AWS Makes Web Search Grounding Generally Available on Bedrock

On August 4, AWS announced the general availability (GA) of "Web Search" on Amazon Bedrock, which grounds model responses in up-to-date web knowledge. Its key feature: it can be enabled with a single API call parameter—no need to separately integrate third-party search vendors, orchestrate external APIs, or clear additional security reviews.

The feature combines a billion-scale web index that Amazon operates itself with a knowledge graph that preserves entity relationships. For fact-heavy queries, high-confidence answers are drawn from the knowledge graph, and search results are passed to the model as semantic snippets optimized for its context. The design reportedly returns grounded answers with source citations in a single round trip while keeping latency down.

It can be called via an endpoint compatible with the OpenAI Responses API, and calls can be audited through CloudTrail. Query text, retrieved URLs, and page contents are not logged, however, and are treated the same as prompts. Data egress is zero by default, and processing stays within the region. For enterprise AI, this looks like a step toward reducing the operational burden of building RAG.

OpenAI Adds Education Plugins to ChatGPT Work and Codex

Also on the 4th, OpenAI announced that it has begun offering education plugins for ChatGPT Work and Codex. Intended to help K-12 teachers, university educators, and students with learning, research, and building, the move appears to be part of an effort to establish ChatGPT Work in education.

The same day, tech publication Latent Space published a detailed write-up reconstructing ChatGPT Work's internal structure from the outside. It maps out how Memory, Proactivity, Scheduling, Browser Use, Plugins, Skills, and Tools work in concert, and the design philosophy of "an agent for a billion users" comes through clearly.

Liquid AI Releases LFM2.5-2.6B, a 2.6B-Parameter Agent-Focused Model

Liquid AI has released LFM2.5-2.6B, a 2.6B-parameter model from LFM2.5, its family of hybrid-architecture models designed for on-device operation. It supports a 128K context and has undergone post-training for agent workloads.

Its inference efficiency is striking: 220 tok/s on an Apple M5 Max and 113 tok/s even on AMD Ryzen CPUs, with memory usage staying under 2.5GB. The company claims performance rivaling models 4x its size on tool use, instruction following, and multi-step agent tasks. Another notable point is that reinforcement learning was run inside major agent frameworks to improve compatibility. The release quickly became a talking point in Reddit's LocalLLaMA community, with some seeing it as a small model for "simple, high-frequency tasks" like processing large volumes of documents.

LLMs Struggle with "Deleting Correct Code," CanItDelete Shows

The paper "To Add Is Machine, To Delete Is Human," published on Hugging Face, lays bare the reality that frontier models struggle with work software engineers never even consciously think about—"carving out correct code and putting a stop to it." The authors built CanItDelete, 200 real tasks that involve deletion only, and use it to distinguish failures such as retaining too much, deleting only partially, deleting too much, or getting the boundaries wrong.

The results are harsh. Even when given the lines to delete, GLM-5.2 failed on 12%, GPT-5.6 Sol on 18%, and Qwen3-235B on 42%, while Claude reached 97.7%; meanwhile, some models reportedly swung toward over-deletion as their localization improved.

Even more interesting is the "Guard-and-Go" pattern that shows up in real repository work. Instead of removing old logic, models leave it tucked behind guards and fallbacks—a phenomenon found in 29% of major SWE-bench Verified submissions. Tests may pass, but the code is not of mergeable quality. On the other hand, adding just 0.7% deletion-focused data reportedly reduced incomplete deletions by 13.9 points and improved SWE-bench Verified by 5.3 points. Models need to learn not just writing code but "what should no longer exist," the paper argues.

llama.cpp PR Caches Frequently Used MoE Experts

A PR (#26563) has been proposed for the local inference engine llama.cpp that caches frequently used experts from MoE (mixture of experts) layers in VRAM. Whereas the choice used to be binary—put every expert on the GPU or offload them all to the CPU—the new mechanism keeps only the heavily used "hot" experts in VRAM while cold experts keep running on the CPU.

In the submitter's own measurements running Qwen3.6-35B-A3B on 8GB VRAM, throughput went from 33.25 to 56.0 tok/s (1.68x) at Q2_M and from 17.34 to 35.93 tok/s (2.07x) at Q5_K_P. For the local LLM crowd, it is a notable improvement, suggesting that large MoE models could run at practical speeds even on consumer GPUs.

Pulitzer Prizes See Record Number of AI Disclosures

At the 2026 Pulitzer Prizes, a record eight entries disclosed their use of AI (five of them winners). Outlets including the Wall Street Journal and AP reportedly used LLMs mainly to speed up searching through large troves of documents. At the same time, Pulitzer administrator Marjorie Miller reiterated the policy that using AI to write or edit articles is not permitted. The line between AI as a "search and analysis aid" and AI as "production" in the newsroom appears to be gradually taking shape.

Anthropic Appoints Cuéllar as First Chief Global Affairs Officer

Anthropic announced the appointment of Mariano-Florentino (Tino) Cuéllar as its first Chief Global Affairs Officer, overseeing policy, international strategy, and government relations. A former president of the Carnegie Endowment for International Peace, he has also served as a justice of the Supreme Court of California and as head of a research institution at Stanford University. He was also a trustee of the company's Long-Term Benefit Trust, which he will step down from upon taking the role. With debates over AI regulation and industrial policy intensifying across major countries, the move looks like an effort to build out the organization for stronger engagement with governments.

---

August 4 was a day when three things came into view at once: the tooling of cloud AI infrastructure falling into place (AWS), a small agent model that runs at the edge (Liquid AI), and research that makes a model "weakness" visible (CanItDelete). The infrastructure, model, and evaluation layers each appear to have taken one step forward.