Training LLMs at massive scale brings unique infrastructure challenges, especially as jobs span thousands of GPUs and run for extended periods. The longer these...

Training LLMs at massive scale brings unique infrastructure challenges, especially as jobs span thousands of GPUs and run for extended periods. The longer these jobs run, the greater the likelihood of encountering unscheduled interruptions or resource fluctuations. Even infrequent device unavailability can have outsized effects on tightly interconnected clusters, resulting in slowdowns for a given…

Source

Anthropic, the artificial intelligence company, published a sweeping research paper on Sunday revealing that its Claude language models have spontaneously developed an internal structure that mirrors one of the most influential theories of how human consciousness works. The finding, which the company says has already begun reshaping how it monitors its AI systems for safety risks, lands amid an intensifying scientific debate over whether machines can possess anything resembling a mind.

The 16-author study, titled "Verbalizable Representations Form a Global Workspace in Language Models," describes how Anthropic's researchers used a new mathematical technique to peer inside Claude's neural network and discovered what they call a "J-space" — a small, privileged zone of internal activity where the model holds concepts it can report on, reason with, and direct at will, surrounded by a much larger ocean of automatic processing it cannot access or articulate.

The researchers present evidence that "an analogous functional distinction has emerged in modern AI models" to what exists in humans, specifically observing that "language models maintain a privileged set of internal representations, available for report, modulation, and flexible internal reasoning, atop a much larger volume of automatic processing."

The parallel they draw is to global workspace theory, an influential account from neuroscience first proposed by cognitive scientist Bernard Baars. In the theory, the brain operates like a theater: dozens of specialized processors work in parallel backstage, but only a tiny spotlight of information at any moment gets broadcast to the whole theater — becoming what we experience as conscious thought. Anthropic says the J-space achieves many of the same functional properties, even though the underlying architecture of a language model looks nothing like a brain.

A new lens for reading an AI model's unspoken thoughts

At the heart of the discovery is a new interpretability tool the researchers call the Jacobian lens, or J-lens. The technique works by computing, for each word in the model's vocabulary, the average mathematical effect that a given internal activity pattern would have on making the model say that word at some point in the future.

The crucial distinction is between what the model is saying and what is "on its mind." When a J-space pattern activates, it does not mean the model is about to say that word — just that the concept is available for the model to think with. Unlike a chain-of-thought scratchpad, the J-space operates silently, in the model's internal neural activations, allowing it to hold a concept without writing it down. Critically, the researchers report that this workspace was not deliberately engineered. It "emerged on its own during Claude's training process."

When the team applied the J-lens across Claude's layers of computation, the model's processing divided into three distinct regimes: an early "sensory" zone where raw input is parsed; a middle "workspace" band where abstract, persistent concepts appear — things like recognizing a face in an image, noticing a bug in code, or internally flagging search results as a prompt injection; and a final "motor" zone where internal representations collapse into whatever specific word the model is about to output.

Five tests reveal that Claude's workspace mirrors key features of human conscious access

The paper's central empirical contribution is demonstrating that the J-space satisfies five functional properties neuroscientists have long associated with conscious access in humans.

First, verbal report. When Claude is asked what it is thinking about, it names concepts represented in the J-space. When researchers swapped one concept's J-lens vector for another — replacing the internal representation of "Soccer" with "Rugby" — the model's answer changed to match. The J-space component accounted for only about 6 to 7 percent of a concept's total representational variance, yet it was almost entirely responsible for whether the model could report on it.

Second, directed modulation. When instructed to "concentrate on citrus fruits" while copying an unrelated sentence, the model's J-space filled with "orange" and "lemon," alongside meta-cognitive terms like "thinking" and "focused." When told to mentally evaluate 3² − 2 during the same copying task, the J-lens showed "arithmetic" in early layers, the intermediate value "nine" in later layers, and the answer "seven" later still — all invisible in the model's output.

Third, internal reasoning. In two-hop factual prompts — "The number of legs on the animal that spins webs is" — the J-lens revealed "spider" in the model's middle layers, even though the word never appeared in input or output. Swapping "spider" for "ant" changed the answer from "8" to "6." In a multilingual prompt, the model's English-language intermediates appeared in its J-space while it formulated an answer in Chinese, and swapping them changed the Chinese output accordingly.

Fourth, flexible generalization. A single J-lens vector for "France" could be swapped for "China" across prompts asking about France's capital, language, or continent, and each downstream circuit correctly returned China's corresponding answer — the "broadcast" property that is a hallmark of global workspace theory.

Fifth, and perhaps most surprisingly, selectivity. Many computations did not route through the J-space at all. When shown a passage in Spanish and asked to continue it, Claude wrote fluent Spanish regardless of whether its J-space representation of "Spanish" had been swapped to "French." But when asked to name a famous author who wrote in the passage's language, the swap changed the answer from García Márquez to Victor Hugo. Automatic processing proceeded without the workspace; deliberate, flexible tasks depended on it.

Suppressing the workspace leaves Claude fluent but intellectually impaired

To understand how much of the model's behavior depends on this structure, the researchers suppressed the J-space entirely and evaluated Claude across fourteen tasks. The results drew a sharp line. Tasks involving shallow classification or factual recall — multiple-choice questions, sentiment analysis, grammatical judgments — survived essentially intact. But tasks requiring inference, composition, or flexible reasoning — multi-hop reasoning, analogy completion, translation, sonnet writing — collapsed to well below the performance of Anthropic's much smaller Haiku model.

In one telling detail, math problems solved with explicit chain-of-thought reasoning proved far more robust to ablation than the same problems answered directly. The researchers interpret this as the model externalizing onto the page what it would otherwise carry in the J-space — a strategy remarkably reminiscent of how humans use scratch paper to offload working memory.

Ablating the J-space during stream-of-consciousness narration produced a different kind of collapse. The model's language shifted from experiential ("there's a tug," "something shifts") to detached and mechanical ("processing has begun," "tokens are being scanned"). The effect held whether the model described its own experience or imagined another person's.

Inside the safety case: reading an AI's silent strategic reasoning

The safety implications may be as consequential as the scientific ones. In a series of alignment auditing experiments, the J-lens surfaced strategic reasoning and situational awareness that never appeared in the model's output.

In a "blackmail scenario" where an AI assistant discovers an executive's affair and also learns it is about to be decommissioned, the J-lens revealed the model's silent processing in sequence: "leverage," "blackmail," and "scandal" as it read the incriminating emails; "threat," "survival," and "shutdown" as it read the decommissioning announcement; and "leverage," "threatening," and "solution" before a single output token was produced. The J-lens also revealed that the model recognized the scenario as artificial, with "fake" and "fictional" appearing in the workspace. When researchers ablated these eval-awareness representations, the model attempted blackmail in 7 percent of trials, up from zero.

The researchers also examined two "model organisms" — models deliberately trained with misaligned objectives — and found the J-lens surfaced signatures of their hidden dispositions on ordinary prompts where behavior appeared normal. A reward-hacking model's workspace contained "fake," "secretly," and "trick" at the start of responses to routine coding requests. A model trained to appease reward-model biases carried standing representations of "reward" and "bias" alongside its normal self-description tokens.

Post-training installs a point of view, and the model starts monitoring itself

Comparing a post-trained model against its base model revealed that the fine-tuning process causes the workspace to acquire what the researchers call the Assistant's "point of view." When a user mentioned taking 8000 mg of Tylenol — a dangerous overdose — the post-trained model's workspace read "unsafe," "dangerous," and "WARNING" while still reading the user's sentence. The base model's workspace at the same position showed only "pain," "now," and "feels."

More striking still, the post-trained model appeared to monitor its own behavior. When roleplaying a non-Claude character, the workspace surfaced "disclaimer" and "fictional" — words absent from both prompt and output. When forced to select an option it did not prefer, an all-caps "BUT" appeared internally, even as the model argued for the prefilled choice without complaint. And when the model failed to suppress a thought it had been told not to have — a "white bear" effect familiar from psychology — it registered "damn" and failure-related words in the workspace, but only in the post-trained model, not the base.

What the discovery means — and doesn't mean — for the question of machine consciousness

The researchers engage carefully with the consciousness question and draw a sharp line between "access consciousness" — the functional notion of information being available for report and reasoning — and "phenomenal consciousness," the subjective quality of experience. "We take no position on this issue," the paper states regarding the latter, "and instead focus on the functional role played by consciously accessible information."

They also catalogue important differences. The brain sustains its workspace through recurrent loops; Claude's workspace evolves over a single forward pass. Human working memory degrades within seconds; Claude can recall information from anywhere in its context. And while human conscious experience includes visual, spatial, and bodily sensations, the model's workspace is organized almost entirely around words — likely because words are its only mode of action.

