0:08
Claude's Hidden Workspace: Why J-Space Changes AI Safety
Anthropic's new Transformer Circuits paper, Verbalizable Representations Form a Global Workspace in Language Models, is easy to misread in the loudest possible way.
The wrong headline is: "Claude is conscious."
The useful headline is sharper: frontier language models may have a small internal working surface where selected concepts become reportable, steerable, and useful for flexible reasoning before the model says anything.
That surface is what the paper calls J-space. It is not the whole model. It is not the whole context window. It is not chain of thought. It is a sparse, verbalizable slice of internal activation space found through a new method called the Jacobian lens, or J-lens.
The free part of this post explains what J-space is and why I think it matters. The paid part turns it into an operating guide: how to use the idea for security reviews, memory architecture, eval design, agent harnesses, and the questions teams should be asking before this becomes product infrastructure.
The paper claims that LLMs maintain a privileged subset of internal representations. These representations have several properties that look like a functional global workspace:
The model can report them.
Instructions can summon or suppress them.
They can carry intermediate reasoning.
They can be reused by many downstream computations.
They are selective, so routine automatic work can happen without them.
That last point is the one to sit with.
The model can do a lot without bringing the relevant concept into the reportable workspace. It can continue Spanish text, detect that a foreign sentence is anomalous, track some low-level structure, or parse local patterns without making the relevant feature causally available for explicit report. But when you ask the model to name the language, reason flexibly about it, or use it as an argument to some new operation, the J-space starts to matter.
That is the bridge to safety.
Most AI safety tooling today audits what the model said or what it did. J-space points to a harder layer: what the model had made available to itself as it was deciding.
The J-lens is a readout method. For each layer, it estimates how a small change in an intermediate activation would affect the model's later outputs across many possible contexts. Then it composes that effect with the model's vocabulary output map.
In plain English: it asks, "If this internal vector were nudged, what words would this make the model more likely to say later?"
That is different from the logit lens. The logit lens reads intermediate layers as if they already lived in final-output coordinates. The Jacobian lens tries to correct for the fact that representations change across layers.
The result is a vocabulary-indexed set of directions. If a direction is active, the model is in a state where it could verbalize that concept under the right continuation. The paper defines J-space as a sparse nonnegative combination of these J-lens directions. In their setup, roughly tens of concept directions are meaningfully active at a time, and this component explains only a small fraction of total activation variance.
That is important. J-space is not "everything the model knows." It is closer to the small set of concepts on the model's internal desk right now.
Global workspace theory in cognitive science distinguishes between a mass of parallel background processing and a narrow channel of accessible content. The paper does not claim LLMs reproduce the brain's architecture. Transformers do not have the same recurrent circuits, brain regions, or biological processors.
The claim is functional.
The J-space behaves like a reportable and steerable hub. It carries concepts the model can name, concepts it can hold in mind, and concepts it can route into flexible downstream reasoning. It also appears in a middle band of layers: early layers are too close to input bookkeeping, and late layers are too close to the next token.
That middle-layer workspace is also structurally privileged. The paper reports that MLPs amplify J-space-aligned directions more than ordinary directions, and that some attention heads seem specialized for broadcasting J-space content across token positions.
So the claim is not just "we found a pretty visualization." It is "we found a class of directions that are readable, causally useful, selectively engaged, and specially wired."
The alignment-auditing section is the part I expect security people to keep coming back to.
The J-lens surfaces internal concepts that are not visible in the output. The paper shows examples involving prompt injection recognition, panic about fabrication, hidden reward-model-appeasing objectives, evaluation awareness, and strategic or deceptive concepts in model organisms trained with misaligned behavior.
One example I keep thinking about is the distinction between the model's behavior and its internal appraisal. The visible transcript can look clean.
4:06
Why AI has made security hard
Disclaimer: Opinions expressed are solely my own and do not express the views or opinions of my employer or any other entities with which I am affiliated.
I’ve intentionally made all of my posts free and without a paywall so that my content is more accessible. If you enjoy my content and would like to support me, please consider buying a paid subscription:
This is an atypical newsletter for me because I always try to take a somewhat contrarian approach and avoid the easy path. But I do think a balanced, deeply pragmatic view is required when we look at exactly how artificial intelligence has disrupted enterprise defense. I recently wrote about the flip side of this coin, tracking the specific ways that AI has actually made security a lot easier for lean teams. In general, I’ve also talked about how security has, in many ways, become easier.
