Anthropic just pulled a move that’s… let’s just say, not going to win it many fans among power users. One of the most popular ways to supercharge Claude, using OpenClaw, is effectively being paywalled out of existence. And yeah, it’s as messy as it sounds.

Anthropic just made OpenClaw way more expensive

According to Anthropic Claude Code exec Boris Cherny, Anthropic has changed how Claude subscriptions work. Starting now, your regular Claude subscription no longer covers usage through third-party tools. Instead, that usage gets kicked into a separate pay-as-you-go billing system.

So what does that mean? Essentially, the hack that many users relied on, which was using Claude credits inside OpenClaw to run more advanced workflows, is basically dead. If someone still wants that setup, they’ll now have to pay extra on top of their subscription, either through usage bundles or API access.

And it’s not like OpenClaw was some niche experiment either. It blew up because it could handle real-world tasks like emails, calendars, and even flight check-ins, turning Claude into something closer to an actual assistant. But that popularity seems to have backfired, reportedly putting pressure on Anthropic’s infrastructure and forcing this clampdown.

This feels less like a tweak… and more like a crackdown

Let’s be real, this isn’t just a pricing change, it’s a pretty clear signal. Anthropic seems to be drawing a hard line: if you’re using Claude in ways they didn’t design (or monetize properly), expect that door to close quickly.

There’s also a bit of strategy baked in here. By making third-party usage more expensive, Anthropic nudges users toward its own ecosystem, like Claude Cowork. That’s great for control, but not so great if you liked mixing and matching tools to build your own workflows. To soften the blow, the company is offering a one-time credit equal to a month’s subscription and discounted bundles. But let’s be honest, that feels more like a transition cushion than a real solution.

from the good-bad-decision dept

The Supreme Court’s decision last year in U.S. v. Skirmetti, upholding a law depriving young trans people the healthcare they need, is insupportable, rendering people unequal in a way the Constitution cannot possibly suborn. But its new decision in Chiles v. Salazar regarding the First Amendment standard to use regarding Colorado’s law regarding conversion therapy is different. Despite its similar subject matter relating to sexual orientation and gender identity sounding similar to Skirmetti, it’s actually another 303 Creative, another case that endorsed bigoted views unacceptably hostile to LGBTQ+ people. But for much the same reason that 303 Creative was an important articulation of the First Amendment’s expansive protection—despite the apparent prejudice the plaintiff (and the Court) advanced—so is this decision.

That’s what’s good about this decision, that it recognizes that the First Amendment operates in the professional licensing space and requires heightened scrutiny before states can be permitted to constrain licensing when those constraints are predicated on viewpoints expressed by the licensee, including as part of the provision of services. Heightened scrutiny is what makes the First Amendment’s protections meaningful, and the Court has not always been consistent or coherent in requiring it, particularly with respect to licensure. But when heightened scrutiny isn’t required, it becomes much harder to fight censorial actions taken by the government, including those driven by animus, and including those driven by anti-LGBTQ+ animus—which would also include those actions targeted at therapists supporting LGBTQ+ patients, such as those recently announced by Ken Paxton in Texas. This Supreme Court decision now makes it much, much harder for him to get away with silencing those therapists whose therapy affirmed their patients’ identity by putting their license at risk if they do.

The main problem with this decision however is that the Court picked a law prohibiting conversion therapy as the moment to finally articulate that heightened scrutiny applies with respect to licensing, including medical licensing. Conversion therapy, as Justice Jackson described in her dissenting opinion, is a scientifically-discredited approach “designed to ‘convert’ a person’s sexual orientation or gender identity, so that the person will become heterosexual or cisgender.” [Dissent p.3]. Historically it has been provided via “aversive modalities,” that many have likened to torture, such as “inducing nausea, vomiting, or paralysis in patients or subjecting them to severe electric shocks to telling patients to snap an elastic band on their wrists in response to nonconforming thoughts.” [Dissent p.3]

Importantly, however, to the extent that any law prohibits these practices, those laws remain in force—this decision does not affect such laws. (“The question before us is a narrow one. Ms. Chiles does not question that Colorado’s law banning conversion therapy has some constitutionally sound applications. She does not take issue with the State’s effort to prohibit what she herself calls ‘long-abandoned, aversive’ physical interventions.” [Majority p.7]). But it does reach conversion therapy delivered via talk therapy, where therapists “seek to encourage patients to change their behavior in an attempt to ‘change’ their identity” still are. [Dissent p.3]. As Jackson explained, this approach also causes real harm. [Dissent p.4-5]. And it’s a kind of harm that states like Colorado, who passed the law challenged here, have an interest in stopping. [Dissent p.5-7].

