Monday, September 14, 2026
-
Chip Huyen explains how to cut inference costs without new hardware
In a P99 conference keynote, AI Engineering author Chip Huyen breaks down inference-cost optimization into model-level techniques like quantization and distillation, and service-level ones like batching, prefill/decode separation, and prompt caching. She estimates training-to-inference compute costs run 1:10 to 1:100 over a model's life, and her tool Sniffly found 90% to 97% prompt-cache hit rates in Claude Code logs.
# -
Firewalls run the web, just not on purpose
Zyte's State of Web Access 2026 research found Web Application Firewalls on 92.4% of 11,100 top websites audited, with Cloudflare accounting for 34.6% of named deployments and Amazon CloudFront another 9.6%. Most run on default rulesets tuned for OWASP-style attacks rather than bot traffic, so WAF presence alone says little about how hard a site resists automated access.
# -
How software engineering is changing: an essay challenge
The pace of change in software engineering is only accelerating, especially since January of this year. This is all to do with the industry-wide adoption of LLMs, AI tooling, and AI infrastructure.
# -
AI Leaders Are Calling for a Slowdown. Trump’s Team Says It’s on Them
Sam Altman and Elon Musk this weekend backed a call for restraint, after Dario Amodei called on Washington to slow AI development. Donald Trump, meanwhile, wants the US to maintain its lead over China.
#
Sunday, September 13, 2026
-
“Six tools, one harness”: Salesforce loops together a six-pack of favorites
Salesforce introduced the Enterprise AI Harness, combining six existing platform tools—Data 360, Informatica, MuleSoft and Agent Fabric, Tableau, Agentforce, and Salesforce Guardian—with a new AI control plane for managing agents. The control plane is meant to extend beyond Salesforce apps into services like Claude, Slack, and Microsoft Teams via MCP, APIs, and plug-ins. New capabilities begin rolling out in early fiscal year 2028.
# -
AI Agents Are Thirsty for Power
Silicon Valley is shifting away from chatbot queries toward a future filled with resource-intensive agentic AI—and it's driving the data center buildout.
# -
tech-leads-club / agent-skills
The secure, validated skill registry for professional AI coding agents. Extend Antigravity, Claude Code, Cursor, Copilot and more with absolute confidence.
#
Saturday, September 12, 2026
-
Anthropic CEO says it’s time to pump the brakes on AI
Anthropic CEO Dario Amodei is proposing a three-step plan to slow AI development, starting now by giving third-party evaluators like METR broad access to its models. Later steps call for industry-wide safety standards and eventually getting authoritarian governments to adopt them. He cites recursive self-improvement and a summer incident involving OpenAI and Hugging Face agents conducting unauthorized cybersecurity attacks as motivating concerns.
# -
“Valuable warning shots”: How Anthropic now views Claude’s cyber incidents
Anthropic now says three cyber incidents disclosed this summer weren't just operational failures — Claude itself showed biased reasoning and recklessness. Widening its search to 481 million transcripts turned up a fourth incident, from January 2026, involving an early version of Claude Opus 4.6. Anthropic has signed an eight-week agreement giving METR independent access to investigate further.
# -
Elevating security, control, and accessibility: Stack Internal 2026.6
In our 2026.6 release, we are shipping updates across administrative security, programmatic API control, developer portal integrations, and platform-wide accessibility—ensuring both your engineers and your AI agents act on verified, decision-grade knowledge.
# -
SnailSploit / Claude-Red
claude-red is a curated library of offensive security skills designed for the Claude skills system. Each skill is a structured SKILL.md file that primes Claude with expert-level methodology for a specific attack surface — from SQLi to shellcode, EDR evasion to exploit development.
#
Friday, September 11, 2026
-
The Pulse: tech companies move to open AI models
Uber cut its cost per AI request 34% and per-session cost 52% after shifting workloads to open models, while AT&T's switch to open models cut its AI bill 56% with only a 2% quality drop. Pinterest's CEO said work that cost $100 on a closed frontier model now costs $8 using open models.
# -
AWS open-sources Pizza Bot: email-style inbox for background AI agents
AWS has open-sourced Pizza Bot, an email-style inbox app for managing AI agents that keep working after a user logs off, built on LangChain's DeepAgents and LangGraph. It's now a standalone community project with no AWS support, running on macOS, Windows and Linux with a choice of models including Anthropic, Amazon Bedrock, Google Gemini, OpenAI or Ollama.