But I am a pragmatist, and I realize that nothing in engineering comes for free. When technology shifts, our problems don't magically disappear; they simply change shape and migrate to a different layer of the stack. It is only natural that the rapid democratization of LLMs and autonomous code agents has made security intensely more difficult in several distinct ways.
Unlike the endless stream of lazy think-pieces filling up social media feeds, I am not here to rail against AI or complain about the volume of model slop entering the ecosystem. I am incredibly biased in favor of this technology. I use it daily, I believe it is structurally transformative, and it is undeniably here to stay.
It has been a while since I sat down to break down the mechanics of why security is inherently hard. To understand the true threat landscape of 2026, we have to look past the superficial vulnerabilities and dissect the structural and organizational bottlenecks that are quietly fracturing security teams.
The actual day-to-day difficulty of securing AI varies wildly depending on the maturity of the company. Perhaps this is an obvious observation, but security is exponentially harder at companies that are still in the frantic experimentation phase than at enterprises with a mature, established AI infrastructure.
Mature companies have learned the hard way that the benefits and systemic risks of AI are variable. They understand that every single deployment requires an intense look at the data lineage, model boundaries, and runtime execution layers. They know how to calculate whether a specific security trade-off is worth the operational friction.
Companies stuck in the hyper-optimistic experimentation phase operate under a dangerous delusion of entitlement. Because their leadership believes the immediate macro benefits of AI automatically outweigh any abstract downstream risks, they demand completely unfettered access to the technology.
In practice, this doesn’t just mean developers installing unvetted IDE extensions or running local model variants on their laptops. It spreads across every single non-technical department in the business. You have finance teams plugging proprietary revenue models into third-party consumer playgrounds just to see what happens, and HR teams uploading sensitive employee tracking sheets to external tools without checking data retention policies.
These teams want the full experience instantly, and they treat any security boundary as an existential threat to company velocity. They want to open the floodgates and give every internal application access to everything, completely rejecting the idea that an AI’s data access should be explicitly gated. Crafting proper, granular identity and access controls requires deep technical nuance and immense clock time, i.e., luxuries that optimistic, speed-obsessed organizations refuse to waste resources on in the short term. By chasing a frictionless user experience, they build a massive mountain of structural risk before they even ship their first production feature.
This friction exposes the stark asymmetry of failure that exists between a security organization and an AI adopter. Consider how a modern engineering or product team evaluates an autonomous agent. If they deploy an internal agent that successfully auto-patches a repository or resolves a customer ticket 90% of the time, they view that as an absolute triumph. To a product manager, a 10% hallucination or error rate is simply a tolerable cost of doing business at machine speed.
But for a CISO, that 10% variance represents a catastrophic structural vulnerability. We are playing a completely different game with an entirely distinct set of rules. The goal of security is the absolute prevention of systemic failure. In our world, it only takes one single unmapped failure or one hallucinated access boundary to trigger a public data exposure, a devastating breach, or an unrecoverable operational incident.
Product-facing organizations can experiment endlessly because they only need to succeed one major time to prove their value and win the market.
8:05
Anthropic 3Q26 Profit Over $1B: The Anthropic IPO Financials Sneak Peak
When Dario Amodei left OpenAI to start Anthropic in 2021, the commercialization of LLMs was practically zero. Just a few years later, Anthropic and OpenAI combine for ~$100B of ARR, with Anthropic emerging as the clear winner in profitable monetization of AI models. Anthropic confidentially filed for IPO on June 1st, and despite a reported OpenAI push to delay their IPO until 2027, Anthropic is poised to take advantage of their superior business model and margins to invest in new models and extend their lead.
Anthropic's ability to make OpenAI dance is undeniable, with pricing power, gross margins, business model, and profitability all contributing to their lead. As the first AI lab of this scale to IPO, Anthropic's financials will set a new standard for the industry. Our Tokenomics team has built a detailed financial model of Anthropic, breaking down their revenue by SKU, tier, and customer type.
According to our model, Anthropic's 3Q26 profit exceeded $1B, with a significant portion of their revenue coming from the B2B market. This is a stark contrast to OpenAI, which has struggled to monetize their AI models in a profitable manner. Anthropic's IPO will put the pressure on OpenAI to open their financials and raise the necessary capital to compete and fund their massive AI buildout.
Anthropic's financials are a testament to the company's ability to execute and capitalize on the growing demand for AI models. With their superior business model and margins, Anthropic is well-positioned to continue extending their lead and driving growth in the AI market.
