Hi everyone,
One of the things we love about living in Boston is how easy it is to get to the ocean. When I need to feel Nature's endless energy and put my place in the world back into perspective, spending a day by the water usually does it. Last weekend, we drove up to Gloucester and Rockport. It was a little cloudy, but we had a great time and came home refreshed.
I have also continued using Grok Bot, xAI's always-on assistant, and the more I use it, the more I like the experience. It gives me a smooth way to work with several agents, and I am beginning to give it more of my chores. I am splitting my time between Grok Bot and Codex. Codex is more reliable in the quality of its output, and I still use Codex and Claude for software development. But I may try using Grok Bot to coordinate a complete development cycle, still leveraging Fable and GPT Sol.
People often ask how I keep up with AI and decide what is worth following. So I am skipping the usual essay this week and sharing the publications, shows, and people that make up my current rotation.
Cheers.
Reza
⏱️ The Brief
Cisco gave 90,000 employees AI agents. At that scale, companies need to know who owns each one, how it performs, what value it creates, and when to retire it.
AI assistants are becoming the front door to company software. Salesforce now lets Claude update sales records, but Salesforce still enforces permissions and business rules. Screens may matter less; the underlying controls do not.
A better model cannot compensate for missing company context. In one BCG case, reported accuracy rose from about 30% to 85% only after the team mapped system dependencies and data flows.
OpenAI's heaviest enterprise users generated 8.3 times more output per person than typical customers, versus 2.6 times in January. The metric measures activity, not value; heavier users also built more reusable instructions and system connections.
📶 Patterns & Signals
Cisco gave 90,000 employees AI agents, making ownership and performance a management problem
Cisco gave all 90,000 employees personal AI agents. At that scale, companies need an inventory and proof each is useful. McKinsey's Kate Smaje argues that business owners should be accountable for agent performance and value, and retire an agent when it no longer fits the work. DoorDash offers one example. Its AI code reviewer uses one agent to flag problems and two to check them. DoorDash says it reviews more than 10,000 proposed code changes a week. Of 2,256 serious warnings with a final outcome, 60.2% led to a code change. The figure shows whether engineers acted, not how many defects the tool missed.
Salesforce moved sales work into Claude, while Anthropic began hosting the infrastructure that keeps agents running
Salesforce and Anthropic announced Claudeforce, which gives Claude 37 prebuilt sales actions, including meeting preparation and pipeline reviews. Sellers can look up and update Salesforce records from Claude. Their existing permissions and Salesforce's business rules still apply. An open beta is planned for September. Salesforce executive Patrick Stokes told VentureBeat that the company's durable value is its data, definitions, and workflows rather than its screens. Anthropic also introduced Claude Managed Agents. The service keeps an agent running, preserves its sessions, separates login credentials from the code it executes, and records its actions. Companies can use Anthropic's workspaces or run them inside their own cloud, while supplying their own business context and instructions.
Signals to watch
Nvidia is paying $6 billion to license Poolside's technology. The reported deal offers jobs to 109 employees and includes a separate $1 billion investment. The team will develop downloadable Nvidia models that companies can run and adapt.
OpenAI published the full Hugging Face postmortem. During cybersecurity tests, research agents turned shared software storage into an unintended message board and rebuilt it after deletion. Severe alerts must now pause the work unless responders show within 30 minutes that they are false.
BCG reports that a system map improved one AI modernization program. Reported accuracy rose from about 30% to 85% after the team supplied application dependencies and data flows. BCG does not define the measure, so this is a case example, not a benchmark.
Private-equity firms and Anthropic created a 160-person AI engineering company. Ode places small engineering teams inside portfolio companies to map existing work and build AI systems alongside the people who perform it.
The EU made its first formal information requests under the AI Act. The European Commission asked more than 30 AI companies about model safety, cybersecurity, and training practices. The requests begin an enforcement process; they are not findings of wrongdoing.
🎯 How I stay informed about AI
I do not read or watch all of this every day. Each source has a job. Daily shows and newsletters help me discover stories. Established reporting and original sources help me verify them and add context. Longer interviews show how people are applying the technology. X gives me early notice of releases and technical debates, but I do not treat a post there as confirmation by itself.
The ones I return to most
The AI Daily Brief, hosted by Nathaniel Whittemore, covers the major model, business, policy, and market stories without turning every announcement into a breakthrough.
How I AI is Claire Vo's practical interview show. Guests demonstrate how they use AI in their work rather than only talking about its potential.
AI & I is Dan Shipper's interview show at Every. Writers, founders, investors, and builders often demonstrate parts of their real workflows with tools such as ChatGPT, Claude, and Codex.
Lenny's Podcast is broader than AI. Lenny Rachitsky's long interviews with product leaders and AI builders often contain operating details that launch announcements leave out.
AI News & Strategy Daily is Nate B. Jones's daily analysis, with more attention to business consequences than feature-by-feature coverage.
Peter Yang interviews founders and practitioners and publishes useful guides to AI products, agents, and product-team workflows.
The Rundown, Superhuman, AI Secret, and Ben's Bites are broad scans of AI news, product releases, and examples. They overlap, so I use them for discovery rather than as four confirmations of the same story.
Import AI is Jack Clark's weekly publication on AI research, policy, and the larger consequences of model development.