Making it hard for states to do so raises a number of concerns, such as that the decision will give a veneer of legitimacy to conversion therapy and stoke the hostile anti-LGBTQ+ attitudes driving it, as well as create the risk that conversion therapy, at least insofar as it includes talk therapy, might be something that minors could be legally subjected to in Colorado and elsewhere. There is also the fear that even if the Court has now articulated a good rule about heightened scrutiny it will only remember to apply it in cases like these where it will lead to results consistent with the Court majority’s biases—in other words, while the Court may be happy to subject Colorado’s anti-conversion therapy rule to strict scrutiny, there is the fear that it will conveniently forget to apply it to, say, Texas’s law trying to punish those who refuse to engage in it.

It also raises a collateral concern even on the speech-protection front, that subjecting licensure requirements to strict scrutiny could have the practical effect of diluting the standard. As Jackson also noted, we have long allowed states to regulate medical professionals, [Dissent p.8], as well as other licensed professionals like lawyers, and much of the regulation is directed to how licensed practitioners speak in some way as they provide their services. Perhaps all these efforts could actually pass strict scrutiny. In fact, it’s even still possible that Colorado’s law might yet survive it; although Justice Gorsuch’s majority opinion casts some doubt, the case is not over.

Rather than deciding it for themselves, the Court remanded the case back to the lower courts to this time apply the more exacting strict scrutiny standard rather than the less-demanding rational basis review they originally applied. Presumably there will be more opportunity for briefing and argument to show how the particular harm of conversion therapy creates the compelling state interest Colorado needed to act, and that its prohibition of licensed therapists from providing it via talk therapy is a remedy that is sufficiently narrowly tailored.

But the problem with applying strict scrutiny to so much regulation targeting licensing is that it might start to become too easy to satisfy when there are strong policy reasons to favor the government action, and as a result strict scrutiny will no longer be useful as a standard if it essentially allows everything, instead of being a meaningful filter. There are after all always compelling reasons for the government to care about the quality of the services licensees deliver via their professional expression, but just because the government has a valid reason to regulate does not mean that everything it does to regulate is constitutional.

Strict scrutiny also requires that the state action be narrowly tailored, in addition to being motivated by a compelling reason, and it’s too easy for courts to skip that part of the analysis, as we saw with the TikTok ban when it was somehow blessed by the DC Circuit. And the fear is that the more strict scrutiny is applied to what is fairly ordinary state regulation—of licensed practitioners—the more likely it will have the practical effect of creating precedent that dilutes the standard so that it is no longer so strict when we need it to be, especially for state action that is more exceptional. (On the TikTok ban the Supreme Court had greenlighted it using a lesser standard, which was itself extremely problematic as the ban should have been found unconstitutional, but at least the tool that should have applied to it remained sharp for future use, rather than dulled by this bad decision.)

On the other hand, a decision upholding the lower courts’ use of rational basis review would have done no one any favors. As Justice Kagan wrote in her concurrence, joined by Justice Sotomayor, it is easy to imagine a law that mirrors what the Colorado one does, prohibiting talk therapy that accepts LGBTQ+ identity instead of challenges it, and now advocates are left with a much more powerful tool to challenge it.

Of course, it does not matter what the State’s preferred side is. Consider a hypothetical law that is the mirror image of Colorado’s. Instead of barring talk therapy designed to change a minor’s sexual orientation or gender identity, this law bars therapy affirming those things. As Ms. Chiles readily acknowledges, the First Amendment would apply in the identical way. [Concurrence p.3]

As Texas shows, such a situation is not hypothetical. But now with this decision people challenging such censorial government efforts can turn to long-established First Amendment doctrine in their fight. And the doctrine remains stable, rather than something now swiss-cheesed with bespoke exceptions tied to certain policy preferences. No matter how valid those preferences, if they can be given special constitutional treatment then so can the bad ones. This decision helps buttress the guardrails preventing speech from being protected or not based on whether the government likes it, which is the whole reason we have the First Amendment, to make sure government preferences cannot dictate what views people can express.

Which is especially important when the courts cannot be trusted to overcome their biases to have good sense about which policy preferences are good and bad. The Supreme Court of course only has itself to blame that the public is so primed to believe that its decisions are driven by its biases and not neutral, sustainable doctrine. But nevertheless this decision still stands as an important declaration of law that is consistent with existing First Amendment jurisprudence and one that will ultimately leave everyone, including those challenging government actions attacking LGBTQ+ interests, far better off than if the Court had let the lower courts’ decisions invalidating the law stand after using a less speech-protective rule. In fact it will be an important one for anyone fighting censorship in any context, including those we generally talk about here, to use, because with this decision, the rule that has long been the rule remains the rule: when a government action non-incidentally touches on speech, is content-based, and is not viewpoint neutral, strict scrutiny applies.

Per this decision, a law targeting what therapists can say inherently involves speech, and not in an incidental way. And it targets it in a way that is not viewpoint-neutral; it has a specific preference, that conversion therapy is bad. As a result, as a law that targets the content of speech in a way that is not viewpoint-neutral, strict scrutiny, a more exacting standard than the rational basis review the lower courts had used, is required.

