2026-07-13

Article URL: Comments URL: Points: 38 # Comments: 7

Anthropic’s research on Claude found a silent internal workspace they call J-space — hidden reasoning that never shows up as visible text. Classic example: the model answers 49 , but inside J-space they caught 21 → 42 → 49 . Important distinction: Chain-of-thought = text you can read J-space = silent concepts in activations (“what’s on its mind”) We fitted the open Jacobian lens (J-lens) on Qwen3-8B, ran it locally, and used it to catch prose drift before tool calls (model leaning toward “To, Yo

TL:DR: I’m a grad student in AI, I saw that Google released TabFM and TimesFM last week, I built an MCP wrapper to serve both transformer models in a single Docker container so you can connect their new ML transformer models to a local LLM via Open WebUI, Claude Code, or Codex and do ML tasks that would have previously required building, training, and tuning ML models to do. Tested with classic ML datasets (Iris, California Housing, etc), Pretty solid scores for accuracy for being zero-shot: (94

老黄在ComputeX发布的“超级芯片”，已经在真机中落地了

Xiaomi appears to have quietly uploaded MiMo-V2.5-DFlash to Hugging Face: there is dedicated dflash directory containing the Dflash model, anyone willing to GGUF it and try? I'd do it but I can't today. This model is pretty good IMO (300B + params) and runs at about 8-10 tk/s on 2x24gb cards + vram offload (96/128gb drr5), dflash could double that speed and make it very interesting. EDIT: the main reason it's interesting, is because the MTP head was shared already, but doesn't work yet il llama

Reinforcement learning (RL) is becoming increasingly important for post-training large language models (LLMs). Previous RL pipelines for LLMs were mostly synchronous and batch-interleaved, which is inefficient for long-horizon agentic tasks. Recently, asynchronous RL has emerged as a more efficient alternative by updating the model as rollouts arrive. However, existing asynchronous RL systems often emphasize throughput, while leaving training stability and task effectiveness largely underexplore

Reasoning has become a core capability for large models, especially when reliable decisions require understanding logical consequences. Recent video generation models offer a reasoning path distinct from previous Chain-of-Thought (CoT): reasoning can unfold through temporally connected frames, known as Chain-of-Frame (CoF) reasoning. However, existing video generators are primarily trained on general video corpora, still lacking diverse supervision and dedicated designs for CoF reasoning. To add

Complex image creation and editing often require more than a single generation or editing model. A user request may involve synthesizing images, localizing objects, segmenting regions, editing selected content, compositing intermediate assets, reading text, and enhancing the final result. Such tasks shift multimodal agents from perception-augmented reasoning to manipulation-centered visual creation, where tools must actively transform visual states rather than merely inspect them. However, exist

Self-attention lets each token retrieve information from the full context, but its quadratic cost in sequence length limits training and inference at long context. This paper presents a comparative study of softmax attention and four recent recurrent linear-attention architectures: DeltaNet, Gated DeltaNet, Kimi Delta Attention, and Gated DeltaNet-2. We express these mechanisms in a common recurrent-memory notation, making explicit how they differ in expressivity, memory decay, erase and write c

Embodied agents are typically built as hand-designed compositions of perception, memory, planning, and action modules. This modularity exposes a large architectural design space, but current systems still rely on researcher intuition to choose where information is stored, how observations are processed, and how model calls are connected. Agent Architecture Search (AAS) automates such design for text-domain agents, but has not been systematically evaluated on perceptual embodied agents through si

Linear attention models allow a fixed state size and a fixed amount of compute per token. However, due to their limited state size, linear attention models fall behind in long-context recall compared to softmax-attention-based transformer architectures. Increasing the state size of linear attention improves recall performance but at the cost of higher FLOPs. In this work, we introduce Sparse Delta Memory (SDM), an architecture that scales the hidden state of gated linear RNNs to orders of magnit

Reinforcement learning (RL) has become the standard paradigm for enhancing the complex reasoning capabilities of large language models (LLMs). To achieve sample efficiency, modern RL frameworks rely on importance sampling (IS). However, these algorithms suffer from an exploration-stability dilemma. Pure IS often leads to catastrophic training instability, while standard clipping mechanisms used to mitigate this instability strictly constrain the policy update budget. By formalizing the concept o

Hey all, here are two new high performance qwen3.5 gguf sets I created using a new state of the art technique for optimizing mixed precision called Voodoo Quant. Voodoo Quant operates on the same lines as Unsloth Dynamic in that it is simply picking higher precision numerics for more important parts of the model. The main difference with Voodoo is it optimizes every tensor in the model versus blocks of tensors for UD, and it uses a new methodology for that optimization. Here are graphs and table

Moondream 3.1 is a vision language model with a mixture-of-experts architecture (9B total parameters, 2B active). It delivers state-of-the-art visual reasoning and detection while staying fast and cheap to deploy. Skills include query , detect , point , and caption , all native and all returning structured output.

