A New Era: More AI Bots on the Web Than Humans

For the First Time, More Autonomous Agents Than Humans Are Browsing the Web

Matthew Prince, CEO and co-founder of Cloudflare, has announced that for the first time in internet history, bots have surpassed human traffic. The ratio of HTTP requests is currently 57.5 percent for machines and 42.5 percent for humans. It's important to distinguish these from classic bots like crawlers, search indexers, or fraud bots: Cloudflare is counting autonomous agents that browse the web on behalf of humans. These agents read product pages, compare prices, book flights, scrape content for AI models instead of search engines, and handle orders or customer service as personal assistants. Cloudflare only began classifying traffic by signed agents and verified bots last year, which is why the charts don't go back very far. In terms of pure time spent in apps, streaming, and endless feeds, humans remain the primary users, as these formats do not generate the rapid succession of page views that automated agents do. → www.tomshardware.com

Synthszr Take: In May 2024, I wrote that a tough selection process would begin as soon as AIs started browsing the web: mediocrity would disappear, and only excellence would survive. Now, the tipping point has been reached, two years earlier than Cloudflare CEO Prince himself expected. When 57.5 percent of all HTTP requests come from agents, it means an agent is reading your product page, checking the price, and comparing flights for its human. It couldn't care less about your pretty hero image. For every brand, this means your site must be cleanly readable by machines, with clear data and pricing—something everyone can add to their backlog tomorrow morning. By the way, Prince is measuring requests; time on site is a different story. We humans are still the main customers when it comes to feed-scrolling. But for comparing, booking, and ordering, the machine is in the driver's seat. Nostalgia for the old click-web won't get us anywhere, so let's build sites that an agent can understand and a human can still love.

Sam Altman: 'Proactive AI' is the Next Big Thing

At an OpenAI enterprise event, Sam Altman outlined a three-phase thesis on AI product development. Phase one was chat models like ChatGPT, phase two was agent-based systems like Codex, and phase three he calls 'proactive AI'—artificial intelligence that constantly runs in the background. 'If I were to prepare for one thing in the next year, it would be this,' Altman said. The agent phase has been the largest category so far, driven by customer demand, but many users are overwhelmed when they have to use chat, Codex, or an API. In parallel, OpenAI is working on a kind of super-app that bundles the agentic functions of Codex and ChatGPT. There's also a usability problem: 'most people' simply don't know how to use AI well; the 'activation energy' is too high. 'Proactive AI' is OpenAI's answer to this as well: if you don't learn to work with AI, automation will work around you. → Techpresso

Synthszr Take: Altman's third phase is, above all, an admission. People aren't using this stuff because the 'activation energy' is too high, as he himself says. Instead of teaching people how to ask, the AI is supposed to just run constantly in the background and make itself useful. Sounds convenient. The only problem: this pushes the second issue right where it hurts. An AI that is connected to the entire company context 24/7 burns even more compute than any chat, and Uber already burned through its entire annual budget in Q1 without continuous operation. We had the cost confession back in February in the Claws Club: folks, it's going to be expensive, very expensive. My advice for anyone who doesn't want to wait a year for 'proactive AI': lower the activation energy yourself. Half a day of real onboarding per team will yield a better digitalization return today than any background automation OpenAI is promising for next year.

Altman and Amodei Want to Regulate the DNA Supply Chain

On June 3, the heads of the largest AI companies signed an open letter calling on the US Congress for more regulation. The demand targets the supply chain for synthetic DNA and RNA, an area outside of their own products. The argument: AI systems are becoming so good at biology that they could make it easier for bad actors to design dangerous pathogens. Specifically, providers should screen every DNA and RNA order against databases of known hazardous sequences, mandatorily verify customers, and conduct a risk assessment before shipping. Since early 2024, OpenAI has been conducting internal red-teaming to test whether its own large language models could assist in assessing biological threats. The results were apparently alarming enough that voluntary industry standards are considered insufficient. Beneficiaries would include manufacturers of biosecurity technology and established players with existing compliance infrastructure. → Techpresso

Synthszr Take: Sam Altman and Dario Amodei agree on almost nothing; their companies are in an existential race. Yet, on the DNA letter, they signed the same line: every order of synthetic DNA and RNA in the US should be screened and tracked. This is a genuine signal, as OpenAI has been internally testing since early 2024 whether its own models could help someone assess biological threats. The results were apparently so uncomfortable that voluntary industry standards are no longer sufficient. What strikes me is where the guardrail is supposed to be placed: on the DNA synthesis providers, while their own language models are left out. This is convenient because screening requirements will push smaller vendors with thin margins out of the market, ultimately benefiting the big players with ready-made compliance infrastructure. Nevertheless, the move is the right one: the DNA supply chain is the one physical bottleneck where a digital risk can be tangibly controlled.