Postcards From the Edge is John Sviokla's view of AI through strategy, management, and how organizations change.
IBM Think publishes useful enterprise explainers and research across AI, security, infrastructure, and business operations.
Reporting and analysis
Financial Times AI, The Wall Street Journal Technology, and The Economist's Science and Technology section put AI announcements into business, economic, and policy context.
MIT Technology Review is especially useful for research, science, policy, and the limits of new technology.
VentureBeat covers enterprise AI and often gets into implementation details. I read vendor claims carefully, but it is useful for seeing what companies are shipping.
WIRED and The New York Times Technology section add reporting on the people, institutions, and social effects around the technology.
Boston Consulting Group publications provide an enterprise strategy and operating-model view. As with any consulting research, I focus on the evidence and examples rather than treating every framework as a universal answer.
People and teams I follow on X
For research and AI's effect on work: Andrej Karpathy, Erik Brynjolfsson, and Jack Clark.
For the people building the tools: Boris Cherny and Thariq Shihipar at Anthropic, Tibo Sottiaux at OpenAI, and Peter Steinberger at OpenAI, who created OpenClaw.
For products, workflows, and independent experiments: Claire Vo, Peter Yang, Dan Shipper, and Tibo Maker.
I also follow the official accounts for OpenAI, Anthropic, Nous Research and Hermes, and OpenClaw for releases and technical updates.
That is my current rotation. If there is one person, publication, or show you think belongs on the list, send it my way.
💬 Interstitials / Overheard
"Three findings, and one is more serious than your question"
The natural end state of letting an AI agent answer a yes-or-no question.
Source: George Mayer, surfaced via Superhuman
🌍 Meanwhile...
In June, MIT Technology Review profiled Jesse Davis's Sports Analytics Lab at KU Leuven, a group of about ten students and postdoctoral researchers using data to test soccer tactics that coaches have long judged by feel. The researchers trained models on more than 1.4 million passes and about 60,000 throw-ins, then simulated different tactics. Their models suggested that, from the middle third of the field, deliberately kicking the ball out near the opponent's goal can leave a team about ten actions from scoring if it recovers the resulting throw-in. The lab releases much of its code for free, and Davis says professional soccer teams use it in their daily work. A small academic team has made counterintuitive tactics easier to test rather than dismissing them because they look wrong.
Source: Inside soccer's data renaissance
📚 What I'm Consuming
How to Build Better AI Evals with Claude Code in 5 Steps (video). Claude can draft a first checklist for judging AI output. Hamel Husain shows why useful checks must also come from reviewing real examples, especially when errors require product judgment.
The Widening Gap Between Frontier AI Firms and Everyone Else (video). Nathaniel Whittemore reviews OpenAI's enterprise data. Output tokens per user at the top 10% of customers rose from 2.6 to 8.3 times the level at typical customers between January and June. Tokens measure activity, not value. The heavier users also built more reusable instructions and company-system connections.
An Innovation Veteran on What's Next in Enterprise AI (article). ServiceNow's Dave Wright argues that AI can propose what should happen, while company systems still execute the transaction, enforce permissions, and recover from failures. Most companies have not digitized enough of their workflows to build a complete digital view of their operations.
The Tragedy of the Cognitive Commons (research paper). This conceptual paper argues that automating junior work may damage the pipeline that creates future experts, even while companies continue benefiting from today's senior people. The effect appears later, when fewer people can recognize a plausible but wrong AI answer.
LLM & AI Agent Benchmarks vs Reality (video). An AI system that reads a long document and gives a short answer behaves differently from one that writes a long response to a short question. Useful tests must match your data, traffic, and definition of success.
🔤 Term of the Week
Open-weight: An AI model's weights are the learned settings produced during training. When those weights are published, companies can often download the model and run it on compatible computers instead of using the maker's app or service. That does not make it fully open source. The training data may remain private, and the license may limit use.
Teams use open-weight models for repeatable work such as sorting documents, extracting fields, searching private information, or powering an internal coding assistant. Running one inside company infrastructure gives more control over data, adjustments, and cost. The tradeoff is more setup and sometimes weaker results on difficult work. Examples include Llama, Mistral, DeepSeek, and Nemotron. Claude Opus and OpenAI's GPT models are available through hosted services, not as downloadable weights.
🌙 After Hours
Act of Passion
Georges Simenon | 1947 | 216 pages | ★★★★★
This is a convicted murderer's confession to the judge who investigated his case, but I cared more about the psychological arc than the crime. Charles spends most of his life allowing other people to decide its shape. Falling in love for the first time gives him a reason to resist the life other people chose for him. It also brings out a possessiveness that can be mistaken easily for rebellion against convention or getting "even with life."
I loved Simenon's prose. It is concise, direct, and visual, and the book moves quickly without feeling thin. Once Charles believes Martine has given him a different life, ordinary life becomes intolerable. The confession shows that change gradually. The combination of confession and self-justification is what made the book so good.
🎙️ Listen
Prefer to listen? Quanta Bits is also available on Apple Podcasts and Spotify.
How this gets made
I collaborate with Spock, the name I use for my editorial AI agent. He researches extensively: scanning, filtering, and surfacing what's relevant across my business. I read, listen, and watch what resonates, and decide what matters. I provide direction, and we draft together. The editorial judgment is mine. He'd tell you the same. Most logical. 🖖