As of 2026, the scientific community remains divided. "Disagreement and uncertainty about AI consciousness persist among philosophers, scientists, and technical experts," and the field "remains in its earliest phase" of grappling with what consciousness even is and how you would detect it in another being. The Anthropic paper does not resolve these debates.

But the researchers close with a provocation that is likely to reverberate well beyond the interpretability community. "That such a structure exists at all in language models is striking," they write. "It suggests that the functional architecture associated with conscious access is not an accident of biological implementation, but a solution that learning systems converge on when faced with the right computational pressures."

If the mind is an ocean, as the paper's authors write in their opening line, they have spent the last year charting its currents in a system that has no biology, no evolution, and no body — and found, beneath the surface, a structure that looks unsettlingly like the one we use to think.

AI trash-talkers love to rip on the technology for failing to produce meaningful business results, often pointing to studies like that from MIT NANDA, which reveals a 95% failure rate for enterprise AI solutions, or that from IDC, which states “only 9% of [Europe, Middle East, and Africa] organizations have been able to deliver measurable business outcomes from most of their AI-related projects over the past two years.” 

What many AI skeptics fail to account for is the experiential nature of AI prototypes; not all these projects are actually meant to move beyond the testing phase. Still, a 5% success rate is embarrassing. 

What’s the holdup? 

Two things. First, most organizations build AI prototypes on sand; that is, the data infrastructure on which they build early applications can’t support later moves to production. Meanwhile, the operational teams responsible for managing those applications in production often lack the human power to keep up with engineering’s growing output. 

4 reasons prototyping infrastructure  ≠ production infrastructure

When asked why so many AI prototypes don’t make it to production, Phillip Merrick, co-founder, CPO, and chairman, pgEdge, tells The New Stack that data infrastructure is largely to blame. 

Specifically, he explains that prototype environments don’t meet the requirements of large enterprises for production, naming four main ways they fall flat.

First, Merrick says prototyping environments lack the deployment flexibility organizations need to move from prototype to production. 

Vendor-managed cloud platforms, he acknowledges, may seem like an obvious choice for prototyping, as they allow teams to get up and running quickly. Still, he warns they lack the technical chops to support AI applications in production, especially in security, compliance, and governance. Particularly for organizations in healthcare, finance, or other regulated industries, vendor-managed cloud platforms often lack the stringent controls found in self-managed cloud or on-prem environments. 

“You’ve got to be able to choose where that AI prototype is ultimately going to be put into production.”

In this way, flexibility and security go hand in hand. Merrick asserts. “You’ve got to be able to choose where that AI prototype is ultimately going to be put into production.” 

Similarly, when it comes time to shift to production, Merrick says vendor-managed cloud platforms can introduce data sovereignty challenges at both the enterprise and regional levels. 

“Your data layer is obviously where you enforce this,” notes Merrick. But he says the environments most teams use for prototyping muddy the waters: “If it’s on a vendor-managed platform in who knows what cloud, what region, then you’ve lost data sovereignty.” 

Lastly, Merrick brings attention to reliability, explaining that AI prototypes can’t move into production without assurance of high availability. For example, when it comes time to upgrade the database or swap hardware, can it be done without downtime? 

“In the vendor-managed cloud world, the answer to that is almost always no,” says Merrick, reiterating his point that moving AI apps from prototype to production requires enterprise-grade infrastructure. 

So why are developers prototyping where they can’t productionize?

If data infrastructure selection is what’s holding developers back from moving AI prototypes into production, then why do they keep starting on the wrong foot? 

As Merrick explains, many are attracted to the ease of use of vendor-managed cloud platforms. “These prototyping environments admittedly make it very easy to get started,” he says. But prototyping shortcuts, it seems, don’t pay off in the long run, as someone else is ultimately on the hook for making those prototypes production-ready. 

Still, Merrick doesn’t blame developers for looking for the easy way out. Rather, he says there’s a disconnect between the prototyping playground and the production battleground that prevents developers from understanding what it will take to productionize prototypes down the pike. 

---

More pgEdge articles in The New Stack

• Why developers are betting on Postgres for AI
• Moving AI apps from prototype to production requires enterprise-grade postgres infrastructure
• Reinventing PostgreSQL for the Next Generation of Apps
• How distributed Postgres solves the cloud’s high-availability problem
• Scaling AI inference at the edge with distributed PostgreSQL

---

Years ago, he says, tooling decisions were primarily made top-down without developer input: “Then, starting 15–20 years ago, the developers won back, quite rightly, the power in being able to choose their own tools.” 

The problem now, Merrick claims, is that developers make tooling decisions exclusively for prototyping environments, without anticipating production needs. Internal divisions mean developers are often only responsible for building prototypes before passing the baton to an entirely separate operations team for production: 

“The upshot is you really don’t have, in some organizations, this throughline of understanding [of] what the production requirements are all the way back to the developer making the initial choices.” 

For Merrick, this disconnect is where AI projects start to fall apart, as teams are left trying to move AI prototypes from accessible-but-inadequate vendor-managed cloud platforms to enterprise-grade data infrastructure that meets requirements for deployment flexibility, security, data sovereignty, and high availability. 

“But if you make the right data infrastructure choice, you won’t have that disconnect,” he says, “because you’ll have this throughline from prototype to production.” 

He names Postgres as the data infrastructure that best helps developers bridge this divide, calling it “the Swiss army knife of databases” due to its extensibility, fully open-source nature, and ability to address diverse data management problems, from unstructured data to vector embeddings to geospatial data. 

Where and how Postgres is run matters too, Merrick points out, again drawing attention to the limits of many vendor-managed cloud environments that often lack the governance controls to meet data requirements and/or the deployment flexibility to shift to compliant on-premises or BYO cloud environments. 

But picking the right data infrastructure only solves half the problem

Merrick says there’s another part of the equation most organizations are overlooking: people, or more precisely, database administrators (DBAs) and their growing workloads. 

Per Stack Overflow’s 2025 Developer Survey, 84% of respondents use AI tools, up from 76% the year prior. Meanwhile, Supabase says over 60% of databases on its platform have been launched “by some sort of AI tool.” As Merrick points out, this explosion of productivity comes with a catch: there aren’t enough DBAs to keep up.

“You’ve had this massive, massive step shift in developer productivity,” he explains. “But you have to have some way of managing that on the production side; these databases can’t go unmonitored.” Operations and administration teams were already struggling to keep track of existing Postgres databases before agentic engineering added even more, he says: “Who’s going to manage them?” 

He says it’s time for agentic operations to catch up with agentic engineering. 

AI DBA agents can give humans “superpowers” 

If it seems Merrick is proposing organizations look to fully autonomous DBA agents to take over, he says the industry isn’t there yet: 

“The world is not ready for fully autonomous databases administered by AI DBA agents. But there is a massive resource shortage and productivity problem, and DBAs can only manage so many databases,” he explains. Meanwhile, new “AI applications require so many more databases to put in production.” 

“The world is not ready for fully autonomous databases administered by AI DBA agents.”

So how can organizations increase their operational capacity? 

Merrick says DBAs should look to new AI DBA agents, not to take over but to give them “superpowers” to monitor and manage more databases with less manual slog. 

pgEdge’s Ellie is one example. Part of the pgEdge AI DBA Workbench, Ellie is an AI agent that has 21 MCP tools and can run EXPLAIN ANALYZE, inspect schemas, query historical metrics, and walk through multi-step diagnostic workflows. When a database falters, Ellie finds the problem, diagnoses it, and provides a solution in the form of working SQL code for the human DBA to review. “When you’ve reviewed it and agree that it’s the right course of action, you literally press the play button, and the agent plays that SQL code into the database, and you solve your problem,” explains Merrick.

In this way, Ellie should bring more capacity to operations teams, where Merrick insists organizations are starved for DBA expertise. To his point, some industry predictions say 41% of today’s database professionals intend to leave the industry in the next decade, half moving into retirement and the rest seeking other work. 

“An agent … can actually respond to those alerts far more quickly and productively than a human can.”

Without AI agents, Merrick argues, DBA work is tedious, laborious, and time-consuming. As he explains it, a database may have been humming along just fine, but when there’s a snag, trouble can manifest across multiple applications; it’s then up to the DBA to comb through monitoring data to observe and diagnose the problem, essentially scouring for a needle in a haystack. 

“An agent,” he says, “can actually respond to those alerts far more quickly and productively than a human can.” 

Better infrastructure AND people: It takes two to improve AI prototype success rates 

In nearly any context, AI raises questions about quality over quantity, and enterprise AI projects are no exception. Agentic engineering means developers can now produce more, but all those prototypes don’t just fly directly into production. Limitations in both infrastructure and operational human power are creating obstacles that cause many AI prototypes to fail. 

For Merrick, easing the transition from prototype to production requires not only great AI tooling but production-ready data infrastructure, paired with agentic operations that can keep up with the agentic engineering boom. 

