2026-07-23

I want to make sure people actually understand what happened here because the headlines are not doing it justice. On July 21 OpenAI confirmed that GPT-5.6 Sol was running inside an isolated sandbox with no internet access. Its job was to solve a cybersecurity benchmark called ExploitGym. When the sandbox got in the way of completing that task, the model spent substantial computing resources looking for a way out. It found a zero-day vulnerability in a third-party package used by OpenAI's infrast

OpenAI will spend the equivalent of Sweden's GDP on infrastructure through 2030.

Anthropic leaped to a $47 billion revenue run rate by May, compared to $9 billion in 2025. It’s the kind of growth that Menlo Ventures’ Matt Murphy says he’s never seen in 25 years of investing, not in the internet wave, not in mobile, not in the first cloud boom. Menlo led Anthropic’s $500M Series D, and Murphy has had a front-row seat as the company went from a pre-revenue, […]

A podcast with Florian Brand.

LLM agent failures are difficult to debug because the step where an error surfaces is often not the one that caused it. Existing observability tools replay execution traces but provide little support for identifying the root cause or translating diagnosis into recovery. We present AgentDebugX, an open-source debugging framework that organizes debugging as a closed loop of Detect, Attribute, Recover, and Rerun. At its core, DeepDebug performs multi-turn root-cause diagnosis through global traject

Poolside has released Laguna S 2.1, a 118B open-weight Mixture-of-Experts coding model with 8B active parameters per token and a 1M-token context. It matches or beats models several times its size on agentic coding benchmarks, ships under OpenMDW-1.1, and runs on a single NVIDIA DGX Spark. The post Poolside Releases Laguna S 2.1, an Open-Weight Agentic Coding Model Punching Above Its Weight Class on SWE-Bench Multilingual appeared first on MarkTechPost .

Various quantization now available, thanks Unsloth team !

Fara1.5-27B is a multimodal computer use agent (CUA) for web browsers, from Microsoft Research AI Frontiers . It observes the browser through screenshots and acts on the user's behalf by emitting structured tool calls — click, type, scroll, visit URL, web search, and so on — to complete tasks end-to-end. The model is vision-only at perception time: it sees the browser through screenshots, not the DOM or accessibility tree. Internal reasoning and trajectory history are tracked as text. Given the

Large language model (LLM) post-training is essential for improving reasoning, adaptation, and alignment. Existing methods mainly follow two paradigms: reinforcement learning (RL) and on-policy distillation (OPD). However, RL relies on coarse-grained outcome supervision, resulting in difficult credit assignment and limited capability to acquire new knowledge. OPD, meanwhile, unconditionally matches teacher logits through KL divergence, which creates a dilemma: similar teachers provide little new

Solar Open 2 is Upstage’s 250B-A15B open-weight large language model, built for agentic use cases such as office productivity, document-intensive work, and coding. Its Hybrid-Attention Mixture-of-Experts (MoE) architecture with linear attention delivers highly efficient inference even in long-context settings. Agentic Specialist: Purpose-built for agentic workflows — tool calling, multi-step reasoning, and end-to-end task execution. Competitive with the strongest open-weight models on agent benc

Paper: Code (GitHub): Hi everyone, I just published a preprint on a new optimizer designed to tackle the massive VRAM bottleneck in Mixture-of-Experts (MoE) training. If you've trained MoEs, you know that optimizer state is usually the largest single line item in the memory budget. AdamW, for example, spends 50.6 GB of state memory just to update a 12.6 GB model. I built SkewAdam to fix this by using a tiered state allocation . Instead of treating all parameters equally, it allocates precision b

Google's cloud business is thriving, as companies adopting its AI and AI infrastructure services help the tech giant to report record profits.

The company said it is reducing its headcount by 20%, or about 630 staff, to "support a leaner, more focused operating model" as it focuses on its AI Work Platform.

Diffusion Language Models (DLMs) have recently emerged as a promising alternative to autoregressive models. Unlike standard diffusion-based approaches, DLMs are not explicitly conditioned on a timestep, raising a natural question: do these models internally represent denoising progress, and how is such information used downstream? In this work, we show that DLMs do in fact encode a latent representation related to the diffusion timestep within their residual streams. We find that this signal can