Turning to the merits, both the district court and the Tenth Circuit denied Ms. Chiles’s request for a preliminary injunction. The courts recognized that Ms. Chiles provides only “talk therapy.” And they acknowledged that Colorado’s law regulates the “verbal language” she may use. But, the courts held, the main thrust of the State’s law is to delineate which “treatments” and “therapeutic modalit[ies]” are permissible. Accordingly, the courts reasoned that Colorado’s law is best understood as regulating “professional conduct.” At most, they continued, Colorado’s law regulates speech only “incidentally” to professional conduct. As a result, the courts concluded, Colorado’s law triggers no more than “rational basis review” under the First Amendment, requiring the State to show merely that its law is rationally related to a legitimate governmental interest. Because the State satisfied that standard, the courts held that Ms. Chiles was not entitled to the relief she sought. [Majority p.6]

[…]

Consistent with the First Amendment’s jealous protections for the individual’s right to think and speak freely, this Court has long held that laws regulating speech based on its subject matter or “communicative content” are “presumptively unconstitutional.” Reed v. Town of Gilbert, 576 U. S. 155, 163 (2015). As a general rule, such “content-based” restrictions trigger “strict scrutiny,” a demanding standard that requires the government to prove its restriction on speech is “narrowly tailored to serve compelling state interests.” Ibid. Under that test, it is ” ‘rare that a regulation . . . will ever be permissible.’ ” Brown v. Entertainment Merchants Assn., 564 U. S. 786, 799 (2011) (quoting United States v. Playboy Entertainment Group, Inc., 529 U. S. 803, 818 (2000)).

We have recognized, as well, the even greater dangers associated with regulations that discriminate based on the speaker’s point of view. When the government seeks not just to restrict speech based on its subject matter, but also seeks to dictate what particular “opinion or perspective” individuals may express on that subject, “the violation of the First Amendment is all the more blatant.” Rosenberger v. Rector and Visitors of Univ. of Va., 515 U. S. 819, 829 (1995). “Viewpoint discrimination,” as we have put it, represents “an egregious form” of content regulation, and governments in this country must nearly always “abstain” from it. Ibid.; see also Iancu v. Brunetti, 588 U. S. 388, 393 (2019) (describing “the bedrock First Amendment principle that the government cannot discriminate” based on view-point (internal quotation marks omitted)); Good News Club v. Milford Central School, 533 U. S. 98, 112–113 (2001); Barnette, 319 U. S., at 642. [Majority p.8-9]

[…]

As applied here, Colorado’s law does not just regulate the content of Ms. Chiles’s speech. It goes a step further, prescribing what views she may and may not express. For a gay client, Ms. Chiles may express “[a]cceptance, support, and understanding for the facilitation of . . . identity exploration.” For a client “undergoing gender transition,” Ms. Chiles may likewise offer words of “[a]ssistance.” But if a gay or transgender client seeks her counsel in the hope of changing his sexual orientation or gender identity, Ms. Chiles cannot provide it. The law forbids her from saying anything that “attempts . . . to change” a client’s “sexual orientation or gender identity,” including anything that might represent an “effor[t] to change [her client’s] behaviors or gender expressions or . . . romantic attraction[s].” [Majority p.13]

But even if the law as it stands can’t survive strict scrutiny, in her concurrence, joined by Justice Sotomayor, Justice Kagan suggested ways the law might be amended so that it could be upheld.

It would, however, be less [likely to be unconstitutional] if the law under review was content based but viewpoint neutral. Such content-based laws, as the Court explains, trigger strict scrutiny “[a]s a general rule.” But our precedents respecting those laws recognize complexity and nuance. We apply our most demanding standard when there is any “realistic possibility that official suppression of ideas is afoot”—when, that is, a (merely) content-based law may reasonably be thought to pose the dangers that viewpoint-based laws always do. Davenport v. Washington Ed. Assn., 551 U. S. 177, 189 (2007). But when that is not the case—when a law, though based on content, raises no real concern that the government is censoring disfavored ideas—then we have not infrequently “relax[ed] our guard.” Reed, 576 U. S., at 183 (opinion of KAGAN, J.); see Davenport, 551 U. S., at 188 (noting the “numerous situations in which [the] risk” of a content-based law “driv[ing] certain ideas or viewpoints from the marketplace” is “attenuated” or “inconsequential, so that strict scrutiny is unwarranted”). Just two Terms ago, for example, the Court declined to apply strict scrutiny to a content-based but viewpoint-neutral trademark restriction. See Vidal v. Elster, 602 U. S. 286, 295 (2024); id., at 312 (BARRETT, J., concurring in part); id., at 329–330 (SOTOMAYOR, J., concurring in judgment). In the trademark context, as in some others, experience and reason alike showed “no significant danger of idea or viewpoint” bias. R. A. V., 505 U. S., at 388.

