On July 28, 2026, the Model Context Protocol specification was updated to version 2.0, introducing a stateless design that collapses the two‑request initialization flow into a single HTTP call. This update inspired Simon Willison to create mcp‑explorer and datasette‑mcp tools. The stateless approach removes the need for server‑side session storage, making MCP easier to scale in web applications and usable by smaller language models that run on laptops. It renews developer interest in a standardized tool‑calling interface for LLM agents. The new spec replaces the initialize + tools/call two‑step flow with a single POST that includes MCP‑Protocol-Version: 2026-07-28, Mcp‑Method: tools/call, and Mcp‑Name headers, while client information is placed inside the JSON‑RPC _meta field. Simon built three implementations this week to demonstrate the simplicity.
A postmortem details a soundness bug in Lean's kernel (
In July 2026, Canada signed the United Nations Convention against Cybercrime, a move that critics warn could expand state surveillance powers under the guise of fighting cybercrime. The signature signals Canada’s willingness to cooperate on cross‑border cybercrime investigations, but it also raises alarms about potential overreach that could threaten digital privacy and civil liberties both domestically and internationally. Although 76 states had signed the treaty by May 2026, it will only enter into force after 40 ratifications, and Canada’s signature does not yet bind it legally; critics point to ambiguous language that could enable broad data‑sharing and surveillance measures.
The article 'The Silicon Valley Founder Meat Grinder' critiques the prevailing founder culture, arguing that the chase for wealth and status drives burnout and misaligned motivations among entrepreneurs. This critique highlights growing concerns about founder mental health and the sustainability of startup ecosystems, urging investors and peers to prioritize well‑being over pure financial gain. It cites anecdotes such as a founder’s financial recklessness exemplified by home brewing, and references community comments that stress persistence over intelligence and the dangers of identity‑driven entrepreneurship.
On July 31, 2026, DeepSeek unveiled DeepSeek-V4-Flash-0731, a 304‑billion‑parameter model (≈167 GB on Hugging Face) featuring substantially enhanced agentic capabilities. It is priced at $0.14 per million input tokens and $0.27 per million output tokens, and Artificial Analysis ranks it ahead of larger models such as the 428B MiniMax M3 in cost‑efficiency. The model’s high intelligence‑to‑cost ratio makes it the current best value‑per‑intelligence offering, potentially lowering the barrier for deploying advanced agentic AI applications. Its strong performance challenges the notion that larger models are always needed for top‑tier reasoning, influencing pricing and model selection across the LLM ecosystem. DeepSeek-V4-Flash-0731 is described as an MIT‑licensed MoE model with 284B total parameters that activates around 13B parameters during inference, enabling efficient computation. Default reasoning level produced a poor pelican‑bicycle illustration, while setting reasoning_effort to high yielded a markedly better output, showing the model’s sensitivity to reasoning configuration.
ByteDance has released Seedance 2.5, a new video generation model that enables one-take creation of up to 30‑second clips and supports flexible referencing across text, image, video and audio inputs. The model advances generative video AI by delivering longer, coherent clips in a single pass and allowing creators to maintain visual consistency using multiple references, reducing post‑production effort. Seedance 2.5 can generate high‑quality audio‑video clips up to 30 seconds long in a single pass, supports multi‑round extensions, and accepts up to 50 multimodal references (text, image, video, audio) for flexible referencing.
The article analyzes how Google's 2013 shutdown of Google Reader and related actions contributed to the decline of RSS feed adoption, pushing users toward centralized platforms. It illustrates how decisions by major platforms can undermine open web standards like RSS, affecting user autonomy and content distribution, and highlights the ongoing tension between convenience and decentralization. Google Reader had millions of active users before its July 2013 shutdown; Mozilla removed Live Bookmarks RSS support in Firefox 64 (2018); despite the decline, niche communities still use RSS and alternatives like Feedly have emerged.
Diátaxis offers a structured framework for technical documentation, organizing content into tutorials, how‑to guides, reference, and explanation.
No Starch Press has released a second edition of Randall Hyde’s 'The Art of 64-bit Assembly', an nearly 800-page guide to modern x86-64 assembly language programming using MASM on Windows. The release highlights ongoing interest in low‑level programming for performance‑critical and security‑sensitive software, and the lively Hacker News discussion shows that assembly remains a relevant topic for systems developers. The book focuses on Microsoft Macro Assembler (MASM) for Windows, covers macro invocation arguments, loops, and string processing, while some readers criticized the marketing copy, the choice of MASM over GNU Assembler, and apparent AI‑generated text in the preview.
