2026-07-24

Aggressive training techniques sharpens threat of bad behavior by leading models.

Poolside's co-CEO on how his small team of top researchers built a model factory capable of training Laguna S - a 118B MOE beating Thinky's ~1T open weights model... and this is just the beginning.

I genuinely believe that if you took an open weights model from 2025 and built a pentest harness for it, it could do this kind of sandbox escape and scan/hack in most networks. This is only surprising because you assume OpenAI has sounder sandboxes. — Thomas Ptacek , doesn't think this even needs a frontier model Tags: thomas-ptacek , openai , security , generative-ai , ai-security-research , ai , llms , sandboxing

hey HN, Jonathan and Guy here, creators of OneCLI ( ). OneCLI is an open source vault for AI Agents. Traditional vaults are used to store your secrets and, on demand, provide them to you all in a secure way, trusting the person to keep them safe. We figured that in the agent's world, this is not the case, as you don't know what happens with the secret after it's delivered to the agent, or where it was saved. Or maybe someone even manipulated them to hand them over... From that understanding, we

Hey everyone, Two years ago, we started building Pocket ID because we couldn't find what we needed: a clean and easy to use self-hosted OpenID Connect Provider with passkey support. We never expected it to grow this quickly. Today Pocket ID protects users across thousands of self-hosted deployments. That's exciting, but it's also a responsibility we take seriously. Identity infrastructure doesn't get second chances. A bug in a media player is annoying.A bug in an identity provider can become an

Reinforcement learning (RL) has become a dominant paradigm for enhancing LLMs' reasoning capabilities. However, RL algorithms with PPO-Clip are inherently limited by exploration collapse. Subsequent works remain primarily heuristic and fail to identify the essential cause of PPO-Clip's failure. This work reveals the fundamental flaw of PPO-Clip: it implicitly measures policy discrepancy using Euclidean metric, which is theoretically inconsistent with the intrinsic geometry on the policy Riemanni

from kwaipilot: Following the release of KAT-Coder-V2.5 in July, we are pleased to release the open-weight version KAT-Coder-V2.5-Dev , an MOE model with a total parameter count of 35B and 3B activated parameters, to strengthen communication with the community and showcase our research achievements. KAT-Coder-V2.5-Dev Highlights Performance improvement. Through SFT/RL training, KAT-Coder-V2.5-Dev achieves SOTA results in the field of Agentic Coding among models with similar parameter scales. Opt

Reinforcement learning with verifiable rewards (RLVR) is rapidly advancing the reasoning capabilities of language models, yet the optimization layer that converts reward feedback into weight-space updates remains poorly understood. Building on our prior analysis (Zhu et al., 2025), we study this missing layer through the singular structure of model weights and identify spectral inheritance: RLVR can reuse the base model's weight spectra while acquiring new behavior through changes in the associa

The Media Router is a tool that automatically selects the best image, video, or audio generation model for a request based on whether a developer prioritizes quality, speed or cost.

An updated government advisory warns that Iranian hackers are exploiting systems used by water and energy providers.

Google continues to report big quarterly revenue, but its AI spending has skyrocketed.

Hi HN, we are Marcos and Harrison, cofounders of Palmier ( ). We are building Palmier Pro, an open source macOS video editor, with built-in AI generation and a local MCP server that connects to your agent. Here are a few demos: - Making some AI transitions: - Multicam editing with Codex: - Cutting long form clips into shorts: We built Palmier Pro as an internal tool when we were making AI launch videos for other startups. The main problem it solved in the beginning was the back-and-forth between

I am the developer of an fairly popular OSS project (which has Windows, Mac, Linux binaries and a play store app - all built directly from source code with Github CI/CD) which is completely free to use and 100% private and no ads. Someone copied the project, the only "enhancement" they made was adding personalized in-app ads to the project and released it as their own app with some catchy keywords and the same project name - "Open DroneLog" but with a different identifier com.opendronelogandroid

Starts a fresh Codex worker and critic every cycle Discussion | Link

We present AutoIndex, a framework for learning representation programs: executable transformations that map raw documents into the representations exposed to a retrieval system. Rather than tuning retrievers, rerankers, or a small set of preprocessing hyperparameters, AutoIndex searches over programs that slice, enrich, normalize, reweight, or reorganize documents before indexing. At each iteration, AutoIndex performs validation-guided program search, in which agents diagnose failures of the cur

The reviews were just released, and I downloaded my paper from OpenReview to identify areas that needed improvement. However, GPT warned me that the PDF contained a prompt injection. I never inserted such a prompt. After comparing my original submission with the version downloaded from OpenReview, it appears that the injection may have been added by NeurIPS. I would like to know whether anyone else has encountered the same issue. Also, check your reviews for suspiciously formulaic wording. If a

Reinforcement learning with verifiable rewards (RLVR) has substantially improved the reasoning capabilities of large language models on tasks such as mathematical reasoning and code generation. However, most RLVR methods assign a scalar outcome reward to an entire trajectory, resulting in sparse supervision and limited token-level credit assignment. On-policy distillation (OPD) provides denser supervision by distilling token-level distributions from a stronger teacher model, but requires an addi

Now live on OpenRouter, and free to use through August 3, 2026. Hoping they will going openweight soon~

I've got some time tonight and am excited to run it through my agent loops and form my own opinion. Wanted to hear everyone else's as well, but without the barchart noise.