The same may well be true of content-based but viewpoint-neutral laws regulating speech in doctors’ and counselors’ offices.* Medical care typically involves speech, so the regulation of medical care (which is, of course, pervasive) may involve speech restrictions. And those restrictions will generally refer to the speech’s content. Cf. Reed, 576 U. S., at 177 (Breyer, J., concurring in judgment) (noting that “[r]egulatory programs” addressing speech “inevitably involve content discrimination”). But laws of that kind may not pose the risk of censorship—of “official suppression of ideas”—that appropriately triggers our most rigorous review. R. A. V., 505 U. S., at 390. And that means the “difference between viewpoint-based and viewpoint-neutral content discrimination” in the health-care context could prove “decisive.” Vidal, 602 U. S., at 330 (opinion of SOTOMAYOR, J.). Fuller consideration of that question, though, can wait for another day. We need not here decide how to assess viewpoint-neutral laws regulating health providers’ expression because, as the Court holds, Colorado’s is not one. [Concurrence p.3-4]

Ultimately, despite all of the concerns, the decision is still a good one that will leave everyone better off. And not just for cases that reach the Supreme Court but in every state and federal court hearing every challenge of laws trying to penalize certain views, including those accepting of LGBTQ+ identities. Whereas a decision to the contrary, one that would have allowed a rational basis standard to be the test for the law’s constitutionality, could be used to defend laws that, instead of fighting LGBTQ+ prejudice as this one tried to do, instead advanced it. As Texas illustrates, already there are examples of certain government actors attempting to impose their biased viewpoints via licensing requirements for therapists. This decision, even if it may stand as an individual reflection of LGBTQ+ animus by this Supreme Court, still makes further state action motivated by it that much harder for any government actor to impose.

Filed Under: 1st amendment, chiles v. salazar, colorado, conversion therapy, free speech, strict scrutiny, supreme court

I want to share a story of struggle. Actually, two kinds of struggle.

My father completed his doctorate at the University of Utah in the early 1970s. For his dissertation, he ran a statistical analysis on genealogical records to determine the impact of certain economic conditions on family size.

He accomplished this on one of the most advanced computers of the time. His method? Literally punching out little rectangles in dozens of stiff paper cards, and feeding the stack into the computer.

My father was a lowly graduate student, and because the demand for computing time at the university was sky high, he had to run his analysis in the middle of the night. He spent many nights punching cards and running them through the machine. Even a single mispunch would cause the entire program to stop running and require painstaking troubleshooting, re-punching, and another night at the computer lab.

Unproductive vs. Productive Struggle

The soul-sapping sleep deprivation and endless paper punching that stood between my father and his goals represents the first kind of struggle in my story: unproductive struggle — the challenging, unavoidable tasks we must perform toward a learning goal, but which add no value to the intellectual outcome.

The real intellectual challenge in my father’s work was in deciding which variables belonged in the model, determining how to represent economic conditions over time, and interpreting the data. This is the second kind of struggle: productive struggle. That is, the effort a learner expends to make sense of concepts, to figure something out that is not immediately apparent. This struggle leads to growth and insight. It builds judgment, expertise and understanding.

What is frustrating about my father’s story in hindsight is that so much of his time and cognitive energy were consumed by the unproductive struggle of punching cards and managing the computer. Without those barriers, he would have had more capacity for the productive struggle that leads to meaningful learning.

Thinking About What Matters

When it comes to AI in schools, some educators fear that it will lead to learning becoming too easy. This is referred to as “cognitive laziness.” The assumption is that we will offload our thinking to AI and eventually lose our ability to think critically. This is a risk with any technology that makes our mental work more efficient, and AI is uniquely adept at taking on cognitively demanding tasks. But ceding our reasoning power to AI isn’t a foregone conclusion. And simply not using AI in learning settings doesn’t have to be our solution for preserving our mental capacities.

Just as better computing tools would have freed my father from punching cards without removing the intellectual rigor of his work, today’s tools, including AI, have the potential to offload unproductive struggle, while preserving, and even amplifying, the productive struggle that is central to learning.

Here’s an example: When reading comprehension is not the goal of a lesson but a necessary prerequisite — a student having to read an article to understand the causes of the French Revolution, for example — AI tools can adjust reading levels on the fly to assist learners who are below grade level or for whom English is not their first language. This allows them to focus on the history rather than on decoding the text.

Refining Rigor

So what does this mean for educators who are grappling with how to help students use AI effectively?

First, we need to remind ourselves and help our students understand that the goal of learning has never been to make learning easy. It is to make it meaningful. We must ensure that learners are spending their time wrestling with big ideas, not battling logistics or bogged down by rote tasks.

Second, educators need to face a hard truth about the assignments we give students. Many assignments contain a mix of productive and unproductive struggle, and we are not always very intentional about which is which. Under crushing time and resource pressure, we can become unreflective about the distinction between productive and unproductive work. We inherit assignments, reuse problem sets, and value rigor without always asking where the rigor actually lies.

If AI forces us to confront that, it may be one of the most useful disruptions education has experienced in decades.

For instance, requiring students to write citations according to a set format may feel rigorous, but the cognitive work of formatting has little to do with the intellectual work of evaluating sources and integrating evidence into an argument. This shift requires us to redesign tasks, rethink assessments and, if necessary, let go of practices that feel rigorous but don’t meaningfully deepen understanding.