NetBSD 11.0 introduces a new MICROVM kernel configuration that enables boot times of about 10 ms on modern x86 hardware, and it includes improvements to the npf firewall such as layer‑2 and user/group filtering. The release also brings various hardware support updates and security fixes. The MICROVM kernel targets fast‑boot, lightweight virtualization workloads, positioning NetBSD for edge and container‑like use cases, while the npf enhancements improve its competitiveness as a high‑performance, extensible firewall. Together they show ongoing innovation in the BSD ecosystem, attracting interest from developers seeking alternatives to Linux. The MICROVM kernel strips out most device drivers and runs them in user space, achieving a minimal footprint and sub‑10 ms boot on an AMD Ryzen 7 5800X; npf now supports layer‑2 filtering (e.g., MAC addresses) and user/group‑based rules in addition to traditional IPv4/IPv6 stateful inspection.
Ripgrep compiled with musl libc occasionally crashes with a SIGSEGV when performing extremely large, concurrent directory searches, as reported in GitHub issue
Pgtestdb introduces a template‑cloning technique that creates a clean Postgres database for each test in roughly 100 ms by cloning a pre‑migrated template DB, eliminating per‑test migrations and teardown. Faster test database provisioning cuts CI wait times, allowing teams to run more frequent, realistic integration tests instead of relying on mocks or heavyweight containers. It uses PostgreSQL's CREATE DATABASE … TEMPLATE command to clone a fully migrated template database, works with plain Docker or Testcontainers, and reports setup times around 100 ms per test.
CISA issued an alert warning that threat actors are targeting programmable logic controllers (PLCs) used in water treatment and distribution systems, urging organizations to harden their operational technology (OT) environments. The alert highlights the growing risk to critical water infrastructure, where successful compromise of PLCs could disrupt safe water supply and threaten public health. CISA recommends mitigations such as network segmentation, disabling unused services, applying firmware patches, and monitoring for anomalous PLC traffic, noting that many water sector PLCs run legacy software with limited security updates.
The blog introduces explorative modeling, a training scheme that generates K candidate outputs and selects the best one according to a loss, integrating winner‑take‑all selection into diffusion and flow‑based generative pipelines. By explicitly handling multimodality during training, the approach can reduce the number of inference steps needed for high‑quality generation, potentially making diffusion models faster and more versatile for applications such as image and video synthesis. Training requires K forward passes to produce the K guesses, after which the best guess is selected and used to compute gradients; this adds computational overhead but can enable end‑to‑end generation in few steps. A noted limitation is that the current implementation samples all K modes with equal likelihood rather than proportionally to their true probabilities.
Cursor accidentally removed dollar cost information from its usage page and CSV export while cleaning up an old feature flag, but the issue was quickly fixed after a team clarification. The removal raised concerns about billing transparency for users who rely on Cursor’s cost tracking. It underscores the need for clear usage data in AI coding assistants to maintain trust. According to a Cursor employee, the CSV export was broken due to cleaning up an old feature flag that also displayed a confusing dollar usage graph. The spending page still shows actual billing, and the CSV export has been restored.
Microsoft open‑sourced Flint, a visualization language that lets AI agents generate charts from simple specs and compile them to backends such as Vega‑Lite, ECharts, or Chart.js. Flint bridges the gap between AI‑generated chart specifications and diverse plotting libraries, potentially reducing the token cost and complexity for LLM‑driven data storytelling. Flint provides over 70 semantic chart types, includes an MCP server for agent workflows, and compiles its intermediate specification to Vega‑Lite, ECharts, or Chart.js backends.
Greg Brockman observed that many OpenAI employees have connected their ChatGPT accounts to Slack, but they dislike receiving help requests from a coworker’s AI agent, even when they would be willing to do the same task if asked directly by that coworker. This insight underscores that people value genuine human interaction and want AI to free up time for relationships rather than insert itself as a communication barrier, highlighting a key consideration for deploying AI agents in the workplace. The quote comes from Greg Brockman, President and Co‑Founder of OpenAI, shared via his Twitter account; it notes that while ChatGPT‑Slack integrations are common, AI‑initiated requests are perceived as impersonal and undesirable.