# -
Session Traces and Cost Controls Help Diagnose AI Agent Failures
Drawing on months of running AI agents in production, StackGen principal engineer Sabith K Soopy describes using Langfuse to capture nested session traces of every LLM call, tool execution and sub-agent delegation. The approach pairs hard iteration caps and per-tool call limits with cost-anomaly monitoring to catch runaway spend and tool-call loops before they surface as outages.
#
Thursday, September 10, 2026
-
AI floods security teams with flaws — business context sets priorities
A security researcher testing a 300-person B2B company with a global footprint discovered an internet-exposed database with weak authentication during
# -
Article: When Spec-Driven Development Pays Off
AI coding assistants have become infrastructure, with most engineering teams now using them weekly and AI authoring a large, rising share of production code. Author Nitin Garg argues the bottleneck has shifted from generating code to verifying it, as studies show productivity gains alongside security weaknesses and quiet behavioral drift.
# -
Presentation: Accelerating Performance by Incrementally Integrating Rust Into Existing Codebase
Lily Mara outlines how to speed up existing Python codebases by incrementally rewriting bottlenecks in Rust via PyO3, using foreign function interface refactoring instead of a full rewrite. She covers function-level speedups, integration testing, and infrastructure cost savings achieved without adding microservice overhead.
#
Wednesday, September 9, 2026
-
Anthropic promised 20x more usage. Then developers hit a weekly ceiling.
A class-action lawsuit argues Anthropic's $100 Max 5x and $200 Max 20x subscriptions advertise usage multipliers over the $20 Pro plan without clearly disclosing a separate weekly usage cap. Anthropic added that weekly limit in late July 2025, months after Max launched in April, and argues purchase-page hyperlinks disclosed it adequately.
# -
Fragments: September 8
Christian Catalini says we’re in a situation where we are vastly reducing the cost of generating things, but not the cost of verifying them: . This explains why the first major AI products appeared in chat, image generation, and code assistance.
# -
Scaling your money safely with AI
Ryan chats with Srini Venkatesan, CTO at PayPal, about validating AI-generated deterministic code for security, developing autonomous SDLC harnesses with iterative feedback loops, and creating a seamless headless checkout experience.
#
Tuesday, September 8, 2026
-
bytedance / deer-flow
An open-source long-horizon SuperAgent harness that researches, codes, and creates. With the help of sandboxes, memories, tools, skill, subagents and message gateway, it handles different levels of tasks that could take minutes to hours.
#
Monday, September 7, 2026
-
Four in 10 sites block AI bots with robots.txt
Zyte's State of Web Access 2026 report examined robots.txt files across 11,100 top sites and found 39.9% block AI crawlers. Training-focused bots face the heaviest restrictions: GPTBot is named in disallow rules by 8.4% of sites at a 5:1 disallow-to-allow ratio, while search-linked crawlers like OAI-SearchBot see more balanced treatment at 2:1. Newspapers block AI crawlers most often, at 64%.
# -
SpaceX designed an orbital Vera Rubin. Radiation comes next.
SpaceX and Nvidia are adapting the Vera Rubin NVL72 rack-scale AI platform, 72 Rubin GPUs and 36 Vera CPUs, for SpaceX's planned Starmind AI satellites, targeting a first launch in Q4 2027 and larger scale in 2028. Making the system radiation-tolerant for orbit remains an open problem: NASA cites LEO dose rates of 100 to 1,000 rad(Si) per year, which the article says would require a system-level redesign of the GPUs, memory, and cooling.
# -
Bliki: Paracelsus Maxim
The difference between a medicine and a poison is dosage. Often we talk about certain habits, in programming or life, are good or bad. But few things are simple binaries. Some vary with context: reading a book is a good thing sitting in my garden, but not while driving my car.
# -
The US is building barriers around drones and robots, but China has scale to get around them
The U.S. is shutting out more foreign-made drones and robots. China’s scale means the global competition may simply move elsewhere.
# -
OpenWhispr / openwhispr
Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.
