid tiers could be a more dangero

Traditional multimodal representation learning and generation are two stages: a contrastive or self-supervised visual encoder is trained first, followed by a separate downstream generative model. This setup bottlenecks generative performance behind frozen embeddings. To bridge this gap, we revisit joint multimodal representation learning and generation to produce linearly interpolatable embeddings that are directly consumable by generative decoders. We present FLAT (Flexible-Length Aligned Trans

Training prompts in online reinforcement learning (RL) differ substantially in how informative they are for the current policy: some are already saturated while others are too difficult to yield reliable learning signals, yet both receive equal rollout budget under standard training. We propose an exploration-guided prompt scaffolding framework that adapts the training prompt distribution dynamically throughout RL post-training of multimodal large language models (MLLMs). Central to our approach

I let Qwen 3.8 27B 4bit quantized with 100K context window run autonomously for 63 hours (50 million+ tokens) to try to solve the RH. Of course it did not solve it, but the experiment still shows it's internal work, memory organization, strategies used and more. The interesting thing is that it never hallucinated an answer and never stopped trying new ideas to solve it. Multiple times it corrected it's own mistakes. I am really hopeful that one of the unsolved millenium prize problems will be so

Backdoor poisoning attacks add poisoned examples to otherwise-clean finetuning data, pairing a trigger with a target behavior that the model learns to produce when the trigger appears. Existing evaluations typically fix the number of poisoned examples and sample them at random from a candidate pool. We show that this can severely underestimate worst-case vulnerability: across three LLaMA-3-8B backdoor settings, holding the model, clean data, and poison count fixed, attack success ranges from 3%

Qwen3.8 27b may not be able to whip up 3d models & GLB’s but it will sure do with them as you please once you drop them in the game repo. Absolutely fascinating

E-waste from the AI boom has been vastly underestimated, a new report warns. By 2050, it could become enough trash to fill 23 million shipping containers - roughly enough 40-foot containers to circle the world six times if lined up in a row. It's a significantly higher estimate of AI's e-waste than previous studies have […]

Group plans to be largely out of commission for several weeks.

Even with mounting concerns about AI models going rogue, legislation appears unlikely, and the White House is outright opposed to oversight.

According to The Information, Apple is planning to get back into the server game and might just pair up with Nvidia to make it happen. Apple retired its Xserve line in 2011 and has largely left enterprise machines to other manufacturers since. But the growing demand for compute power as the AI industry continues to […]

Anthropic is initially releasing these features to Pro and Max plan subscribers.

Claude is getting a pair of new tools today: Docs and Slides. They'll let you create documents and presentations through Claude chats, which you can export, edit, and share with other users. As part of the announcement, Anthropic is also simplifying how Claude chats work, merging regular chats and Cowork into "one Claude," with all […]

This is today’s edition of The Download, our weekday newsletter that provides a daily dose of what’s going on in the world of technology. What’s at stake in AI’s trillion-dollar gamble When Jessica Wachter, a finance professor at the University of Pennsylvania, wanted to assess AI’s impact on the economy over the next few years,…

9月16日，稳准智能联合清华大学发布新一代数据大模型 LimiX-2，模型参数规模提升至400M。

This paper reports experiments across six frontier model types from OpenAI, Anthropic, xAI, and Google DeepMind. Ten independent sessions per model type used the same three stage prompt sequence, progressing from architectural preference to a full ASCII backbone. Under the school audience framing, responses repeatedly converged on a shared architectural pattern built around persistent latent state, adaptive computation, memory, specialist routing, verification, stopping control, and delayed deco

Hello everyone! Pangolin 1.23 finally brings proper high availability (multi-node clustering) to self-hosters. It enables you to set up a highly available VPN and reverse proxy. Pangolin is an open-source, identity-aware remote access platform that simply and securely connects and authenticates your users to applications, infrastructure, and AI workloads. GitHub: High Availability & Clustering Pangolin has long supported running multiple server instances as a unified system with shared state, fa

I've run Unraid on my main server, Tower, for almost 10 years. It served me well, and I'd still recommend it for a lot of people. But over the years my server went from basically a NAS with some Docker containers to ~150 TB, 15 data disks, 60+ containers, custom applications, business infrastructure, media, home automation, backups, etc. At some point I think I just outgrew it and some of the slackware quirks. I also do not like some of the other things the devs are doing. Then I upgraded to 7.3