DeepSeek is Showing Up on the Invoices of American Corporations

In the Ramp report for June 2026, DeepSeek is at the top of the list of 'Trending' software providers, which measures growth relative to size. The crucial factor behind this placement is the usage pattern: companies are paying DeepSeek directly and sending their data directly through DeepSeek's systems. This sends a different signal than the adoption of self-hosted open-weight models. Back in January, DeepSeek briefly reached 0.3 percent business adoption but then fell back to 0.1 percent. This time, there's a tangible driver: in May, DeepSeek made the 75 percent discount on its V4-Pro model permanent. Cached input now costs as little as RMB 0.025, about $0.0035 per million tokens, which is about one percent of the comparable cost at Anthropic. Ramp itself warns against overestimating the trend's durability, a warranted precaution. Alongside DeepSeek, Fireworks AI, fal AI, DeepInfra, and GPU provider Vast.ai also appear on the list. In May, OpenRouter already showed that 7 of the 10 most used models were Chinese-made, with a weekly token volume of 25 trillion, up from 5 trillion six months earlier. → Hello China Tech

Synthszr Take: At one percent of Anthropic's cost for cached input, DeepSeek becomes the default choice for any CFO looking at the inference bill. This is exactly what the Ramp report shows: companies are paying directly and transmitting their data through Chinese systems, instead of just talking about prices in developer forums. In our vendor comparison, the compliance question was always the focus with DeepSeek, and a price of $0.0035 per million tokens doesn't make that disappear. Cheap inference increases consumption—that's the Jevons paradox in its purest form: 25 trillion tokens per week on OpenRouter, a fivefold increase in six months. What's forming here is a complete cost tier below the premium providers, with Fireworks AI, fal AI, and DeepInfra as the serving layer. In early April, we wrote that tokens are becoming the new GDP; the bill for that is now landing in American corporate accounts. The task for everyone is concrete and can be decided tomorrow morning: first, clarify which data can even be run through a Chinese model, and only then talk about the cost savings.

Recursive AI Dramatically Accelerates Development

The Anthropic Institute has published figures showing that AI is already accelerating the development of AI; Ethan Mollick summarizes them in his newsletter. Anthropic engineers now deliver eight times as much code per quarter as they did between 2021 and 2025. The term for this is 'recursive self-improvement': a system that designs and trains its own successor. We're not there yet, but the curve is pointing steeply in that direction. The length of tasks that models can reliably handle on their own now doubles every four months, instead of every seven as before. In March 2024, Claude Opus 3 could handle four-minute software tasks; a year later, Sonnet 3.7 was completing one-and-a-half-hour tasks, and today Opus 4.6 manages twelve-hour tasks. On SWE-bench, models climbed from single-digit scores to saturation in two years, and on CORE-Bench, from 20 percent to fully solved in fifteen months. Mollick distinguishes between two work areas: engineering, where Claude can solve an underspecified task on its own, and research, where it matches or surpasses experienced humans on clearly defined experiments. By 2027, he believes AI systems will be capable of tasks that would otherwise take a human weeks. → Ethan Mollick from One Useful Thing

Synthszr Take: The 80 percent is the headline. The real lesson is hidden in the 800 individual fixes Claude made autonomously via the API in April, reducing the error rate by a factor of 1,000. This is the lever anyone can pull tomorrow morning: unleash agents on the legacy debt that's been postponed for years and that no human wants to touch, instead of chasing speculative new features. To their credit, Anthropic openly names the second effect: flooding the system with synthetic code makes human review the bottleneck, and Amdahl's law hits mercilessly. This is precisely why there is now an automated Claude reviewer in the pipeline that, in retrospect, would have caught a third of the bugs that crippled claude.ai. As I described in Code Crash, the engineer's role is shifting from typing to conducting. An 8x increase in code output per engineer is worthless as long as quality assurance sticks to the old pace. If you don't have an automated review architecture, you're now just producing garbage faster.