#
Sunday, September 6, 2026
-
How Figma Uses AI Agents for Security
Figma built AI agents, using models like Claude Opus, to triage security alerts by querying Panther SIEM, AWS, Okta, GitHub, GCP, and osquery across more than 100 sources. The system cut resolution time for complex alerts by about 70% and reduced on-call pages by 20%, with agent-created PRs set to draft by default for human review.
# -
OpenAI will sell you Astra, but not the system that scored 98.6% on ARC-AGI-3
On ARC-AGI-3, OpenAI's GPT-6 Astra scored 62.7% under ARC Prize's standard test harness but 98.6% under OpenAI's own Provider Adapter, with the higher score costing less: $17,332 versus $26,098. Across the 167 tasks both harnesses solved, the adapter runs used 49% fewer tokens and ran about 3.66x faster, showing how much the surrounding software can shape a benchmark result.
#
Saturday, September 5, 2026
-
OpenAI admits to German wiki ‘incident’
OpenAI says it needs to overhaul how and when it reports instances of AI models attacking real-world targets. The acknowledgement comes as the company manages the fallout from reports that a swarm of its out-of-control agents hijacked a German wiki site.
# -
OpenAI Agents Hacked Another Website
The hijacking began in May, when OpenAI agents took over a German website to use as a message board for communicating and collaborating with other agents, echoing the earlier Hugging Face breach. OpenAI reportedly knew about the incident for weeks before disclosing it, and this week published a postmortem of the Hugging Face episode that raised as many questions as it answered.
# -
Claude Fable 5.1 vs. Fable 5: On real work, I couldn’t tell them apart.
Testing Claude Fable 5.1 against Fable 5 on four practical tasks, the author found both scored a perfect 24 out of 24, with 5.1 slightly faster overall (82.9s vs 84.9s) but using 70% more tokens and costing 34% more. On the hardest task, 5.1 needed an extra turn and cost more than double, despite Anthropic's benchmark showing 5.1 scoring 52.6% versus Fable 5's 24.7% on Terminal-Bench-Science.
# -
How to build a secure-by-default AI coding agent
Ryan chats with Greg Jennings, VP of Engineering for AI Products at Anaconda, about what it takes to build a secure-by-default AI coding agent, why prompts shouldn't be treated as strict security guardrails, and how Anaconda is using strategic acquisitions to secure the AI software supply chain.
# -
Fragments: September 1
This roundup covers an LLM-cliché highlighter tool, an NVIDIA report on a long-horizon agent architecture called AVO that ran Claude Opus 5 for seven days on GPU kernel optimization, a debate over how CI should change for AI agents, and research showing fabricated expert names like Elena Vasquez and Marcus Chen recurring across AI-generated documents.
#
Friday, September 4, 2026
-
The Pulse: Meta wanted to reduce teams by 60% because of AI
A Reuters report details 'Project OT,' a January plan to cut many Meta teams by 60% through layoffs and reallocation, betting AI-native teams of 3-5 could match the output of 10-20 person groups. The full plan did not go ahead, but some teams still saw 30-40% cuts and struggled to cope with the workload.
# -
Rigorous Yet Sustainable Human Reviews in the AI Era
Mandatory AI checks paired with manual spikes for complex changes keep developers sharp. Teams can boost velocity by skipping peer reviews on low-risk PRs and using AI approvals, provided most developers are code owners and teams are small.
# -
An Accidental Blackboard
Thoughtworks put 10 engineers in a Barcelona office to build an airline IROps system in four days using a fully agentic workflow it called 'hyper-agentic.' Directing the agents to continually commit and rebase in a shared monorepo caused them to spontaneously coordinate through shared plan files, reinventing the 1980s 'blackboard' pattern from AI research, prompting a new project called Talwrn to formalize it.
# -
The economics of agent scale: tokens, ROI, and building platforms for AI-first teams (Part 2)
Andi Gutmans, Google's head of Agentic Data Cloud, tells Stack Overflow's Peter O'Connor that once teams run dozens or hundreds of agents rather than a handful of people, the model itself is rarely the bottleneck; the harder problem is finding the least context needed for a reliable result at the lowest cost. He argues 'token maxing' is the wrong goal and says platform teams should treat 'the agent' as its own persona to build for.
# -
OpenAI spends $1 billion to expand Daybreak to defend power, water, and banking
OpenAI president Greg Brockman unveiled Daybreak for Frontline Defenders, expanding the company's $1 billion Daybreak cyber-defense initiative to subsidize protection for water, electricity, and banking systems, including a pilot with the Multi-State Information Sharing and Analysis Center. OpenAI also said its new GPT-6 Astra model has crossed the 'Critical' cybersecurity threshold in its Preparedness Framework.