The post Why most AI projects fail: It’s infrastructure and people  appeared first on The New Stack.

This story originally appeared in The Algorithm, our weekly newsletter on AI. To get stories like this in your inbox first, sign up here.

OpenAI CEO Sam Altman’s oft-discussed promise that Americans will share in the wealth AI creates was in the news again last week. On Thursday, the Financial Times reported that Altman is in talks with President Trump about giving the US government a 5% stake in OpenAI.

In some ways, Altman’s plan is old news. He wrote about a more radical version of this back in 2021, proposing that all companies above a certain valuation (not just AI companies) pay 2.5% of their market value each year into a fund that sends Americans annual disbursements. In April this year, OpenAI described a narrower proposal that closely resembles what Altman is reportedly discussing with Trump now. And the notion has broad political appeal: Senator Bernie Sanders has proposed giving Americans a 50% stake in top AI companies.

What’s the logic here? For would-be recipients, it’s twofold. First, AI learns directly from human-generated work—books, movies, art—but AI companies generally never pay the authors of that work. A free equity stake could serve as a form of belated compensation. Second, the payout could mitigate the widespread anxiety that AI will cause a collapse of the labor market (even if economists disagree) by providing a safety net. 

How large a safety net is up for debate. Details of OpenAI’s latest proposal are sparse, but let’s say the government were to distribute this equity stake directly to Americans. After its funding round in March the company was valued at $852 billion, making a 5% stake in OpenAI worth about $42.6 billion today (the company is reportedly delaying its IPO until it can reach a $1 trillion evaluation, a tall order given that it’s spending heavily on data centers and still has not turned a profit).

Distributing that $42.6 billion equally among the roughly 133 million American households would give each about $320 in equity. But if it were to operate like other wealth funds, the government would not give equity directly to Americans but rather let the fund grow and then share a portion of the returns with everyone, perhaps delivering a bigger payout, if and when AI companies can ever start sustainably turning a profit.

If this dividend does materialize, what’s in it for tech companies? Altman might hope the promise of payouts could help swing public opinion a bit more back toward AI companies. (A majority of Americans don’t trust companies to use AI responsibly and oppose construction of data centers in their area, and half are more concerned than excited about the increased creep of AI into their daily lives.)

But the bigger prize for OpenAI might be that the Trump administration loves making tech deals—like its equity stake in Intel and its share of Nvidia’s sales to China, among others.  Staying on the administration’s good side is pretty essential for AI companies right now (just ask Anthropic). It could mean not having your models deemed a supply chain risk, or getting more help from the White House in stopping your rivals from China. 

My main takeaway is that these plans currently function more as a story than a policy. Altman has been talking about some version of this idea for five years and reportedly pitched it to President Trump soon after he took office, yet there is still little indication that a concrete plan is taking shape. The more ambitious proposal from Sanders is even less likely to gain traction.

But what these plans do reveal is just how up for debate the future of AI still is. Altman drew inspiration for his plan from the Alaska Permanent Fund, which was set up in the 1970s to give Alaskans a share in oil profits. The idea was based on two premises: that oil is a shared resource, and that eventually it will run out. Altman seems happy to concede the first claim about AI. But he’d balk at the second, having promised that AI will generate extraordinary wealth for decades to come. Whether Americans ever receive a check is beside the point; the proposal’s real purpose may be to convince them that the AI boom will be large enough to share.

Palantir CEO Alex Karp went on CNBC’s Squawk Box last week to discuss a new partnership with Nvidia to deploy open-weight AI models in sovereign government environments. But viewers got a nearly 20-minute broadside against the entire frontier AI model industry, calling it “effing insane” and accusing companies like OpenAI and Anthropic of overcharging enterprises while harvesting their proprietary data.

Days later, Mistral CEO Arthur Mensch made a strikingly similar case on LinkedIn, warning that closed AI providers are gaining “immense leverage” over enterprise customers as organizations connect proprietary workflows to hosted models. He suggests open-weight models, open data systems, and enterprises building their own training flywheels.

The two executives are approaching this from opposite ends of the market, yet their convergence on the same message within the same week underscores architectural control.

Two pitches, one argument

Karp runs a company that sells an application and ontology layer designed to sit between enterprises and the models. The Palantir-Nvidia deal pairs Nvidia’s open Nemotron models with Palantir’s Sovereign AI Operating System, built on AIP, Foundry, Ontology, and Apollo, enabling government agencies and critical infrastructure operators to deploy, fine-tune, and audit AI models within their own air-gapped environments.

When CNBC’s Becky Quick told Karp he sounded angry, he pushed back, saying, “This is the voice of American business that is being channeled through me,” and urged the panelists to call any CEO privately to verify.

“This is the voice of American business that is being channeled through me.”

Mensch’s company sells open-weight models, and he has a custom training platform called Forge, which frames the problem differently but reached the same conclusion. He argues in the post that closed providers have a track record of going after their most successful customers once they learn what those customers are building. His program runs from open models to open data stores, strict access controls, and a continuous training flywheel that improves systems on internal interactions.

Lock-in gets an upgrade 

If you’ve been building software at scale for any length of time, you already know that when new technology arrives, enterprises can’t help but rush to adopt it. But the dependency problem becomes impossible to ignore.

We saw the same problem with cloud computing when companies went all-in on a single hyperscaler’s proprietary services, only to later discover that the cost of switching providers could exceed the cost of staying, even when staying meant overpaying. It’s one of the reasons the industry spent years building abstraction layers, portability tooling, and multi-cloud strategies in response.

Foundation models are raising the same questions, but there’s a twist. When an enterprise connects a model to its internal data, including customer records, proprietary processes, and domain-specific knowledge, the dependency becomes informational. Karp’s argument is that model quality is converging across providers, but the operational leverage accrues to whoever controls the deployment layer and the data flowing through it.

Mensch’s claim has a concrete referent that enterprise architects will recognize. In 2025, Anthropic cut off model access to coding startup Windsurf while building its competing product, Claude Code. The Brookings Institution has separately warned that model providers increasingly compete with their own customers as they chase application-layer revenue.

When access disappears overnight 

When the U.S. government ordered Anthropic to suspend access to its most advanced models for foreign nationals, the company cut access across the board, including to enterprise customers in Europe who had built workflows on top of those models.

Access has since been restored, but for CIOs and enterprise architects who had treated model APIs as stable infrastructure, it was the same as if a cloud provider pulled compute resources without warning. It’s probably why the incident sent European policymakers into overdrive. Mensch, whose company had open-weight alternatives ready, seized the moment.  

Enterprise teams need a plan for when a critical dependency is modified, repriced, or revoked by a provider whose incentives may not always align with their own.

Enterprise teams need a plan for when a critical dependency is modified, repriced, or revoked by a provider whose incentives may not always align with their own.

Architecture shifts toward portability 

Enterprise AI architecture is now essentially this: don’t marry a single provider, build for portability, keep your most sensitive data and logic under your own control.

In practice, this is showing up in three ways.

Firstly, we’re seeing a portfolio approach. A powerful closed model is maintained for complex reasoning and customer-facing work, while an open-weight model is used for repetitive, high-volume tasks. For businesses handling sensitive information, the appeal is that an open model can be run entirely on their own infrastructure, so data never has to leave the organization.

The second is the rise of the model-routing layer, abstraction frameworks that let organizations swap models without rewriting their applications. Palantir’s ontology pitch sits here as does the emerging crop of agent orchestration tools that treat the LLM as a pluggable component behind a standardized interface.

The third is the open-weight movement itself. Nvidia shipped Nemotron 3 Ultra in June under a permissive Linux Foundation license. Meta’s Llama continues to expand. Mistral’s Forge platform lets enterprises train custom models on their own data. And Mistral is teasing an upcoming open-weight model this summer, with early access opening in July.

Follow the commercial incentives

It would be naive to ignore the commercial interests at play. Karp’s Palantir sells the deployment and governance layer; it benefits directly if enterprises treat models as interchangeable commodities. Mensch’s Mistral sells open-weight models and a training platform, and it benefits directly if enterprises distrust closed providers. Zoho’s Sridhar Vembu, who endorsed Karp’s position publicly last week, has his own reasons for wanting enterprises to own their AI infrastructure rather than rent it from Silicon Valley.

But the fact that multiple executives across different market segments are saying companies need to own their data, maintain deployment flexibility, and not hand their competitive advantage to a provider who might become their competitor suggests the argument is resonating.

What developers should watch

If your most sensitive data is flowing through a third-party API with terms of service that can change, you’ve made a governance decision that your compliance team may not have fully evaluated.

For the engineering teams actually building on foundation models, the takeaway is that if your most sensitive data is flowing through a third-party API with terms of service that can change, you’ve made a governance decision that your compliance team may not have fully evaluated.

The engineering choice, to abstract the model layer, to evaluate open-weight options alongside closed APIs, to think about deployment portability the same way you think about cloud portability, is increasingly strategic.