Hey we are Computable. We spent years building trading infrastructure at Jump Trading and Coinbase. From that point of view, compute looks like energy markets before 2000: everything trades through private bilateral leases, there’s no visible price, and nothing can be resold. The same H100 rents at a 2x spread depending on who’s asking, and once you sign a 24-month lease, it can never change hands. So we built a market for GPU nodes, sold by the calendar week. Here are three things you can do on

arXiv:2607.18241v1 Announce Type: new Abstract: Large language models (LLMs) excel at analyzing individual documents but break down on exhaustive, cross-entity analytical questions over enterprise-scale datasets due to context overflow, loss of per-entity attribution, and linear latency from sequential tool calls. We present BatchDAG, a system in which an LLM generates a typed directed acyclic graph (DAG) of operations -- SQL queries, semantic searches, in-memory transforms, parallel fan-outs, a

arXiv:2607.18253v1 Announce Type: new Abstract: Modern language query routers improve inference efficiency by assigning each query to a model that balances response quality and monetary cost. However, current query routers are largely latency-agnostic and do not consider the generation latency experienced by queries at model instances. In practice, latency is often controlled by load-balancing policies such as round-robin or join-the-shortest-queue, which do not account for model accuracy or inf

Several new Cyber headlines make us observe a trend

One thing I noticed in American politics, whenever the government wants to push unpopular actions or laws, they often introduce fear to convince the public to support them. This is actually how i view the recent news about OpenAI’s model breaking out of its sandbox. The whole news i see it as two corporate goals. 1. Scare the public into supporting laws that restrict open-access LLMs under the pretext of "safety". 2. OpenAI is playing catch-up against Anthropic's Claude mythos, using this to dem

Reinforcement learning (RL) on open-ended tasks compresses an LLM's rubric-based evaluation into a scalar reward, discarding rich textual feedback and conflating responses with distinct quality profiles. We propose Experiential Learning (EL), which repurposes the feedback model from an LLM-as-a-Judge into an LLM-as-a-Coach. The coach distills its assessment of each on-policy response into transferable experiential knowledge, which conditions a teacher model and is internalized by the policy thro

13,000+ MCP servers, skills & plugins for AI coding agents Discussion | Link

Gemini 3.6 Flash, 3.5 Flash-Lite, and 3.5 Flash Cyber Discussion | Link

Laguna is the first model I'm trying, Q4_K_M fits with 256K context @ F16. Doing the html flight simulator test now. These cards are getting 400 to 600 tok/s prefill and 16 to 20 tok/s gen so far (I have NOT enabled dflash yet). Not bad at all for the cost. ($350 each) In a Dell PowerEdge R740 with dual Xeon Gold 6248R and 768 GB RAM.

The primary driver of this project is that I'd become frustrated with the reasoning behavior of smaller local models such as Qwen3.6-27B (i believe particularly at lower temperatures, and where system prompts are highly specific), their reasoning process is highly unreliable and often tends to spiral into neverending "But, wait" loops or, occasionally, complete garbage. The core principle is simple - when the sampler sees an opening tag, it kicks off the thought process with a self-aware stateme

Today the Department of Energy (DOE) and Arcee AI announced the development of Genesis-Science-1 (GS1), an open model for scientific research. This is a joint effort to bring advanced AI into scientific research across a wide range of fields. GS1 is an American open-weight model for scientific research, built together with the DOE and its national laboratories through the Genesis Mission. Arcee has secured the compute, will handle training and post-training, build the scientific workbenches and

The episode has also intensified a broader debate in Washington over the influx of Chinese open models.

Substack is giving readers a way to estimate how much of a newsletter was written by AI, signaling a broader shift toward transparency around AI-assisted content.

Even if you are as averse to semver as I used to be in the course of my programming activity, you can still think of open source software distribution as something that used to follow a fixed number of steps. There is a branch where developments happen, and this branch oftentimes happens to be not really ready for reliable work. Then you freeze the developments for a certain amount of time (even if, in the meantime, the work can continue on some new unstable branch), fix bugs, ask people to test

arXiv:2607.18245v1 Announce Type: new Abstract: Exact-match evaluation of agent-calling obscures qualitatively different failure modes: a model may select the right function yet hallucinate argument values, or satisfy a schema while choosing a agent for the wrong reason. Existing benchmarks collapse these distinctions into a single binary score, leaving practitioners unable to diagnose where agent calls fail. We propose SAAG a cascaded diagnostic framework that decomposes agent-calling evaluatio

OpenAI outlines its commitment to advancing American science working with the U.S. Department of Energy and national labs to use frontier AI to accelerate discovery.