Claude-Mem Gives Coding Agents a Memory: 80,000 GitHub Stars for a Plugin

Claude-Mem transforms Claude-Code sessions into a searchable, dated project memory. The open-source project by Alex Newman captures session events via hooks, compresses them into observations, and makes them retrievable via an MCP server. As of June 2, the repository had collected 80,253 stars and 6,910 forks on GitHub, along with 149 open issues and pull requests. The architecture combines Claude-Code hooks, a Bun-CLI layer, an Express worker, as well as SQLite and Chroma. Newman emphasizes that the tool generates 'new thoughts, new observations' from each tool execution, rather than just saving transcripts. In addition to Claude Code, it also supports Cursor, Gemini CLI, Windsurf, OpenCode, and Codex CLI. With version 13.0.0, the project switched from AGPL-3.0 to Apache-2.0 so that, according to Newman, 'everyone can build the primitives into their systems.' The catch: a plugin becomes a local service with worker logs, provider calls, and rate limits that you have to keep an eye on (with Gemini without billing, we're talking about 5 to 10 requests per minute). → Marcus Schuler

Synthszr Take: Memory is the missing layer between a usable coding agent and one you don't have to explain half the codebase to again in the morning. 80,253 stars at a time when other projects are begging for 200 says everything about the pain point. In January, we wrote that LLMs would get long-term memory; here is the tangible open-source version of it, built from SQLite, Chroma, and a few hooks. The switch to Apache-2.0 is the real statement: Newman is giving away the primitives so that Anthropic, Cursor, and the rest can integrate them directly. This turns the memory layer into a commoditized product before anyone could seriously monetize it. Any engineering lead betting on this now shouldn't underestimate the operational costs, because the local worker with its logs and provider calls needs maintenance. My pragmatic advice: set it up cleanly once, check the worker logs, verify new rows in the database, and then enjoy the fact that your agent finally knows what it broke while fixing something yesterday.

🦾 ChatGPT Now Dreams and Rewrites Your Past in the Process

OpenAI is giving ChatGPT a feature the company describes with the metaphor of 'sleeping on it': the model revises the entire chat history in the background and updates old facts, so it knows your trip is over and you're back home. ElevenLabs goes a step further with its Flows Agent, which creates a complete advertisement from a single brief. The system selects the models, wires up the pipeline, and generates images, video, voice, and music itself; if a change is made mid-process, it only re-renders the affected part. Meta is providing every business with an AI representative that answers questions, recommends products, books appointments, and closes sales via WhatsApp, Messenger, and Instagram. The AWEAR wearable sits behind the ear, reads the brain's electrical signals, and translates stress, focus, and mental load into understandable cues in real-time. Alongside these are two thought-provoking pieces: six experiments show that errors increase and critical thinking diminishes as soon as people trust AI without double-checking. The developers behind Claude Code break down six core patterns Claude uses to build its own workflows in a playbook. There are also smaller tools like Comp AI for SOC-2 audits or Recoverit for data recovery. → TAAFT - There's An AI For That

Synthszr Take: In January, long-term memory in models was the big headline. Now, OpenAI is letting ChatGPT revise its own memory in the background: the trip is over, you're home, the old state is silently replaced. Convenient, no doubt. What I find more interesting is the news three paragraphs down about six experiments on people who stop thinking. As soon as we trust AI without double-checking, errors climb and critical thinking quietly fades away. This is exactly where the two news items intersect: a system that curates its own facts meets users who no longer verify them. Anyone who completely outsources their memory loses control over their own version history.

78 Years of Computer History, Runnable on a Laptop

There is now a virtual museum for operating systems that makes the entire history of the computer, from 1948 to the present day, runnable on a normal laptop. The Virtual OS Museum bundles over 570 different operating systems, more than 250 platforms, and a total of over 1700 pre-installed setups, all as emulated machines for QEMU, VirtualBox, or UTM. A custom launcher starts the systems regardless of the specific emulator, everything is pre-configured, and a snapshot function restores a broken installation with a click. The collection begins with the Manchester Baby from 1948, the first stored-program computer, and extends through CTSS, the ancestor of all modern operating systems, to early Android and iOS versions. It also includes the Xerox Star, the first system with a graphical desktop interface. The catalog covers mainframes, workstations with their Unix variants, home computers like the Commodore and the ZX Spectrum, and PC systems from DOS through OS/2 and BeOS to Windows starting from version 1.0. Classic Mac OS up to Mac OS X 10.5 and research systems like Smalltalk and Oberon are also included. There is a full version that runs completely offline and a lite version that downloads the images on first launch. → The Download from MIT Technology Review

Synthszr Take: Anyone building interfaces today should spend an hour clicking through this. The Xerox Star had the desktop metaphor figured out back in 1981, long before Apple and Microsoft made it the standard. 570 operating systems from 78 years, ready to run on a laptop: this is the most honest history lesson for anyone who thinks they're inventing something new. Much of what we celebrate as a fresh use case was already on a screen in Palo Alto in the eighties. The practical value lies in seeing which ideas have stood the test of time and which have rightly disappeared. Emulation turns software history into running code instead of a screenshot in a textbook, and that's something most product development teams are missing. I'd put it in every junior's onboarding. Those who know history build the future with more humility.