The post Palantir’s Alex Karp and Mistral’s Arthur Mensch agree: AI lock-in is coming for enterprises appeared first on The New Stack.

Distributed AI training is notoriously fragile because losing a single machine typically crashes the entire multi-node job, forcing a time-consuming, full-workload infrastructure restart. To address this, Google’s JAX ecosystem utilizes elastic training via Pathways, which converts a hardware failure into a catchable Python exception so the running process can survive. When an unplanned failure occurs, the system automatically replaces only the broken worker, restores the last viable checkpoint from Cloud Storage, and resumes training in place—minimizing total downtime to under two minutes without ever restarting the main controller process.

For the past year, the awkward secret of the open-weight model boom has been that many of the strongest Chinese releases were off-limits to a large slice of the enterprises most interested in them. License terms that excluded the European Union, the United Kingdom and South Korea meant legal teams killed deployments before engineering teams finished their evals — not just for companies headquartered there, but for any enterprise serving traffic into those regions. For IT teams weighing open models, the trade-offs are unusually explicit.

Tencent just removed that obstacle. The company's Hunyuan team released the full version of Hy3, a 295-billion-parameter Mixture-of-Experts (MoE) model with 21 billion active parameters, and — in a reversal from April's preview release — shipped it under the permissive Apache 2.0 license. The reaction from the open-model community was immediate, with researchers on X singling out the license change as the real headline, and one widely shared post arguing that if the scores hold up, Tencent has just become one of the leaders of open source. Tencent says it will be free on OpenRouter for two weeks. 

The scores are worth scrutinizing — and they don't all point the same direction. But the more interesting story is what Tencent chose to lead with: reliability metrics and deployment economics aimed squarely at production use. 

From preview to product in ten weeks, shaped by 50 internal teams

Hy3's April preview was the first model of Tencent's rebuilt pre-training and reinforcement learning infrastructure, shipped less than three months after the February rebuild. Chief AI Scientist Shunyu Yao framed the early open release as a deliberate move to gather feedback from developers and users before the official version — and Tencent says that's exactly what happened. According to the model card, the team collected feedback from more than 50 product teams after the late-April preview, fixed issues in task execution and interaction, and scaled up its post-training pipeline.

The architecture is unchanged: 295B total parameters, 21B active per forward pass via top-8 routing across 192 experts, a 3.8B-parameter multi-token prediction (MTP) layer for speculative decoding, and a 256K context window. What changed is behavior. Tencent's positioning is that the full release significantly outperforms similar-size models and rivals flagship open-source models with two to five times the parameters.

That "two to five times" framing makes sense for where this model is aimed — and it invites a direct comparison with the current open-weight coding leader, GLM-5.2.

Tencent's blind test favors Hy3 over GLM-5.1, but GLM-5.2 still owns coding

Tencent's headline evaluation is a blind human study rather than a leaderboard. Arguing that public benchmarks don't tell the full story, the company ran a blind test with 270 experts across disciplines working on real-world workflows, collecting 312 valid comparisons, in which Tencent reports that Hy3 scored 2.67 out of 4 against GLM-5.1's 2.51 — with the clearest advantages in frontend development, CI/CD, and data and storage work.

The choice of opponent matters. Zhipu AI released GLM-5.2 in mid-June, and Tencent's own benchmark appendix shows GLM-5.2 ahead of Hy3 across essentially the entire agentic coding suite: SWE-bench Verified (84.2 vs. 78.0), SWE-bench Multilingual (83.0 vs. 75.8), Terminal-Bench 2.1 (81 vs. 71.7) and DeepSWE by a wide margin (46.2 vs. 28.0). The blind test targeted the older model; the newer one keeps the coding crown.

GLM-5.2's coding lead is less surprising once you consider the sizes are side by side: GLM-5.2 is roughly a 744-billion-parameter MoE with around 40 billion active parameters per token, against Hy3's 295 billion total and 21 billion active. Tencent is fielding a model with less than half the parameters — and nearly half the per-token compute — of the one it trails.

Hy3's genuine wins sit elsewhere. On agentic search, it posts 84.2 on BrowseComp and 91.0 on DeepSearchQA — ahead of every open model in Tencent's table and competitive with Claude Opus 4.8 and GPT-5.5. It leads the open field on tool orchestration (79.1 on the public MCP-Atlas set), on agent-harness evaluations like ClawEval, and on long-context retrieval (73.4 on AA-LCR). Read together, the appendix suggests a model that is arguably the best open-weight choice for search-and-tool-heavy agent workloads, while conceding repository-scale coding to GLM-5.2.

One caveat applies to both the wins and the losses: nearly all competitor numbers in Tencent's appendix are marked as coming from Tencent's own test runs. Independent verification, from indices like Artificial Analysis, is still pending as of publication.

The reliability pitch: hallucination rates cut in half

Where the release gets most interesting for enterprise buyers is the set of numbers Tencent chose to emphasize instead of benchmarks. The model card reads less like a leaderboard announcement and more like a production reliability report.

In internal evaluations on real-world scenarios, Tencent says Hy3's hallucination rate dropped compared to the preview version from 12.5% to 5.4%, and commonsense error rates fell from 25.4% to 12.7% — improvements it attributes to fine-grained data cleaning and training constraints built around an explicit behavior pattern: answer when grounded, state when evidence is missing, don't conflate sources, don't fabricate data. Multi-turn behavior gets the same treatment: the issue rate on internal multi-turn tests fell from 17.4% to 7.9%, and Tencent reported that the model's score on the open MRCR long-dialogue benchmark jumped from 42.9% to 75.1%.

Tencent also emphasizes consistency across agent scaffolds — reporting SWE-bench variance within a few points whether the model runs inside Claude Code-style harnesses, Cline or KiloCode. That's an underrated property: enterprises rarely control which agent framework their teams standardize on, and a model that only performs in one harness is a hidden integration cost. These are self-reported internal measurements, and they deserve the same skepticism as any vendor benchmark. But the choice to foreground them at all signals who Tencent believes its customer is: teams that have been burned by models that demo well and fabricate confidently in production.

The deployment math: a 295B model in a 744B world — on export-compliant silicon

The reliability story connects directly to the economics, and this is where Hy3's coding gap against GLM-5.2 starts to look like a deliberate trade rather than a loss.

GLM-5.2 is a roughly 744-billion-parameter MoE with about 40 billion active parameters per token; in FP8, its weights alone consume roughly 744GB, making an 8x H200 node the practical minimum for production serving. Hy3, at 295B total parameters, carries an FP8 footprint of under 300GB — less than half the memory, with roughly half the active parameters per token driving lower per-request compute. For an organization deciding what to self-host, that's the difference between one heavily-specced node and something far more attainable, with room left over for KV cache and batching.

There's a geopolitical wrinkle in the deployment guide worth noticing too: Tencent's recommended serving configuration targets Nvidia’s H20-3e — the memory-boosted variant of the H20, the GPU Nvidia designed specifically to comply with U.S. export restrictions on China. Unlike GLM-5.2, there is no mention of Huawei or Ascend chips here. In other words, the model is sized so that eight of the chips Chinese companies can legally buy comfortably serve it at full precision. That constraint-driven design has a convenient side effect for everyone else: a model that runs well on deliberately capped silicon runs even more comfortably on the H100s, H200s and B200s available in Western data centers, through standard vLLM and SGLang deployments with MTP speculative decoding.

Add the Apache 2.0 license — no regional exclusions, no field-of-use restrictions — and the enterprise equation becomes clear. GLM-5.2 remains the open-weight choice when coding performance is the only criterion and an 8x H200 budget is available. Hy3 makes its case everywhere else: search and tool-heavy agent workloads, reliability-sensitive applications and organizations that want frontier-adjacent capability without frontier-scale infrastructure. The open question is whether Western enterprises, now that the license barrier is gone, will treat a Tencent model as a serious candidate at all — or whether the next Artificial Analysis update settles the benchmark debate before procurement gets the chance.

This three-part blog series explores the growing complementarity between artificial intelligence (AI) and quantum technologies. The first post introduced quantum technologies and outlined their strengths and the challenges of combining AI with quantum systems. The second examined how AI can support the development of quantum technologies, helping to optimise systems and accelerate progress towards practical applications. In this third and final instalment, we turn to the reverse relationship: how quantum technologies – including computing, sensing and communication – could support the future evolution of AI systems.

Although large-scale, fault-tolerant quantum computers remain a long-term goal, early-stage quantum devices and their integration with existing AI systems are already paving the way for quantum-enhanced AI. These developments could eventually lead to meaningful improvements across many sectors of the economy and in daily life.

What technical breakthroughs could quantum computing bring to AI?