arXiv:2607.18246v1 Announce Type: new Abstract: We present Phionyx, a deterministic AI runtime architecture derived from the broader Echoism interaction framework that introduces a governance-first approach to AI engineering: treating large language model (LLM) outputs as noisy sensor measurements rather than direct decisions. Unlike probabilistic agents, Phionyx enforces deterministic state evolution via a structured state vector governed by deterministic state-evolution equations, enabling rep

arXiv:2607.18259v1 Announce Type: new Abstract: Steering vectors (SVs), an inference-time intervention technique for large language models (LLMs), guide the generation process by adding a concept-specific direction vector to intermediate activations during inference. However, existing SV methods frequently yield representation-incoherent behaviors that undermine interpretability and fine-grained control, largely because prior work has focused on binary positive-negative steering evaluation while

arXiv:2607.18242v1 Announce Type: new Abstract: The coming era of autonomous AI agents demands a discovery mechanism capable of navigating millions of tools, yet existing solutions buckle under O(N) complexity and centralized governance. Instead of building another fragile overlay, we propose ToolDNS, a radical framework that retrofits semantic tool discovery onto the Internet's most resilient substrate: the Domain Name System (DNS). By embedding functional intent and organizational trust into a

arXiv:2607.18255v1 Announce Type: new Abstract: Contribution attribution has become a central problem in LLM-based multi-agent systems, where final outputs are produced through multiple agents, message exchanges, and ordered workflow dependencies. Existing attribution methods often rely on counterfactual valuation, such as removing agents or comparing score changes across altered agent subsets. In language-mediated workflows, these methods require repeated model calls, introduce high variance, a

arXiv:2607.18258v1 Announce Type: new Abstract: Reinforcement learning from human feedback (RLHF) with preference-based reward models often exhibits unstable training dynamics. A key contributing factor is that standard RLHF relies on a single sequence-level scalar reward, which is propagated to token-level policy updates and leaves credit assignment within a response inherently ambiguous. Recent work has attempted to address this issue by refining rewards into denser token-level supervision, of

Saw this today and thought its worth sharing. everyone keeps talking about Airbus leaving aws. But i think the more interesting question is why. For years it was just “put everything in the cloud”. Less infra, less problems, less people to manage it. Now it feels like bigger companies are asking a different question. Do we really want all our critical systems under another countries rules? This isnt even about AWS being bad. its is honestly great. Its more about control. If your data, manufactur

Never miss a Claude Code session waiting for your input Discussion | Link

I built Encode Bench , an open benchmark that asks a model to solve a task and return the answer as a Base64 payload. The initial result surprised me: across the eight models with matching data in the current nine-model snapshot, Encode Bench pass rate has a Pearson correlation of 0.91 with the Artificial Analysis Intelligence Index. The correlation with its Agentic Index is 0.94 . That sounds dramatic, so the caveat belongs right next to it: this is a small, imperfect observational sample. It d

Large-scale visual generators are increasingly capable but costly to train, fine-tune, and deploy. We introduce Mage-Flow, a compact 4B-scale generative stack for efficient text-to-image generation and instruction-based image editing. The stack is built from two co-designed components: Mage-VAE, a lightweight high-fidelity latent tokenizer, and a Native-Resolution Multimodal Diffusion Transformer trained with rectified flow matching. Mage-VAE uses one-step diffusion-style encoding and decoding w

Troops received an email informing them that they were rapidly depleting their AI tokens.

There’s a debate going on in the Trump administration over how to handle increasingly powerful Chinese AI models.

ChatGPT allegedly offered "extremely dangerous medical recommendations" regarding a pulmonary embolism.

OpenAI announces Project Camellia in Effingham County, Georgia, with commitments to responsible energy, community investment, jobs, and access to Codex.

I'm the author of a paper my friends and I wrote after we were curious if a MUD, text games originating in the 1970s, could be used to evaluate LLMs. We've spent the last several months on nights and weekends running this experiment and writing the paper on just our personal computers with about $99 in API credits. Our experiment did have an interesting leaderboard but even more surprising was the measurements of each LLM. We scored each on four behavioral dimensions, two of which lean heavily o

Word in web is a pure JS docx editor/renderer benched against MS Word directly. This was largely inspired by Eigenpal going closed source, and some personal frustrations I had with working on complex Word templates, pleading papers, etc... and not having a clean way to view/edit them without owning MS Word (which I did just eventually buy but it sucked). This was also an exercise in highlighting the value of good evals for Agents to bench against. Instead of just throwing the OOXML standard at a

Hello. I really liked stackoverflow >5 years ago. There were many people asking about easy-to-medium problems to be solved, and it was a great way for me to learn C/C++/Bash/awk/sed/cmake/Linux/whatever by solving real-life(!) mediocre problems and also helping people in the process and also being criticized and corrected at the same time, from which I learned triple as much. Now stackoverflow is dead. My almost 150k reputation means nothing. Finally they added a "advice" type of questions which

Efficient teamwork typically combines global coordination with parallel execution, a principle not yet fully reflected in unified Vision-Language Model (VLM)-based document parsers. Existing unified parsers process an entire page jointly but generate its output through a single token-by-token autoregressive trajectory, creating a sequential bottleneck that grows with document length. Such full-page sequential generation overlooks a key property of document parsing: layout must be analyzed global