Sharpening Learning

If we do this well, AI won’t hollow out learning; it will sharpen it. It will give students more space to wrestle with ideas instead of mechanics, more time to interpret instead of transcribe, and more opportunity to make active sense of the world. It will give us a chance to be far more intentional about the kind of struggle we ask students to engage in.

In the end, AI won’t decide whether our students experience cognitive laziness or cognitive growth. We will decide that by how we design assignments and assessments, and by the choices we make about which AI tools to adopt and how we choose to use them.

This is our chance to weed out the punch cards and open up more time for students to struggle over things that truly matter.

Six months after renegotiating the contract that once barred it from independently pursuing frontier AI, Microsoft has released three in-house models that directly challenge the partner it spent $13 billion cultivating. MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 are now available in Microsoft Foundry, and they do not carry OpenAI’s name anywhere on the label.

The models are the first publicly released output of the MAI Superintelligence team that Mustafa Suleyman, CEO of Microsoft AI, formed in November 2025 with a stated mission of pursuing what the company calls “humanist superintelligence.” In a March internal memo first reported by Business Insider, Suleyman wrote that he intended to focus all of his energy on superintelligence and deliver world-class models for Microsoft over the next five years. That ambition now has its first tangible evidence.

MAI-Transcribe-1 is, on paper, the most immediately disruptive of the three. The speech-to-text model claims the lowest word error rate across 25 languages on the FLEURS benchmark, averaging 3.8 per cent, and Microsoft says it outperforms OpenAI’s Whisper-large-v3 on all 25 languages, Google’s Gemini 3.1 Flash on 22 of 25, and ElevenLabs’ Scribe v2 on 15 of 25. It runs 2.5 times faster than Microsoft’s previous Azure Fast transcription service and is priced at $0.36 per hour of audio. Perhaps most revealing is the team that built it: just 10 people.

MAI-Voice-1 completes the audio loop. The text-to-speech model generates 60 seconds of natural-sounding audio in under one second on a single GPU and supports custom voice creation from a few seconds of sample audio. Combined with MAI-Transcribe-1 and a large language model of the customer’s choosing, it forms a complete voice pipeline that runs entirely on Microsoft infrastructure without any dependency on OpenAI’s technology.

MAI-Image-2, the oldest of the three, had already debuted at number three on the Arena.ai text-to-image leaderboard in March, placing it behind only Google’s Gemini 3.1 Flash and OpenAI’s GPT Image 1.5. The model was developed in collaboration with photographers, designers, and visual storytellers, and WPP, one of the world’s largest marketing groups, is among the first enterprise partners building with it at scale.

The strategic context matters more than the benchmarks. Until the September 2025 renegotiation, Microsoft’s original partnership agreement with OpenAI contractually prevented the company from independently pursuing general AI development. The revised memorandum of understanding changed that calculus fundamentally. Microsoft retained licensing rights to everything OpenAI builds through 2032, gained $250 billion in new Azure cloud business commitments, and crucially won the freedom to build competing models. Suleyman acknowledged the pivot directly: the contract renegotiation, he said, enabled Microsoft to independently pursue its own superintelligence.

The timing is deliberate. Jacob Andreou, formerly a senior vice-president at Snap, took over as executive vice-president of Copilot on 17 March, freeing Suleyman from day-to-day product responsibilities. The MAI models landed barely two weeks later. Microsoft also hired Ali Farhadi, the former chief executive of the Allen Institute for AI, for Suleyman’s superintelligence team in March, a recruitment signal that the ambitions extend well beyond transcription and image generation.

For OpenAI, the development creates an awkward dynamic. Microsoft remains its single largest investor and its primary cloud infrastructure provider, and the two companies continue to share a platform in Foundry, which hosts both OpenAI and Microsoft models. But OpenAI’s own push into commercial monetisation is accelerating in parallel, and the relationship is beginning to resemble two companies orbiting the same market with overlapping products rather than a partnership with a clear division of labour. OpenAI’s $110 billion raise in February, backed by SoftBank, Nvidia, and Amazon, valued the company independently of Microsoft at a level that makes the original partnership framing increasingly anachronistic.

The broader AI model market is fragmenting along similar lines. Anthropic’s $30 billion raise at a $380 billion valuation established it as a credible third force in enterprise AI, with run-rate revenue of $14 billion. Google continues to iterate rapidly on Gemini. The era in which OpenAI was the only game in town for frontier AI capabilities, and Microsoft was content to be its exclusive distribution channel, is definitively over.

Microsoft Foundry, the platform formerly known as Azure AI Foundry and before that Azure AI Studio (the second rebrand in twelve months), now serves developers at more than 80,000 enterprises including 80 per cent of Fortune 500 companies. That distribution advantage is what makes the MAI model family strategically significant: Microsoft does not need to beat OpenAI on every benchmark to shift enterprise spending toward in-house models. It needs to be competitive enough that customers choose the integrated option over the third-party alternative, a dynamic that the past year of AI industry consolidation has made increasingly plausible.