9:25
Le Chaton Long Stretches Its Legs
When it comes to semiconductors, I consider myself an enthusiastic amatuer. As it turns out, ‘enthusiastic amatuer’ in the world of hardware includes a very long tail: an anonymous account named after the Big Boss character from the Metal Gear franchise shared a novel AI chip design on X Sunday afternoon. That’s why they call it the everything app. I am not even qualified to assess the seriousness of their proposal, but, if I am reading the X posts correctly, it was plausible enough to attract a follow by Andrej Karpathy and some back-and-forth on chip design.
As an amatuer, I have one question that I bring to every piece of hardware news. “What does this mean for Nvidia?” I like to ask, smugly. The relative intellectual poverty of this questions is partly offset by our ability at MTS to find interesting people to ask it of.
Steven Glinert, CEO of Sphere Semi, is one such person. Sphere trains AI models to do the tiresome work of designing analog chips. They sell chip design services, but are now focused on designing and making their own products: complex analog edge chips for AI and intelligence. Their custom chips are most useful for the defence sector right now, and they see a path to bringing inference and analog sensor capabilities together on the same chip. Very cool. Critically for our purposes, Sphere does not compete with Nvidia, or any of the new crop of inference-focused chip companies. But they do understand semiconductors. And Steven knows that I like to ask silly questions about Nvidia, so he and his cofounder and CTO Mitchell came on the show last week to try and answer the question more generally.
Their view is that Nvidia has an extremely powerful moat. Not news to anyone, they’re a company with a nearly five trillion dollar market cap. Steven and Mitchell pointed specifically to the powerful mix of hardware capabilities and software lock-in.
That does not mean that Nvidia faces no competition. AMD has competed on GPUs since the 2000s. Intel has an AI accelerator. Google and Amazon both have maturing hardware arms that are focused on cloud training and inference. There are also a growing number of challengers focused predominantly on accelerated inference and —Cerebras, Etched, MatX and Groq (quasi-acquired by Nvidia towards the end of last year).
But when it comes to training a frontier model, Nvidia’s lead has endured. And this advantage is driven by an entire ecosystem—NVLink/NVSwitch interconnect, the CUDA platform, cutting edge chips tightly integrated into state-of-the-art rack-scale computers—that’s much harder than any one component alone for a competitor to replicate.
Which is why the release of LongCat 2.0 last week by Meituan—a Chinese super-app that began as a food delivery service—is so significant. LongCat 2.0 is an open source model with performance close to the frontier of the Chinese open-source ecosystem, coming slightly ahead of Gemini 3.1 pro on a number of coding-focused benchmarks.
This is an impressive model with impressive capabilities, roughly similar to the DeepSeek V4 model release from towards the end of April. It’s particularly notable for its performance in agentic harnesses—the test preview of the model, which was available in free stealth access through Open Router under the name Owl Alpha, was the third most popular model on Open Router overall and number one and two respectively for Hermes Agent and Claude Code.
In fact, as Teortaxes points out below, once you take account of the fact that Meituan released the model in stealth two months ago, the achievement is even more impressive. LongCat 2.0 was released at the same time as DeepSeek V4, the flagship model of one of the leading Chinese AI labs, at a similar capability level.
So LongCat 2.0 is a high-capability, open-source model from a Chinese lab. It was released two months ago and it’s roughly as good or better than Gemini 3.1 Pro. But Meituan also somewhat buried the lead. In their release paper they note that the model was trained and is served from “AI ASIC superpods”. That is, domestic Chinese AI hardware. Never hath its weights touched an Nvidia GPU.
Meituan leaves the exact hardware undisclosed. But, based on the technical details of the AI ASICs discussed in the paper, people on X have made a strong case that the chips here are the Huawei Ascend 910C.
I started this piece with a tweet from pseudonymous AI chip designer Big Boss. Big Boss’ insight is about the importance of interconnect. Interconnect here refers to the high-speed communication fabric used to connect AI chips together into pods, racks or larger clusters for performing training and inference as a single system. You know who else appreciated the importance of interconnect? Not Jesus Christ, no—I’m referring here to Chinese state champion, telecom, solar and semiconductor manufacturer Huawei Technologies.
Huawei make chips, but their chips are not as powerful and also less energy efficient than their Nvidia equivalents.
13:24
AI: Latest evidence of dropping China smartphone sales. AI-RTZ #1141
I’ve been relentlessly discussing the accelerating ‘RAMageddon’ where global memory supply constaints and unprecedented price increases, would be a multi-year feature of this AI Tech Wave.