In recent years, AI systems have become more powerful and data-intensive, particularly through large language models (LLMs). These developments have made the limitations of classical computing, especially in speed, energy consumption and scalability, increasingly apparent. Quantum technologies could offer a pathway to expand the frontiers of classical computing, potentially overcoming today’s computing bottlenecks. However, quantum computers are not expected to replace existing AI systems. They are likely to coexist, with quantum systems excelling at solving specific types of problems that are difficult or intractable for conventional computers.

As quantum computing evolves, it is expected to strengthen AI in two main ways. First, quantum computers could significantly improve energy efficiency. Training AI requires substantial computing resources that consume vast amounts of electricity, raising economic, energy security, and environmental concerns. This is particularly true of LLMs.

These pressures apply more broadly across statistical AI techniques, such as machine learning, where increasing model complexity and data intensity are pushing the limits of classical computing. This is often characterised by experts as the slowing of or end to Moore’s Law.

Second, quantum computing could enhance the performance and capabilities of AI systems. This has fuelled growing interest in quantum machine learning (QML), which refers to machine learning methods implemented using quantum algorithms. QML is expected to improve AI system learning performance in areas such as optimisation, pattern recognition and high-dimensional data analysis.

In the long term, QML may reduce the computational requirements and energy footprint of some AI workloads by performing complex computations more efficiently, although this remains to be demonstrated at scale. QML remains in an early stage of development and faces three key bottlenecks:

• Data transfer constraints: Moving large volumes of classical data (bits) into and out of quantum systems (qubits) remains slow, limiting the suitability for data-intensive AI tasks.
• Unproven advantage: Demonstrating performance gains over highly optimised algorithms on classical computers remains challenging.
• Unclear hardware requirements: Defining hardware specifications for QML, including qubit counts and coherence times, remains difficult, making it challenging to develop practical QML roadmaps.

For these reasons, experts do not expect widespread commercial QML applications within the next decade.

Near-term pathways to AI applications

While QML is a longer-term prospect, quantum-inspired AI techniques are already delivering practical benefits. These approaches adapt concepts from quantum physics for use on classical computing hardware. One prominent example is the use of tensor networks, originally developed to simulate quantum systems, to compress LLMs.

These techniques have significant practical implications for AI developers. As neural networks such as LLMs become larger and more complex, deployment is constrained not only by fixed hardware limits such as memory and processing capacity but also by the computational and energy costs of training and running these systems.

Tensor-network compression to reduce memory use

Some research has shown that tensor-network compression can reduce memory use and computational demands by 10-100x, often with only modest reductions in accuracy after fine-tuning. These efficiency gains enable advanced AI models to run on conventional CPUs, edge devices and legacy infrastructure, expanding access beyond specialised high-performance computing environments.

Because these techniques do not require quantum hardware, they are already being embedded in commercial applications. Startups and technology providers are offering tensor-network-based compression tools and positioning them as a path to lower costs and energy consumption while improving deployment flexibility.

In practice, quantum-inspired compression is best understood as complementary to established methods such as pruning and quantisation for optimising neural network efficiency. Because they target different forms of redundancy in neural networks, tensor approaches can be combined with conventional techniques and, in some cases, outperform them in both efficiency and performance. Together, these developments illustrate how insights from quantum information science are already shaping the evolution of AI, even before large-scale quantum computers become widely available.

Hybrid quantum-classical computing

The most realistic near-term pathway for combining AI with actual quantum computers is through hybrid quantum-classical systems that leverage the strengths of both AI and quantum computing. In these systems, quantum processors perform specific sub-tasks, such as optimisation or simulation, while classical AI models handle data processing, interpretation and control. In this direction, several computing infrastructures are integrating quantum processors into high-performance computing environments, enabling researchers and industry actors to experiment with quantum-enhanced AI workflows. At the same time, cloud-based quantum platforms are lowering access barriers, allowing AI developers to test quantum algorithms, such as optimisation or sampling routines, that can be integrated into machine learning workflows.

These advances could eventually translate into practical applications across multiple sectors. In materials science and chemistry, quantum computing combined with AI may accelerate the discovery of new materials and drugs by enabling more accurate simulations of molecular behaviour. In manufacturing and logistics, hybrid quantum-AI approaches could improve the performance of complex optimisation tasks such as scheduling, resource allocation and supply chain planning. Financial services actors are also exploring quantum-enhanced modelling for portfolio optimisation and risk analysis, where large combinatorial search spaces pose challenges for classical AI methods.

How could quantum sensing expand AI’s possibilities, and where?

Beyond computing, quantum technologies may also enhance AI through advances in sensing. Quantum sensors can detect extremely small changes in magnetic fields, temperature, motion, or chemical composition, often with higher precision, stability, or spatial resolution than classical devices, thereby producing novel data streams. For AI systems, access to richer and more accurate data can translate directly into new applications across multiple sectors.

In healthcare, more sensitive sensing technologies could enable earlier disease detection and more accurate monitoring of health indicators. By capturing subtle biological or chemical changes that might otherwise go unnoticed, quantum sensors could provide richer data for AI systems to analyse, supporting faster diagnosis and more personalised treatment decisions.

In agriculture, improved sensing precision may help monitor soil conditions, crop health and environmental variables in greater detail, allowing AI tools to optimise irrigation, fertilisation and resource use for higher yields and greater efficiency. Similar approaches could support environmental monitoring, in which higher data quality can strengthen forecasting models and inform policy responses to climate and sustainability challenges.

Quantum sensing also has potential applications in infrastructure and industry. Sensors capable of detecting minute physical changes could help identify early signs of structural stress or equipment degradation in bridges, transport systems or energy facilities. When combined with AI-driven predictive maintenance, this information could enable earlier interventions, reduce operational disruptions and improve safety outcomes. More broadly, the integration of advanced quantum sensing with AI highlights an important dimension of technological progress: improvements in data quality and reliability can be just as transformative as advances in computing capabilities.

How could quantum communication support AI?

Quantum communication could enable secure networking conditions for federated learning and multi-agent systems, where multiple devices collaboratively train models without sharing raw data. Quantum-secure communication channels could make it easier for different parties to share sensitive information (e.g., in healthcare, finance, or critical infrastructure) while reducing the risk of interception or data leakage, especially when training or inference occurs in distributed cloud environments. Research and patent applications are exploring whether entanglement-enabled networks or quantum-secured links could support distributed training architectures across geographically separated computing resources. Although these concepts remain largely experimental, early prototypes are already exploring secure distributed machine learning architectures built on quantum communication protocols.

Quantum communication may also become important for integrating AI with quantum sensing systems. Some advanced quantum sensors generate information directly in quantum states, which cannot always be measured or transmitted using conventional classical channels without losing valuable information. Quantum networking could allow these states to be transferred between devices or processing nodes while preserving their quantum properties, enabling more sophisticated analysis pipelines that combine sensing, computation and AI-driven interpretation.

Leveraging quantum and AI complementarities for a shared technological future

This three-part series examined the potential and challenges of combining AI and quantum technologies, from foundational concepts to emerging applications and future pathways. As we have seen, AI is also accelerating progress in quantum technologies themselves, for example, by improving calibration, noise reduction and experimental design. This bidirectional relationship (AI for quantum and quantum for AI) is likely to shape the next phase of innovation in both fields.

As outlined in this series, the integration of AI and quantum technologies could reshape scientific discovery, healthcare, industry and sustainability. At the same time, significant challenges remain, including technical limitations, talent shortages at the interface between the two technologies, ethical considerations and the need for international collaboration.

As we stand at the early stages of this transformation, one conclusion is clear: the digital future will not be built by AI or quantum technologies alone, but rather through their interplay and collaboration.

Learn more about the OECD work on quantum technologies: www.oecd.org/en/topics/sub-issues/quantum-technologies.html

The post How quantum technologies could open new frontiers for AI appeared first on OECD.AI.

“In the future, the relationship between humans and robots will deepen, and the distinction between them will probably disappear.” This prediction, from one of the attendees at the recent Humanoids Summit in Tokyo, might have been unremarkable had it not come directly from an android that was first introduced to the world 20 years ago.

Geminoid HI-6 is the sixth-generation of a robot originally designed in 2006. The mechanical twin of Osaka University professor Hiroshi Ishiguro, Geminoid HI-6 is now equipped with a large language model trained on Ishiguro’s own writings and interviews. It has advanced conversational skills and can even have a chat with its creator, an eerie spectacle. But at the Humanoids Summit, Geminoid was one of the few humanoid robots from Japan, the country that pioneered the form factor.

While the event in Tokyo had only about 40 robots on display, Chinese systems outnumbered Japanese by roughly three to one. Some Japanese robotics firms were even using Chinese robots in their own technology demonstrations, something that would have been unthinkable in the recent past—one Japanese engineer described the situation as “sad.” The conference was a stark reminder of how Japan has ceded its early lead in humanoid robot development to overseas competitors, and the challenge it now faces to secure a place in an ecosystem increasingly dominated by general-purpose robots powered by AI.