Suleyman has said it will take another year or two before the superintelligence team produces frontier-class language models. What landed this week is the foundation: a multimodal toolkit that gives Microsoft its own voice, ears, and eyes independent of OpenAI. The $13 billion partnership is not ending. But the premise on which it was built, that Microsoft needed OpenAI to compete in AI, is being quietly dismantled one model release at a time.

Strands is the NYT’s latest word game after the likes of Wordle, Spelling Bee and Connections – and it’s great fun. It can be difficult, though, so read on for my Strands hints.

Want more word-based fun? Then check out my NYT Connections today and Quordle today pages for hints and answers for those games, and Marc’s Wordle today page for the original viral word game.

AI vibe coders have yet another reason to thank Andrej Karpathy, the coiner of the term.

The former Director of AI at Tesla and co-founder of OpenAI, now running his own independent AI project, recently posted on X describing a “LLM Knowledge Bases” approach he’s using to manage various topics of research interest.

By building a persistent, LLM-maintained record of his projects, Karpathy is solving the core frustration of “stateless” AI development: the dreaded context-limit reset.

As anyone who has vibe coded can attest, hitting a usage limit or ending a session often feels like a lobotomy for your project. You’re forced to spend valuable tokens (and time) reconstructing context for the AI, hoping it “remembers” the architectural nuances you just established.

Karpathy proposes something simpler and more loosely, messily elegant than the typical enterprise solution of a vector database and RAG pipeline.

Instead, he outlines a system where the LLM itself acts as a full-time “research librarian”—actively compiling, linting, and interlinking Markdown (.md) files, the most LLM-friendly and compact data format.

By diverting a significant portion of his “token throughput” into the manipulation of structured knowledge rather than boilerplate code, Karpathy has surfaced a blueprint for the next phase of the “Second Brain”—one that is self-healing, auditable, and entirely human-readable.

Beyond RAG

For the past three years, the dominant paradigm for giving LLMs access to proprietary data has been Retrieval-Augmented Generation (RAG).

In a standard RAG setup, documents are chopped into arbitrary “chunks,” converted into mathematical vectors (embeddings), and stored in a specialized database.

When a user asks a question, the system performs a “similarity search” to find the most relevant chunks and feeds them into the LLM.Karpathy’s approach, which he calls LLM Knowledge Bases, rejects the complexity of vector databases for mid-sized datasets.

Instead, it relies on the LLM’s increasing ability to reason over structured text.

The system architecture, as visualized by X user @himanshu in part of the wider reactions to Karpathy’s post, functions in three distinct stages:

1. Data Ingest: Raw materials—research papers, GitHub repositories, datasets, and web articles—are dumped into a raw/ directory. Karpathy utilizes the Obsidian Web Clipper to convert web content into Markdown (.md) files, ensuring even images are stored locally so the LLM can reference them via vision capabilities.

2. The Compilation Step: This is the core innovation. Instead of just indexing the files, the LLM “compiles” them. It reads the raw data and writes a structured wiki. This includes generating summaries, identifying key concepts, authoring encyclopedia-style articles, and—crucially—creating backlinks between related ideas.

3. Active Maintenance (Linting): The system isn’t static. Karpathy describes running “health checks” or “linting” passes where the LLM scans the wiki for inconsistencies, missing data, or new connections. As community member Charly Wargnier observed, “It acts as a living AI knowledge base that actually heals itself.”

By treating Markdown files as the “source of truth,” Karpathy avoids the “black box” problem of vector embeddings. Every claim made by the AI can be traced back to a specific .md file that a human can read, edit, or delete.

Implications for the enterprise

While Karpathy’s setup is currently described as a “hacky collection of scripts,” the implications for the enterprise are immediate.

As entrepreneur Vamshi Reddy (@tammireddy) noted in response to the announcement: “Every business has a raw/ directory. Nobody’s ever compiled it. That’s the product.”

Karpathy agreed, suggesting that this methodology represents an “incredible new product” category.

Most companies currently “drown” in unstructured data—Slack logs, internal wikis, and PDF reports that no one has the time to synthesize.

A “Karpathy-style” enterprise layer wouldn’t just search these documents; it would actively author a “Company Bible” that updates in real-time.

As AI educator and newsletter author Ole Lehmann put it on X: “i think whoever packages this for normal people is sitting on something massive. one app that syncs with the tools you already use, your bookmarks, your read-later app, your podcast app, your saved threads.”

Eugen Alpeza, co-founder and CEO of AI enterprise agent builder and orchestration startup Edra, noted in an X post that: “The jump from personal research wiki to enterprise operations is where it gets brutal. Thousands of employees, millions of records, tribal knowledge that contradicts itself across teams. Indeed, there is room for a new product and we’re building it in the enterprise.”

As the community explores the “Karpathy Pattern,” the focus is already shifting from personal research to multi-agent orchestration.