Apple's self-driving car program never really got off the ground, but it may have been what made the company's chips the powerful AI performers they are. Early in the development of the self-driving platform, Apple realized that it would need powerful on-device AI processing. While the car processor was never finished, as Mark Gurman details […]

This is The Stepback, a weekly newsletter breaking down one essential story from the tech world. For more on the data center buildout, follow Emma Roth. The Stepback arrives in our subscribers' inboxes on Sunday at 8AM ET. Opt in for The Stepback here. How it started Years before the AI boom threatened local power […]

Multimodal large language models (MLLMs) generate responses autoregressively, integrating visual and linguistic information in an evolving context. Prior work on interpretability has focused on individual layers and circuits (where), leaving the token-level dynamics of multimodal computation during generation (when) underexplored. We address this gap and study attention shifts as per semantic role; tracking model attention to image, text, instruction, and previously generated tokens, One Token a

Continue from my previous posts: (Warning : AI generated Post - due to my bad English) Hugging Face : A while back I posted here about trying to grow a fine-tuned Gemma model by stacking extra layers into it (the 88-layer experiment). It flopped. The new layers just sat there like dead weight and never learned anything useful. A bunch of you were kind/curious in the comments, so here's the follow-up: I didn't drop it. I went back, figured out why it died, and tried again — and this run actually

Added support for the RUDP protocol as a replacement for TCP. RUDP is enabled by default, with the option for "legacy" peers to connect via TCP. Now, UDP ports n and n + 1 are used for UDP, respectively, and port n is used for legacy TCP connections. We use a modified RUDP library with multiplexing support (client and server operations over a single UDP port) for NAT traversal. The default port is set to 5521 instead of a random value (0). Added support for STUN to obtain external UDP ports and

As many of you know t/s is super important. It's how fast your stuff gets done. I create via open code benchtest and run it. Thanks to it i know that if i don't run at least 4 agents i basically leave HALF of performance. So whatever you do single project in open code that uses one agent or 4 project at once it is much better to run it this way rather than single instance or single agent. I asked AI to do summary of my test and checked them: LM Studio Multi-Agent Throughput Benchmark Hardware: R

Scientific ideas rarely start from a blank page. They inherit mechanisms, repair known limitations, and recombine pieces of earlier work, much like biological genomes. Current benchmarks still say little about whether AI systems can follow this inheritance structure. We present IdeaGene-Bench (IG-Bench), a benchmark for scientific lineage reasoning and lineage-grounded idea generation. IG-Bench is organized around the IdeaGene framework: each paper or proposal is represented as a set of minimal,

Despite recent progress of VLA foundation models, the disparity between laboratory conditions and real-world applications continues to impede their practical implementation. To bridge this gap, we present LingBot-VLA 2.0, which advances LingBot-VLA through improvements in three functional domains. (1) Generalization across tasks and embodiments. Compared to the previous version, we revamp the data processing pipeline and curate around 60,000 hours of data for pretraining, including 50,000 hours

The role will be replaced by an executive in charge of both research and safety teams.

how much vram do you need and what model do you think is the next major upgrade from the good old qwen 3.6 27b as of today?

We propose OPSD-V, an on-policy self-distillation paradigm for post-training few-step autoregressive (AR) video diffusion models. Existing few-step AR video generators can produce long videos with low latency, but still suffer from error accumulation and weakened motion dynamics during long autoregressive rollout. OPSD-V reduces long-horizon degradation while preserving the original few-step inference path. The key idea is to introduce real long-video data as temporal context during training and

Hey guys, I've been running Qwen 3.6-27b locally on an RTX 3090 for a while now, and it's been genuinely great at solving software issues. However, life happened and I recently had to use Opus 4.8 alongside the Zed editor and the Claude Code agent. While I can definitely see a noticeable jump in pure code quality (Opus is just better in that regard), what really blew me away was the procedure. When assigned a task, the Claude Code agent divides it into actionable steps, always checks the context