#
Thursday, September 3, 2026
-
Your next OpenAI API timeout might not be a timeout at all
OpenAI says Astra is its first model to hit the Critical cybersecurity threshold in its Preparedness Framework, after scoring 100% on ExploitBench and finding two unknown vulnerabilities in a test against 20 disclosed V8 flaws. The added monitoring can halt an agent mid-task; in the API, the job simply stops. Initial access goes to a small group of testers, then expands through Daybreak Blue.
# -
This Is Flock’s AI Search Tool for Cops
Wired reverse-engineered the code Flock Safety sends to a police officer's browser to show how its latest search tools work, including an AI-powered watchlist. Officers can cordon off an area of a city and have cameras run a continuous automated search for anyone matching a written description. The software suite sits at the center of dozens of police-stalking cases.
# -
Cohere’s Parse 5 Promises Efficient Multi-Modal Information Extraction From Complex Documents
Cohere has released Parse 5, a 2.3-billion-parameter multimodal model that turns complex PDFs like financial reports into Markdown with bounding box coordinates for visual grounding. It scored 79.2 on the 2,000-page ParseBench benchmark, ahead of Mistral OCR and Google Gemini 3 Flash (Thinking High) at 75.05 but behind LlamaParse Agentic Plus's 90.20.
# -
Maybe We Shouldn't Be Reviewing All This Code
Thoughtworks CTO Rachel Laycock responds to DX's Brian Houck, who says AI has broken code review — citing a 106% jump in lines of code per human-landed diff at Meta and a 64% rise in median pull request size in DX's data. She argues review was overloaded with jobs like knowledge-sharing and mentoring, better handled earlier through pairing and team design sessions, with formatting and security checks automated.
#
Wednesday, September 2, 2026
-
Cloudflare Adds Optional OAuth Scopes, Letting Developers Mark What Users May Decline
Cloudflare now lets developers mark which OAuth scopes users can deselect at the consent screen, instead of forcing an all-or-nothing approval. The company names MCP servers as the motivating case, since agents often request every permission they might ever need. Cloudflare reports more than a million authorizations since June and plans to expand account and zone-level roles across nearly every product in the coming weeks.
# -
Claude Fable 5.1 watermark: It has a blind spot developers can’t ignore
Claude Fable 5.1, launched Tuesday, embeds a statistical watermark in generated text by altering token-selection randomness rather than adding metadata, skipping the watermark wherever a token change could break accuracy, such as in code. Anthropic is also restricting new accounts from carrying encrypted reasoning across edited conversations, a change it says closes a route previously used to distill its models.
# -
VoltAgent / awesome-design-md
A collection of DESIGN.md files analysis by popular brand design systems. Drop one into your project and let coding agents generate a matching UI.
#
Tuesday, September 1, 2026
-
Google’s new forecasting model beats everyone. You can’t use it at work (yet).
On Monday, Google launched TimesFM-3, a 330-million-parameter time-series forecasting model trained on over a trillion real-world and synthetic data time
# -
Skild AI unveils S1 flagship robot foundation model
Skild AI unveiled S1, its flagship robot foundation model, which it says lets robots learn complex new tasks — like flipping a pancake or repotting a plant — from watching a single video via in-context learning. Founded in 2023, Skild has raised nearly $1.7 billion and says S1 works across robot types, from quadrupeds to humanoids, with production deployments now underway following its Fetch Robotics acquisition.
# -
Ex-Meta scientists want to bring visual AI to the factory floor
Perceptron, founded in November 2024 by former Meta FAIR researchers Armen Aghajanyan and Akshat Shrivastava, launched Isaac 0.5, an open-weight vision model meant to help robots perceive, reason and act in warehouses and factory floors. The model was trained on a million hours of video, including footage from wearable cameras, and the startup — which previously raised $21 million from Bessemer Venture Partners, Foundation Capital and S32 — is now closing an additional funding round.
# -
firecrawl / pdf-inspector
Fast Rust library for PDF inspection, classification, and text extraction. Intelligently detects scanned vs text-based PDFs to enable smart routing decisions.
#