The Summer 2026 Edition

Recent autoregressive video diffusion methods are increasingly built upon Self Forcing, where the student is trained on histories produced by its own rollout rather than ground-truth video contexts. This reduces exposure bias, but the historical key-value cache is still used by future frames only as frozen rollout state. As a result, future losses cannot supervise how earlier generated latents should be written into more useful keys and values for later video-latent generation. We call this the

arXiv:2607.18252v1 Announce Type: new Abstract: Machine learning methods have shown that data-driven policies can accelerate mixed-integer linear programming (MILP) solvers, but many such approaches remain difficult to inspect, adapt, and deploy because the learned policy is represented as an external predictor or other opaque model. By contrast, explicit solver logic is easier to understand and integrate, but is usually hand-designed rather than learned from solver feedback. We study whether th

具身智能正在从「参数竞赛」进入「架构竞赛」。

The power of Codex with local, self-hosted models and voice Discussion | Link

The Python Package Index (PyPI) now rejects new files being uploaded to releases that are older than 14 days. This restriction was put in place to prevent old and long-stable releases from being poisoned in case publishing tokens or workflows of PyPI projects were compromised. As far as we are aware this has not yet been abused, but there is no technical reason beyond that attackers weren't aware it was possible. — Seth Larson , PyPI blog Tags: packaging , python , supply-chain , pypi , seth-mic

Run GPT-5.6 on OpenClaw & Hermes safely on your Mac Discussion | Link

The news about Anthropic settling a class action lawsuit for $1.5B over training data isn't just a legal headache for them, it's a massive warning sign for engineering teams relying entirely on closed API vendors. When you route core business logic, proprietary codebases, and customer data through third party APIs, you're exposing your stack to three massive risks: The massive fine(not rly massive for anthropic) is not gonna come out of their pockets, they gonna charge you more to make up the co

Finetuning a pretrained vision-language model (VLM) on robot demonstrations via behavior cloning (BC) has become the standard recipe for vision-language-action (VLA) policies. However, BC finetuning progressively overwrites the pretrained representations that support visual and semantic generalization. Co-training on web image-text data, a common remedy, does not prevent this; it applies language and action losses to separate observations, leaving VLAs with language-action misalignment that stan

Just saw this on huggingface: The benchmarks claim that it's quite a bit better than qwen3.6 27B original and that they reduced the amount of necessary tokens by more than 90%. It would make 27B running on my old laptop at 3tps feel more like 30tps for the thinking part, if true. Couldn't test it yet.

Evaluating the factuality of long-form generations has focused predominantly on precision, measuring whether the claims a model makes are correct. The dominant decompose-search-verify pipeline catches incorrect claims well but says little about whether a response contains all the information it should. Measuring factual completeness, the missing half of factuality, is harder: it requires enumerating the full set of facts a complete answer should contain, and these facts rarely form a flat list.

At the moment MLX (and Llama.cpp for Macs) run 16bit activations everywhere. Despite this, the M5 generation silicon actually does support INT8 activations - it actually allows w4a8 d_type. It's just that no inference backends are using them yet I built some w8a8 kernels and have managed to get 1.4x speed up on Gemma4 prefill tasks; on my M5 MacBook Air it brings baseline prefill for the E2B from 2193 tps stock to 3,029 tps for 130,173 tokens of input* *Even faster at small context lengths; it a

Link to their official GGUF repo: All the GGUFs received this fix 5ish hours ago - correct yarn_attn_factor to 1.0 (llama.cpp derives mscale) And the chat template fixes a lot of broken thinking, preserve thinking, and tool calling Chat template: So far the model seems to be doing MUCH better.

ethernet." src=" title="FYI You dont need expensive networking for multi-node gpu. 30t/s laguna Q2_K_XL (39.7GB) on 2x4060+1x4060 using a $20 usb->ethernet." /> Turns out a regular ethernet cable between 2 nodes can run laguna UD-Q2_K_XL (39.7GB) using a direct point to point network. Interestingly on `nvidia-smi dmon -s pucvmet -d 2`, the inter/intra gpu traffic is not really capped in this setup - Uses ~30-70MB/s at peak # gpu pwr gtemp mtemp sm mem enc dec jpg ofa mclk pclk pviol tviol fb bar

App is installed through Windows Update when certain LG monitors connect to a PC.

Users can also integrate their personal data from services like Apple Health, Function, and MyFitnessPal.

Bill would let Homeland Security chief decide when an AI should be shut down.

Designing and developing a new medicine is an expensive, failure-prone scientific challenge. A new drug can take many years to develop, at the cost of a significant investment. And even then, most possible candidates never reach the patient. For biologic medicines, therapies made from engineered proteins rather than synthetic chemistry (which are often used to…

Asynchronous reinforcement learning improves throughput by decoupling rollout generation from optimization, but staleness is an inevitable byproduct compounded by policy lag, engine delays, and mixture-of-experts routing. From a trust-region perspective, this mismatch is critical: training-inference divergence governs approximation error in finite-horizon bounds, whereas PPO clipping only gates sampled outward updates, acting as a sampled surrogate rather than a full-policy constraint. As a resu

models, writers, and routers

Ship AI analytics in your product, powered by GPT-5.6 Discussion | Link

Health in ChatGPT now lets eligible U.S. users securely connect medical records and Apple Health to get more personalized insights and better understand their health.

This is rather counterintuitive, since banning Chinese models would benefit them the most. Jensen Huang is also against the ban , although the interests here are more obvious. Do you think that if Arcee, Cohere or Mistral release an open source GPT/Claude level model, they will also be accused of 'unsafety', 'distillation' and other deadly sins?