A recent architectural breakdown by @jumperz, founder of AI agent creation platform Secondmate, illustrates this evolution through a “Swarm Knowledge Base” that scales the wiki workflow to a 10-agent system managed via OpenClaw.

The core challenge of a multi-agent swarm—where one hallucination can compound and “infect” the collective memory—is addressed here by a dedicated “Quality Gate.”

Using the Hermes model (trained by Nous Research for structured evaluation) as an independent supervisor, every draft article is scored and validated before being promoted to the “live” wiki.

This system creates a “Compound Loop”: agents dump raw outputs, the compiler organizes them, Hermes validates the truth, and verified briefings are fed back to agents at the start of each session. This ensures that the swarm never “wakes up blank,” but instead begins every task with a filtered, high-integrity briefing of everything the collective has learned

Scaling and performance

A common critique of non-vector approaches is scalability. However, Karpathy notes that at a scale of ~100 articles and ~400,000 words, the LLM’s ability to navigate via summaries and index files is more than sufficient.

For a departmental wiki or a personal research project, the “fancy RAG” infrastructure often introduces more latency and “retrieval noise” than it solves.

Tech podcaster Lex Fridman (@lexfridman) confirmed he uses a similar setup, adding a layer of dynamic visualization:

“I often have it generate dynamic html (with js) that allows me to sort/filter data and to tinker with visualizations interactively. Another useful thing is I have the system generate a temporary focused mini-knowledge-base… that I then load into an LLM for voice-mode interaction on a long 7-10 mile run.”

This “ephemeral wiki” concept suggests a future where users don’t just “chat” with an AI; they spawn a team of agents to build a custom research environment for a specific task, which then dissolves once the report is written.

Licensing and the ‘file-over-app’ philosophy

Technically, Karpathy’s methodology is built on an open standard (Markdown) but viewed through a proprietary-but-extensible lens (note taking and file organization app Obsidian).

• Markdown (.md): By choosing Markdown, Karpathy ensures his knowledge base is not locked into a specific vendor. It is future-proof; if Obsidian disappears, the files remain readable by any text editor.

• Obsidian: While Obsidian is a proprietary application, its “local-first” philosophy and EULA (which allows for free personal use and requires a license for commercial use) align with the developer’s desire for data sovereignty.

• The “Vibe-Coded” Tools: The search engines and CLI tools Karpathy mentions are custom scripts—likely Python-based—that bridge the gap between the LLM and the local file system.

This “file-over-app” philosophy is a direct challenge to SaaS-heavy models like Notion or Google Docs. In the Karpathy model, the user owns the data, and the AI is merely a highly sophisticated editor that “visits” the files to perform work.

Librarian vs. search engine

The AI community has reacted with a mix of technical validation and “vibe-coding” enthusiasm. The debate centers on whether the industry has over-indexed on Vector DBs for problems that are fundamentally about structure, not just similarity.

Jason Paul Michaels (@SpaceWelder314), a welder using Claude, echoed the sentiment that simpler tools are often more robust:

“No vector database. No embeddings… Just markdown, FTS5, and grep… Every bug fix… gets indexed. The knowledge compounds.”

However, the most significant praise came from Steph Ango (@Kepano), co-creator of Obsidian, who highlighted a concept called “Contamination Mitigation.”

He suggested that users should keep their personal “vault” clean and let the agents play in a “messy vault,” only bringing over the useful artifacts once the agent-facing workflow has distilled them.

Which solution is right for your enteprise vibe coding projects?

The “Vector DB” approach is like a massive, unorganized warehouse with a very fast forklift driver. You can find anything, but you don’t know why it’s there or how it relates to the pallet next to it. Karpathy’s “Markdown Wiki” is like a curated library with a head librarian who is constantly writing new books to explain the old ones.

The next phase

Karpathy’s final exploration points toward the ultimate destination of this data: Synthetic Data Generation and Fine-Tuning.

As the wiki grows and the data becomes more “pure” through continuous LLM linting, it becomes the perfect training set.

Instead of the LLM just reading the wiki in its “context window,” the user can eventually fine-tune a smaller, more efficient model on the wiki itself. This would allow the LLM to “know” the researcher’s personal knowledge base in its own weights, essentially turning a personal research project into a custom, private intelligence.

Bottom-line: Karpathy hasn’t just shared a script; he’s shared a philosophy. By treating the LLM as an active agent that maintains its own memory, he has bypassed the limitations of “one-shot” AI interactions.

For the individual researcher, it means the end of the “forgotten bookmark.”

For the enterprise, it means the transition from a “raw/ data lake” to a “compiled knowledge asset.” As Karpathy himself summarized: “You rarely ever write or edit the wiki manually; it’s the domain of the LLM.” We are entering the era of the autonomous archive.

Nick Shirley—the right-wing creator whose YouTube investigation sparked the Trump administration’s immigration crackdown in Minnesota—claims that his most recent video about alleged fraud in California was bolstered by data provided by none other than Edward Coristine, one of the first members of the so-called Department of Government Efficiency (DOGE) known online as “Big Balls.”