OpenAI used an internal version of its upcoming Astra model to attempt solutions on ten longstanding mathematical problems, reporting a cost of less than $2,000 per problem using GPT-5.6 Sol token pricing. The results include Lean 4 formalizations, a descriptive paper, and an LLM-generated reasoning walkthrough PDF. If verified, this shows that advanced language models can contribute to solving deep mathematical questions at low cost, potentially accelerating research and supporting the vision of 'big mathematics' where AI handles technical work. It also highlights the growing role of AI in theoretical computer science and pure mathematics. The claims are based on an unpublished internal Astra model, with costs calculated at GPT-5.6 Sol rates; no peer review or details on failed attempts are provided. Supporting artifacts include a GitHub repository with Lean 4 proofs, a PDF paper, and another PDF that reconstructs the model's reasoning traces.
Simon Willison released llm-mcp-client version 0.1a0, an early‑alpha Python client that enables LLMs to communicate with external tools via the Model Context Protocol (MCP). The package is available on GitHub and PyPI. This release gives developers a ready‑to‑use MCP‑compliant client, lowering the barrier to integrate LLMs with file systems, databases, and other tools in a standardized way. As MCP gains adoption among major AI providers, such clients will become essential for building interoperable AI applications. The client is written in Python, distributed as the llm-mcp-client package on PyPI, and raises llm_mcp_client.MCPToolError for MCP errors. It is tagged as version 0.1a0 (early alpha) and its source code is hosted at github.com/simonw/llm-mcp-client.
The UT Austin Radionav Lab published a 2023 paper that characterizes the signal structure and modulation of Starlink’s Ku-band downlink in the 10.7–12.7 GHz band, presenting a blind signal identification technique and detailing the embedded synchronization sequences. Understanding Starlink’s downlink signal structure aids RF and satellite engineers in interference analysis, signal detection, and potential navigation applications, while contributing to the broader knowledge of non‑geostationary satellite constellations. The paper describes an OFDM‑based signal with specific subcarrier spacing, reveals the exact values of synchronization sequences that can be exploited for pseudorange measurements, and provides a simulator for generating realistic Starlink Ku‑band downlink signals.
Researchers used TruffleHog to scan 7.6 petabytes of publicly available Hugging Face training data, uncovering 787 live Hugging Face tokens, including 237 with write access and 70 with org-admin privileges. The findings highlight widespread secret leakage in AI training datasets. Exposed tokens can allow attackers to inject malicious models or poison datasets, compromising the integrity of AI systems and exposing organizations to supply‑chain risks. The discovery underscores the need for better data hygiene and secret scanning practices in machine learning pipelines. The scan was performed with TruffleHog, an open‑source secret scanner, across 7.6 PB of data, revealing 787 live Hugging Face tokens, of which 237 grant write access and 70 provide organization‑admin rights. Such tokens, if committed to public datasets, could be abused to push malicious model weights or alter training data at scale.
Kaisel is a new Dart 3‑native Flutter router that treats routes as first‑class values using sealed types and pattern matching, eliminating string‑based paths and code generation. By leveraging Dart 3’s records, patterns, and class modifiers, Kaisel offers type‑safe navigation that integrates naturally with Flutter’s reactive framework, potentially reducing runtime routing errors. The router uses a stack‑as‑state model, sealed route classes, and Dart 3 switch expressions with patterns to match routes, requiring no string paths or build‑time code generation.
Datasette‑apps version 0.2a0 was released, introducing two new tools: app_debug() which lets the Datasette Agent test apps inside an invisible iframe, and app_list() which lists the apps a user can edit so the agent can modify them. These tools improve the developer workflow for the Datasette Agent by enabling automated testing and editing of Datasette Apps without manual intervention, streamlining app development and debugging within the Datasette ecosystem. app_debug() renders the target app in an iframe with CSS opacity:0 and pointer-events:none, then runs agent‑provided JavaScript via the new context.browser_task() mechanism introduced in datasette‑agent 0.4a0; app_list() returns a list of editable apps for the current user.
The article describes a project that successfully boots a minimal Linux kernel on a TI‑84 Plus calculator, outlining the steps taken to adapt the OS to the device’s Z80 processor and constrained memory. It demonstrates that even very limited 8‑bit platforms can run a Linux‑based system, offering insights for embedded developers and the maker community about pushing hardware boundaries. The hack relies on porting U‑Boot to the TI‑84 Plus’s Z80 processor, compiling a tiny Linux kernel for the Z80 architecture, and fitting a minimal userland into the calculator’s approximately 128 KB of RAM.