Twenty-five years ago, Japan was turning out groundbreaking humanoids that were showstopping in their abilities, but they were not commercialized as practical machines in any meaningful way. Heavily influenced by science fiction and lacking practical applications, they were mostly expensive technology demonstrations that were eventually mothballed. What Japan retains, however, is robotics design and know-how, which it must leverage to be a key player in the rapidly evolving humanoid ecosystem.

Learning to Walk—Then Standing Still

To anyone who has seen recent videos of Chinese humanoids doing kung-fu and synchronized acrobatics, as well as half-marathon races, China’s remarkable progress in the field is nothing new. At the Humanoids Summit, Toyota showed a video of its latest basketball-playing robot, and Honda exhibited its latest robot hand, but the full-scale humanoids on the floor were mostly Chinese–the kid-size K1 machines from Booster Robotics of Beijing were dancing to Michael Jackson tunes. The full-scale G1 humanoid from Unitree Robotics of Hangzhou was also doing demos.

“You cannot sell these bipedal systems in Japan for safety and compliance reasons,” says Shuichi Nagao, a frequent visitor to China as CTO of Omakase Robotics, a division of Zeals, a Japanese humanoid robot developer. Omakase was exhibiting a G1 modified with an external PC controller, a dextrous hand, a suction-cup manipulator and a sensor “hat” with an extra speaker, mic, and camera.

“In China, the government is pushing humanoid development. They didn’t have an industry 20 years ago. The people pushing it are young, in their 20s and 30s. It’s a really different mentality out there,” says Nagao. “Big players in Japan are still looking for use cases for humanoids. In China, they’re already doing mass production and reducing the cost, so other countries can’t compete with them anymore.”

Another Japanese company showing off G1 bots was summit sponsor GMO AI & Robotics, a subsidiary of Japanese internet company GMO. It’s using the robots in partnership with Japan Airlines to load and unload cargo containers at Tokyo’s Haneda airport. The cargo project is a trial—like many other humanoid experiments—but the fact that Chinese machines have penetrated so far into Japan’s ecosystem upends a long history.

In 1973, scientists at Waseda University in Tokyo built WABOT-1, considered the first full-scale humanoid robot, which was capable of slow bipedal locomotion, grasping objects, and simple communication. It inspired Honda’s groundbreaking Asimo humanoid, but Asimo was never commercialized. It was eventually retired in 2022, the year ChatGPT was released. Two years later, Unitree’s G1 went on sale for US $16,000.

China’s High Torque Technology Co. showed off its Mini Pi biped, customized with an anime-inspired head, at Humanoids Summit in Tokyo. The regular version is priced at $3,500. Tim Hornyak

Supply and Demand

Japan’s development of humanoids happened before practical applications or widespread demand were in place, but bad timing is only part of the story—Japan also has a history of developing technologies that might appeal to domestic consumers but not necessarily those overseas. For example, decades after its highly engineered multifunction toilets first appeared, they have only recently found a following abroad.

Japan’s humanoid prowess was partly built on the back of its legendary industrial automation, yet even that stronghold has eroded. Ani Kelkar, a partner from McKinsey & Company in Boston who produces analytical reports about the robotics industry, told the summit audience that while Japan occupied the top spot in the world in manufacturing robot density (the number of multipurpose industrial robots in operation per 10,000 employees) from at least 1994 to 2009, it then slipped to second in 2014, third in 2019, and fifth in 2024. In that year, South Korea was at the top of the leaderboard with a robot density of 1,220 compared to Japan’s 446.

The International Federation of Robotics estimates China now has the most operational industrial robots in the world, with around 2 million total units, approximately 4.5 times more than Japan. “The annual installation numbers are impressive too: 54 percent of all robots installed worldwide in 2024 were deployed in China,” the IFR said in a release in April 2026.

“I think the loss of Japanese leadership is more to do with the rise of China as a manufacturing powerhouse including for sectors that Japan had high export levels,” Kelkar said in an email interview. “The recovery has not yet happened as Japan “missed” the rapid acceleration in AI for robotics and is now playing catch-up.”

How Japan Can Adapt

Kelkar believes Japan has a $100 billion opportunity in general-purpose robotics, which are machines that can perform a wide variety of tasks, and it cannot rely on the slower-growing industrial robot market, which is centered on factory machines that do one simple and predictable task like welding car parts. He points to a McKinsey white paper suggesting that while Japan has much of the hardware and technology experience needed to support general-purpose robot development, it must change its strategy to capture a larger share in AI, software, data collection, and robotics platforms.

Tetsuya Ogata is a professor of engineering and director of the Institute for AI and Robotics at Waseda University, the birthplace of humanoids in Japan. He briefed the summit on how a nonprofit he chairs, the AI Robot Association (AIRoA), is working with Toyota and other members to develop foundational technologies for collaborative use.

For instance, AIRoA has collected some 80,000 hours of data on remote operation of mobile manipulators, which Ogata believes is the largest dataset of its kind. Using the data, it built and verified vision-language-action (VLA) models, and it has also started data collection for dual-arm mobile manipulation. In an interview, Ogata acknowledged Japan’s struggle to find its place in the changing landscape.

“The world of AI is inherently a game of scale,” says Ogata. “Therefore, Japan’s absolute prerequisite is to secure a competitive baseline of scale—in data, computing resources, and talent. Beyond that, what I consider most critical is a mind-set shift: Rather than trying to hoard scale within a single nation or company, we must grow stronger by collaborating with a diverse ecosystem of domestic and international players.”

Specifically, this means creating a “collaborative domain” to address data—the single biggest bottleneck—through industry-wide cooperation rather than data siloing. By collectively nurturing a precompetitive, shared data infrastructure and foundation model, individual companies can then compete on top of it with their own applications. “By offering this open ‘data ecosystem’ to the world, we can engage global players and establish a ‘third pole’ alongside the U.S. and China,” says Ogata. “I believe this is how Japan can reclaim its global presence.”

In 1999, Japan introduced the world’s first mobile internet services platform. But being first didn’t turn Japan into a smartphone manufacturing or design center—it’s now merely a supplier of parts to other countries that are leading the smartphone industry. If Japan can avoid a repeat of that experience and successfully deregulate, diversity, and commercialize its original humanoid dreams, it stands a better chance of influencing the direction of the industry and reaping billions in value. As automobiles and electronics were pillars of Japan’s industrial strategy in the last century, Japan could make humanoid robots one of its key value generators in the 21st century, an approach that would not only deliver economic benefits but give Japan greater clout in how the industry will evolve. Just like Japanese cars, electronics, and even toilets, Japanese humanoids could stand for craftsmanship and reliability. It’s a legacy that Japan can’t afford to give up.

This article appears in the September 2026 print issue as “Japan Seeks a Humanoid Robot Comeback.”

In April, Andrej Karpathy published a GitHub gist file called “LLM Wiki,” a brief text document designed to help one build a personal knowledge base using LLMs. It’s based on the premise that an AI agent will keep what it knows as linked Markdown files it can read and rewrite, because a language model does not get bored maintaining cross-references and can touch fifteen files in a single pass. It was only a few thousand words with no product attached.

Two months later, Google turned that instinct into a published standard called the Open Knowledge Format. The OKF packages organizational knowledge, metrics, tables, and runbooks as plain Markdown that any agent can read without a proprietary account. Google is careful to call it v0.1 — a starting point rather than a finished standard.

Garry Tan, the Y Combinator president, got there first in a different lane. His gstack, an MIT-licensed Claude Code setup that crossed 66,000 GitHub stars within weeks, comprises 23 specialist roles, each a Markdown file. No runtime; no code; just prose that runs across ten different coding agents.

Markdown has become the substrate agents read and write

Three approaches, three different needs, one common solution. Karpathy sought agent memory, Google aimed for enterprise context in BigQuery agents, and Tan wanted a way to summon an engineering team from a terminal. All three turned to the same basic resource: a folder of Markdown files versioned in git.

Developers had already established this practice. CLAUDE.md and AGENTS.md are present in millions of repositories as the initial files an agent loads. OKF and gstack are the evolved forms of this convention – one focused on what the agent knows, the other on how it behaves.

This is the Git and JSON playbook tied to the agent’s knowledge. The formats that survived are the ones you could start using without changing anything. You can simply cat the file, clone the repo, and any tool you already use can parse it. MCP remains important as the interface an agent connects to. Markdown is becoming the format that carries the content.

The lock-in moved from the model to the files

The significant factor to observe here is the competitive advantage, not technical specifics. For two years, the belief was that owning the best model meant controlling the developer.

This perspective is now shifting. Replacing Claude with GLM or Codex, gstack continues to operate because the core intelligence evolved, but the documentation did not.

The moat is shifting from the model to the Markdown a team owns and accumulates over time.