Coristine, who joined DOGE at 19 years old with no prior government experience, was staffed across several agencies including the Social Security Administration (SSA) and the Small Business Administration (SBA). Before joining DOGE, Coristine worked at Elon Musk’s Neuralink for several months and founded a startup known for hiring black hat hackers.

In an interview with Coristine published on Shirley’s YouTube channel on Thursday, Shirley claims that Coristine personally pulled data on Medicaid spending for businesses based in California as potential targets. Coristine nodded along, telling Shirley that the government must create more opportunities to crowdsource fraud investigations.

The information Coristine allegedly pulled for Shirley was from a dataset published by the DOGE team at the Department of Health and Human Services (HHS) in February. In a post to X at the time, the HHS DOGE team referred to it as “the largest Medicaid dataset in department history.” The post also claimed that the dataset could be used to “detect” large-scale fraud.

“After that, I went to California based off that dataset you had helped me extract, and these fraudsters also weren’t even trying to hide it,” Shirley told Coristine in Thursday’s interview.

Coristine said that by open-sourcing data on government spending, vigilante investigators like Shirley who are “more well-positioned” could uncover fraudulent payments. “You are someone who actually went to the places where we were spending all this money and confronted the people and got to know the truth. I think we just have to create more opportunities for that to happen. We have to continue to open source data,” Coristine said.

The intersection of the right’s favorite fraud influencer and one of the most notorious DOGE engineers exemplifies the next evolution of DOGE and the Trump administration’s fight against “waste, fraud, and abuse.”

Shirley’s videos have become key pieces of evidence for the Trump administration’s fraud and immigration crackdowns. When Shirley released his December video claiming to have uncovered more than $100 million in Somali-run childcare fraud in Minnesota, figures like vice president JD Vance shared it. A surge of immigration agents were then sent to Minnesota, resulting in mass arrests, detainments, and the deaths of two protesters, Renee Good and Alex Pretti.

Early in their YouTube video, Shirley and Coristine directly tie fraud to immigrant communities and foreigners. “A lot of the money is being stolen and siphoned out of the country,” Coristine says, without providing evidence. “Once that money is in a suitcase to Somalia, that’s never coming back,” Shirley replies.

Later in the video, Shirley and Coristine cite specific examples of “waste and fraud” identified by DOGE, including funding for a “Sesame Street style children’s TV program in Iraq” and “tax policy consulting in Liberia.” Both programs were supported by the US Agency for International Development (USAID), which DOGE effectively shut down in the early months of 2025. Coristine also alleged that the SBA “did a terrible job,” particularly with loans during the height of COVID, and that there were “no checks at all on who’s receiving money, not even the most basic checks of like, if [a Social Security number] is real.”

OpenAI announced a major reorganization on Friday as the company’s CEO of AGI deployment, Fidji Simo, takes medical leave to focus on her health. OpenAI president Greg Brockman will handle the product teams in Simo’s absence. Simo’s previous title was CEO of applications.

Brad Lightcap, the chief operating officer and one of CEO Sam Altman’s top deputies, is transitioning to a “special projects” role. Kate Rouch, the chief marketing officer, is taking a leave of absence to focus on her health. Rouch has been undergoing treatment for breast cancer. When she returns, it will be in “a different, more narrowly scoped role,” according to a note Simo shared with OpenAI staff which was viewed by WIRED.

“As I shared when I joined, I had a relapse of my neuroimmune condition a few weeks before starting the job,” Simo said in the note which was sent in OpenAI’s “core” Slack channel. “It’s been a bit of a rollercoaster since, and the last month has been particularly rough health-wise. For my entire time here, I’ve postponed medical tests and new therapies to stay completely focused on the job and not miss a single day of work. I took time off for the first time two weeks before the break for some medical tests, and it’s now clear that I’ve pushed a little too far and I really need to try new interventions to stabilize my health.”

Simo is expected to take “several weeks” of leave according to her internal post.

In his new role, Lightcap will be in charge of the company’s forward-deployed engineers, which embed within enterprise organizations and help integrate OpenAI’s technology, among other duties.

OpenAI will begin searching for a new CMO, Simo said. The company is also looking for a chief communications officer to replace Hannah Wong, who left her position in January. Chris Lehane has taken over as the leader of the communications team in the interim.

“We have a strong leadership team focused on our biggest priorities: advancing frontier research, growing our global user base of nearly 1 billion users, and powering enterprise use cases,” said an OpenAI spokesperson in a statement. “We’re well-positioned to keep executing with continuity and momentum.”

Simo joined OpenAI in August 2025, where she took over many of the company’s consumer-facing products, including ChatGPT, Codex, and the social-video app Sora. She recently shuttered the Sora app and told staff that the company needed to cut side projects and refocus around its core products.

The decision comes as OpenAI eyes an IPO as soon as this year. The company recently raised $122 billion in the largest funding round the tech industry has ever seen, which valued the company at $852 billion.