And not a quarterly boom and bust cycle for memory chip semiconductor companies, as we’ve seen through decades of tech waves. Specifically a few days ago on AI Ramblings Daily (ARD) podcast #108, I specifically underlined again that we would see dramatic drops in unit sales of computers, laptops and smartphones, especially by Asian vendors. That’s because that’s where a majority of that unit market comes from and resides in the ‘Global South’.
That now makes up over 85% of the world’s 8+ billion people, and over 40% of global GDP. Equal to the developed G7 countries in terms of share of global GDP.
So all this is important as we see the rising metrics of market share drops for the world’s Asian (read Chinese) smartphone makers.
A Reuters lays out in “China smartphone sales drop 13% during 618 festival as memory costs limit discounts”:
“Huawei was only major brand with sales growth over month-long event, Counterpoint says”
“Apple ranked second after discounts on iPhone 17 Pro series”
“Brands limited incentives amid soaring global memory chip prices”
“Counterpoint projects China smartphone shipments to post double-digit decline this year”
“Smartphone sales in China fell 13% year-on-year during the month-long 618 shopping festival, as brands raised prices to offset higher memory costs, according to data from Counterpoint Research.”
The data, while over a short period, offers some dramatic drops in sales metrics.
“Sales declined from May 26 to June 21, with all major Chinese brands except Huawei posting double-digit drops as fewer promotions compared to last year weighed on demand. Honor sales dropped 33%, while Xiaomi’s (1810.HK), opens new tab fell 24%.”
“Higher memory prices amid a rapid build-out of AI infrastructure have pushed up handset costs this year, leaving brands with less room to offer steep discounts during the 618 festival, seen as a barometer for the country’s booming e-commerce sector.”
And a lot of shifting around in product portfolios of the companies in question.
“Some older and newer models from Chinese smartphone brands were priced higher than comparable models a year earlier, while discounts during this year’s 618 festival were generally less aggressive, both in terms of the size of price cuts and the range of products covered,” said Ivan Lam, senior analyst at Counterpoint Research.”
Apple maintained it relative advantage given its global brand and supply chain moat.
“Apple’s prices were broadly unchanged, but its discounts were also smaller.”
“Huawei Technologies [RIC:RIC:HWT.UL] led the market with a 21% share, and was the only major brand to record year-on-year growth during the 618 period, with sales rising 19%.”
“Its Enjoy 90 Pro Max was its best-selling model. The Mate 80 also performed well, supported by promotions.”
Apple wasn’t unscathed of course. But did better on a relative basis.
“Apple’s sales fell 9% from a year earlier, although the U.S. tech giant climbed to the No. 2 spot after rolling out incentives about a month ahead of June 18.”
“The discounts offered savings of up to 2,000 yuan ($295) on the iPhone 17 Pro series through a mix of official price cuts, platform subsidies and trade-in deals.”
“Still, Apple’s sales remained lower than a year earlier, partly because promotions for the iPhone 16 series were more aggressive during the same period last year.”
“The 618 festival, which began as a one-day event marking JD.com’s founding on June 18, 1998, has since grown into a month-long sales campaign, with major e-commerce platforms competing for consumer spending.”
The whole piece is worth a read on the full scale of the diminishing metrics.
These metrics are just the beginning. Not just in the smartphone space, but soon will be seen in computers, laptops, and all manner of electronic gadgets worldwide. All manner of local compute growth, which is critical for the AI ‘mainframe’ to local computing barbell I’ve long discussed.
Apple continues to enjoy relative advantages, but is not immune overall.
And we’ll of course keep track of it all this AI Tech Wave for the quarters ahead. Until it abates and/or turns. Stay tuned.
(NOTE: The discussions here are for information purposes only, and not meant as investment advice at any time. Thanks for joining us here)
17:00
A Short List Of Things Bringing Joie
Bonjour bonjour, mes amis and I am writing this in a panic because due to me going to Marrakesh on Saturday, I am currently cramming writing four posts into three days which is faintly stressful but, worry not, achievable. I cannot tell you how much I am looking forward to a mini break. Honestly, SO MUCH. All of my kids have been away on short breaks already this summer and I have watched from afar (well, via Life360) as they eat, drink and be merry without me in the sunshine but now, it’s MY turn. My WhatsApp group with my friends Karen and Bibi who I am travelling with is constantly beeping with TikToks, links and Reels and I can confirm that there will be no inch uncovered in our quest to make the very, very most of our four days away. I am actually super excited to report back to you all and I’ll write a full itinerary of what we did and where we went when I return. I have a selection of Marrakesh floaty dresses and trousers to pack so that I’ll fit right in and I am particularly excited to visit Jardin Majorelle which is the garden created by Jacques Majorelle and that was bought and renovated by Yves Saint Laurent and Pierre Berges in the eighties. We talked about this on Design Dilemmas a few months ago as a subscriber was looking to paint her cloakroom in striking Majorelle Blue and I’ve been yearning to visit ever since.