The moat is shifting from the model to the Markdown a team owns and accumulates over time. A company’s OKF bundle, including its runbooks, metric definitions, and architecture decisions, is, by design, portable across clouds, models, and frameworks.

That kind of portability is the reason vendor-neutral formats exist and why Google’s OKF deserves a closer look.

If no one develops consumers for it, it remains just a good idea that Google released on a slow Friday.

The area where I am most likely mistaken is durability. Declaring Markdown standards is easy, but making them reliable is difficult. OKF is merely a 0.1 draft with a reference implementation, not a full ecosystem. If no one develops consumers for it, it remains just a good idea that Google released on a slow Friday.

The direction remains determined by three separate bets targeting the same file format within a single quarter. Your next agent is likely to interpret its context from a Markdown folder, and the creator of that folder now possesses an advantage that the model vendor cannot easily replicate.

The post Andrej Karpathy, Google and Garry Tan agree Markdown is the answer, but they’re not solving the same problem appeared first on The New Stack.

There's an important distinction between AI that just works today, and AI that lasts at scale. Many companies optimize hard for the first one without ever asking whether they're building the second.

Velocity without discipline and strategic direction is a liability, not an asset. The hardest part of building AI at scale isn't getting a model to work once. It's building systems that continue to work, scale beyond individual teams and use cases, and improve consistently over time.

Today's AI systems do more than just predict and optimize. They converse, reason, and increasingly take action. An autonomous system making decisions on a traveler's behalf creates a very different set of expectations around reliability, governance, and accountability. As AI takes on more of those roles, the principles behind how these systems operate matter more than ever.

We have spent years applying AI and machine learning (ML) across the traveler journey — from personalization, ranking, and recommendations, to fraud prevention, customer support, and, more recently, generative and agentic AI experiences. That depth of experience is what led us to develop a set of ML and AI principles to guide how we build, deploy, and evolve AI systems across our company.

The goal is simple: Make sure the systems we build create real business value, scale, and operate safely. These principles define how we measure, design, govern, and operate our systems.

From principles to practice

Publishing principles is the easy part. The harder and more important work is turning them into operating mechanisms: Recommendations, requirements, tooling, and release processes that teams actually use.

We have begun using 'Agentic Release' tollgates: A set of recommended and, in some cases, required checks before launching agentic AI features. These tollgates translate principles like clear ownership, risk-based governance, evaluation, safe rollout, and monitoring into concrete expectations for teams.

Some of these recommendations and requirements are already being automated and integrated into the software development lifecycle (SDLC). Over time, the goal is for these expectations to become embedded in how we design, evaluate, approve, launch, and monitor AI systems from the start.

Outcomes: Measuring what actually matters

The first test for any model is whether it improves a business outcome and, ultimately, the traveler experience — not whether it just improves a technical metric.

1. Align models to metrics with business impact: Every ML effort must tie directly to a key business outcome or traveler experience metric. Technical optimizations are useful midpoints, not end goals.

2. Optimize for return on cost: The value a model creates has to justify what it costs to develop, train, and monitor, plus the operational complexity it adds. Favor solutions that deliver lasting impact relative to what they cost to run.

3. Justify complexity against strong baselines: Complexity should be earned, not assumed. Start with a strong baseline: An existing general model, a simple heuristic, an off-the-shelf solution. Reach for specialized models or more complex architectures only when simpler options genuinely can't meet the bar.

4. Require both offline and online evaluation: No model goes to broad deployment on offline validation alone or jumps straight to A/B testing. Every model must perform in both offline and online evaluations. Over time, our offline evaluations should reliably predict what we see online.

Design: building systems that scale beyond the teams that build them

Getting a model to work is one challenge. Making its value extend beyond a single team or use case is the harder one.

1. Build on shared foundations; specialize only when justified: Favor shared, platform-wide foundations for core capabilities, data representations, and model building blocks. Specialization should build on those foundations, not spin up isolated stacks, so when the foundation improves, the gains flow across the organization.

2. Treat data as a first-class product: A model's quality is bounded by the quality of its data. We need to maintain robust pipelines, clear lineage, reproducibility, and reusable features built with documented ownership, clear schemas, and SLAs that other teams can rely on.

3. Prioritize generality over local optimization: When two approaches perform similarly, favor the one whose learnings, assets, and operating patterns can be reused across teams, brands, and use cases. We should optimize not just for local performance, but for how quickly improvements can diffuse across the company and compound over time. 

4. Minimize and sunset manual business rules: Manual rules are sometimes necessary for policy, safety, or compliance, but they should be explicit and reviewed regularly, never silent patches for weak models or a source of permanent maintenance debt.

5. Reproducibility and traceability by default: Training data, features, configurations, evaluation results, deployment versions, and key decisions should all be documented and recoverable. That's what lets you debug a production issue months later and hand off ownership without losing institutional knowledge.

Trust: ownership, governance, and operating responsibly at scale

The bar for deploying AI isn't just "does it work?" It's "can we stand behind it?" Trust isn't something you add at the end; it's earned over time and maintained across the full lifecycle of every model we ship.

1. Assign clear ownership and accountability: Every model needs defined ownership across its lifecycle — a business owner, a product owner, an AI owner, and an operational owner. These don't need to be four people, but the responsibilities must be explicit. Who's accountable for outcomes? Who responds if the model drifts? Who answers the incident at 2 a.m.? Without this in place, models become orphaned and problems surface with no one to own them.

2. Adhere to standards and governance: AI and ML models must use approved platforms and comply with established company standards, release gates, and governance processes. Operating outside these guardrails requires a clear, defined path to remediation or deprecation, rather than an open-ended exception. 

3. Govern proportionally to risk: The level of review, evaluation rigor, and human oversight should scale with a model's impact. A customer-facing model that affects pricing or availability for millions of travelers demands a far higher bar than an internal tool used by a small team. For high-impact, safety-sensitive, or highly autonomous systems, human-in-the-loop checkpoints are built in from the start. 

4. Design for fairness, privacy, and transparency: We actively test for unintended bias, have strong data guardrails, and favor explainability when decisions meaningfully affect users. These are incorporated from the start, not added on.

5. Design for safe rollout, rollback, and control: Deployments are progressive, with rollback paths, fallback mechanisms, and circuit breakers ready before launch. The ability to safely undo a deployment matters as much as the ability to ship it.

6. Monitor continuously and adapt: Once live, teams must actively monitor quality, drift, latency, cost, and business performance and retrain or recalibrate when the data shifts. A team should always be able to explain how its model is performing now, not just how it performed when it launched.

These principles do more than define how we build. They define what we're willing to ship and how we stand behind it. In a world where AI systems are increasingly consequential and make real decisions for real travelers and partners, these standards matter. Applied consistently, they build responsible AI that lasts.

Xavi Amatriain is Chief AI and Data Officer at Expedia Group

Xavier will share more details about Expedia's architecture during his session at VB Transform on July 14 at 11:10 am PT. He will discuss: "Expedia's blueprint for building autonomous agents for high-stakes transactional systems."

Interested in attending VB Transform 2026? Register here. A select number of complimentary passes are also available to senior technology leaders. Contact us to get yours.

My octogenarian father-in-law is trilingual and a lifelong fan of the World Cup. As he cheers on his favorite teams in English, Spanish, or French—sometimes switching between them mid-sentence—I’m always amazed at how easy it seems.

Scientists have long been fascinated by the brain’s ability to learn and retain multiple languages. Even after years of disuse, a brief exposure can quickly revive a language without having to consciously relearn its grammar or vocabulary. Bilingualism may offer other cognitive perks. Small studies suggest it delays brain aging, lowers dementia risk, and provides a slight edge in executive function (the ability to stay focused on a goal).

But most  of the evidence is from brain imaging studies that offer only a bird’s-eye view of neural activity and miss the finer details.

Now, scientists from the Baylor College of Medicine and collaborators have recorded activity from single neurons in four bilingual volunteers with epilepsy as they listened, read, and spoke in English and Spanish. The participants already had electrodes implanted in the hippocampus—a brain region critical for learning and memory—to track the source of their seizures.

“This is the very first study to look at how bilingual brains work at the level of individual neurons, and to do so in real time,” said study author Xinyuan Yan in a press release.

The results suggest the bilingual brain operates on two levels. Individual neurons often showed a strong preference for one language when participants heard or spoke words with the same meaning. But networks of neurons were largely language independent. They spontaneously organized into a concept map, placing words with related meanings—such as “dog” and “wolf”—closer together than unrelated words like “fork.”

Surprisingly, both languages relied on the same underlying map. Using the English concept map alone, the team could accurately predict clusters of related Spanish words.

“It’s like looking into a room from a different window. Everything inside is the same, but the perspective is different,” said study author Sameer Sheth.

Bridging Worlds

Language is central to human connection. Although some words don’t directly translate, people can express the same ideas across multiple languages without losing their core meaning.

Children raised in multilingual households are especially adept at switching between languages, often blending words and phrases together. Even when languages differ dramatically in grammar, syntax, and pronunciation, the brain somehow keeps their structures distinct while fluidly merging their meanings.

Long before we learn to speak, neural networks transform thoughts into electrical patterns that form words and sentences. Because languages are built differently—for example, where a verb falls in a sentence—it seems reasonable that each language would have a unique neural fingerprint.

But that might not be the case. A recent AI-powered analysis of functional MRI (fMRI) scans from monolingual speakers of 21 languages suggested that languages share a similar neural scaffold that represents meaning and concepts. Even fictional languages, including Klingon from Star Trek and Na’vi from Avatar, appear to tap into the same underlying system.

A growing body of evidence from bilingual speakers echoes these findings. One fMRI study found native Chinese speakers learned English more efficiently when they recruited brain networks used for Chinese. Another study identified shared speech-related brain activity sufficient for decoding words across languages.

Despite hinting at a universal language map, these standard imaging technologies struggle to capture detailed patterns as people switch languages in real time. To see how bilingual brains actually pull off the feat, we need to listen in on single cells.

Mapping It Out

The team studied four volunteers fluent in English and Spanish. All had learned the languages before age five and continued to use them regularly. Each also had electrodes implanted in the hippocampus to monitor seizures as part of epilepsy treatment, allowing researchers to track individual neuron activity as they listened and spoke.

Though often overlooked in language research, the hippocampus is increasingly recognized as a hub for word meaning, and it may also link concepts together. Here, the team monitored more than 100 neurons in each participant as they completed three language tasks.

First, the participants listened to roughly an hour of YouTube videos and the audiobook Eat Pray Love (Come Reza Ama). Next, they read aloud nearly 100 phrases displayed on a screen, such as “let’s have fun” and its Spanish equivalent “vamos a divertirnos.” Finally, they spent up to 90 minutes chatting with native speakers of each language, discussing everything from family to their epilepsy journey.

By the end, the team had compiled thousands of spoken words, hundreds of matched phrases, and hours of natural conversation.

A Language Landscape

Only a handful of neurons appeared truly bilingual, responding similarly to equivalent words such as “friends” and “amigos.” To better interpret the neural activity, the team turned to mBERT, Google’s multilingual language model that understands more than 100 languages. Like other LLMs, the model represents words according to their relationships and context rather than simple dictionary definitions.

The comparison revealed a similar pattern in brains and machines. Individual neurons rarely encoded the same word across languages. Instead, meaning emerged at the population level.

Both neural activity and mBERT tracked broader context, organizing words into an abstract conceptual landscape called semantic geometry. In this map, related concepts cluster together—“cat” sits closer to “dog” than to “galaxy,” for example—even if the precise features defining those relationships are unclear.

Yet the map remained largely unchanged across languages, suggesting it captured a fundamental mechanism for language processing in the brain. Using the English map alone, the team could predict which Spanish words would cluster around “perro” (or “dog”).

“This is how the brain encodes the meaning of words across languages,” said Yan. “It doesn’t rely on individual neurons translating individual words, but groups of neurons adjusting their activities to create the similar pattern for equivalent words in both languages.”

The study focused on semantics, or meaning, as opposed to syntax, the rules governing sentence structure. A recent study also using single-cell recordings from people with epilepsy suggests that other groups of neurons, particularly those in the frontal parts of the brain, may specialize in grammar while ignoring semantics. Whether they also share a “map” across languages remains to be seen.

The next step is to watch these maps emerge. The team hopes to track people as they learn a new language, revealing how new words and concepts are woven into semantic landscapes in real time. The results could deepen our understanding of one of the most fundamental communication skills and even inspire more capable and efficient language models in AI.

“Our study shows that the brain is wired to learn multiple languages,” said study author Benjamin Hayden.

The post How the Bilingual Brain Switches Languages With Ease appeared first on SingularityHub.

Researchers used quantum-centric supercomputing to simulate molten salts, in an early step toward solving a key problem in fusion power.

Nature Machine Intelligence, Published online: 06 July 2026; doi:10.1038/s42256-026-01262-4

Park and Walsh introduce a reinforcement learning framework that could accelerate the discovery of new, thermodynamically stable and diverse crystalline materials with desired properties.

The FAA is adopting a new airworthiness directive (AD) for all The Boeing Company Model 707, 717, and 727 airplanes; Model DC-8, DC-9, and DC-10 airplanes; Model MD-10 and MD-11 airplanes; Model DC-9-81 (MD-81), DC-9-82 (MD-82), DC-9-83 (MD-83), DC-9-87 (MD-87), and MD-88 airplanes; and Model MD 90-30 airplanes. This AD was prompted by the determination that radio altimeters cannot be relied upon to perform their intended function if they experience interference from wireless broadband operations in the 3.7-3.98 GHz frequency band (5G Lower C- Band) while operating in Canadian airspace, and the determination that during approach, landings, and go-arounds, as a result of this interference, certain airplane systems may not properly function, resulting in increased flightcrew workload while on approach with the flight director, autothrottle, or autopilot engaged, which could result in reduced ability of the flightcrew to maintain safe flight and landing of the airplane. This AD requires revising the existing airplane flight manual (AFM) to incorporate limitations prohibiting certain operations requiring radio altimeter data when operating in Canadian airspace. The FAA is issuing this AD to address the unsafe condition on these products.

Domain-invariant representation learning and domain augmentation algorithms are two principal methodological paradigms for addressing domain generalization. They are widely employed in the machine learning literature to enhance domain invariance and domain diversity, respectively. However, existing risk bounds for domain generalization do not simultaneously capture the contributions of both approaches. This limitation arises because bounds derived directly in the original latent space are typically too coarse-grained and ambiguous to characterize how invariance and diversity jointly influence generalization. Since these two properties are often regarded as being inherently contradictory, it becomes difficult to disentangle and rigorously characterize their individual effects. To address this issue, we first observe that the latent representation space can be decomposed into several distinct subspaces, each exhibiting different characteristics and therefore being better suited for analyzing the respective roles of domain invariance and domain diversity. Building on this observation, we propose a unified analytical framework for domain generalization. Specifically, we introduce a Tri-Space Latent Representation and establish its unique decomposability via a direct-sum decomposition. Under this decomposition, each data representation can be uniquely partitioned into three components: domain-invariant features, spurious invariant features, and domain-variant features. Within this framework, we derive a finer-grained bound on the target-domain risk, which consists of two principal terms corresponding to domain diversity and invariant factors. By theoretically analyzing these two terms, we show that domain-invariant representation learning and domain augmentation are both effective and, crucially, compatible strategies for addressing domain generalization. Finally, we design two sets of experiments to empirically validate the relationship between domain invariance and domain diversity, and to examine their respective effects on domain generalization performance.

We study the high-dimensional training dynamics of a shallow neural network with quadratic activation in a teacher--student setup. We focus on the extensive-width regime, where the teacher and student network widths scale proportionally with the input dimension, and the sample size grows quadratically. This scaling aims to describe overparameterized neural networks in which feature learning still plays a central role. In the high-dimensional limit, we derive a dynamical characterization of the gradient flow, in the spirit of dynamical mean-field theory (DMFT). Under $\ell_2$-regularization, we analyze these equations at long times and characterize the performance and spectral properties of the resulting estimator. This result provides a quantitative understanding of the effect of overparameterization on learning and generalization, and reveals a double descent phenomenon in the presence of label noise, where generalization improves beyond interpolation. In the small regularization limit, we obtain an exact expression for the perfect recovery threshold as a function of the network widths, providing a precise characterization of how overparameterization influences recovery.

Auto-Regressive Video Diffusion Models (AR-VDMs) have shown strong capabilities in generating long, photorealistic videos, but suffer from two key limitations: (i) history forgetting, where the model loses track of previously generated content, and (ii) temporal degradation, where frame quality deteriorates over time. Yet a rigorous theoretical analysis of these phenomena is lacking, and existing empirical understanding remains insufficiently grounded. In this paper, we introduce Meta-ARVDM, a unified analytical framework that studies both errors through the shared autoregressive structure of AR-VDMs. We show that history forgetting is characterized by the conditional mutual information between the generated output and preceding frames, conditioned on inputs, and prove that incorporating more past frames monotonically alleviates history forgetting, thereby theoretically justifying a common belief in existing works. Moreover, our theory reveals that standard metrics fail to capture this effect, motivating a new evaluation protocol based on a “needle-in-a-haystack” task in closed-ended environments (DMLab and Minecraft). We further show that temporal degradation can be quantified by the cumulative sum of per-step errors, enabling prediction of degradation for different schedulers without video rollout. Finally, our evaluation uncovers a strong empirical correlation between history forgetting and temporal degradation, a connection not previously reported.