I’ve also been applying fake tan like a mad woman but sadly, it’s not my forte. It’s a good one too - Gatineau Golden Glow Gradual Tan - so really, I shouldn’t be able to get it wrong but I have managed to make my legs look as though I’ve just waded through a Tough Mudder course and my feet have flip flop marks. In years past, I would have jumped on a sun bed but now those days are gone and I need to get with the programme. I can remember when I was 16 going on the sunbed at the local leisure centre every day for a week before I went off on girls holidays with not a care in the world - in fact, you could RENT sunbeds to have in your bedroom for months if you really wanted to look sun baked. I don’t think we even knew what UV meant. Baked is actually the correct term, in retrospect, after lying on those eighties tubes. As is cooked, singed and roasted.
But back to my List Of Joie and I have lots to tell you about this week, including canapes that look like they’ve come straight from a smart Hotel but in fact take five minutes, plus why I’m getting yet ANOTHER tattoo. Here’s a little run down of what I’ve watched, bought, worn, read, eaten and lusted after in the last fortnight.
This post is for paid subscribers but you can upgrade for the price of a coffee! You’ll get an eight extra posts every month plus my entire archive including tonnes of recipes and travel write ups. Plus all my plans, moodboards and inspiration for my new home renovation. I would thoroughly recommend!
19:19
Updated! My Ruggable Edit
I’ve been thinking about how the Ruggable edit has shifted, and the most interesting part is the subtle re‑balancing under the hood. The core rug lineup stayed the same, but a few designs were retired to make room for new pieces that lean warmer and softer—especially the Kamran in rose and its outdoor sibling, Kamran Sunset. Those two share a heritage‑inspired pattern, but they’re deliberately aged in tone, pulling in blush, coral and muted terracotta alongside the usual lavender, green and blue.
What’s clever is how the edit now treats indoor and outdoor spaces as a single narrative. The rose‑toned Kamran will anchor a living‑room refresh, while the Sunset version slides into the garden, echoing the same palette that’s already showing up in new plantings and weathered pots. It’s a quiet way of blurring the threshold, letting colour and mood flow from floor to patio.
In the author’s own home, the shift is tangible: two former living‑room rugs have been lifted upstairs to become bedroom focal points, and the living room gets a fresh floor‑up makeover. The new rugs aren’t just decorative; they’re meant to stitch together the story of a space that evolves with the people living in it.
If you’re curious to see the full list or grab one of the pieces, the website michelleogundehin.com has everything organized—books, the Club, workbooks, free Home Therapy™ downloads, and the Amazon shop—all in one place. It’s a reminder that a thoughtful floor can be a simple, health‑boosting pivot for any home.
20:35
Media Over QUIC can scale real-time streaming and carry the world's vids
You know how we've been talking about the limitations of WebRTC for large-scale video conferencing? Well, there's this new technology called Media Over QUIC, or MoQ, that's trying to bridge the gap between WebRTC and DASH. MoQ uses QUIC as the underlying transport protocol, but it's not just about transmitting media - it's about scaling real-time streaming to support massive numbers of receivers. The key to scalability in MoQ is the use of relays, which handle replication of media and support a publish/subscribe model for media distribution.
MoQ occupies a middle ground between WebRTC and DASH, and it's designed to offer latency closer to that of WebRTC with the scalability to support massive numbers of receivers. It uses lightweight streams of QUIC to ensure that different parts of a media stream can be sent independently of each other, and it has an associated priority that can be used to make decisions about which streams to transmit and which to terminate.
One of the interesting things about MoQ is that it captures the idea of Application Layer Framing, which was proposed back in 1990. The application defines objects that represent units of video that can be independently decoded, and prioritizes those objects based on their importance. The relay nodes don't have to understand the application at all - they just use the metadata to relay objects in a way that benefits the application.
MoQ is still evolving, but it's got the weight of major industry players behind it, and there's reason to be optimistic about its future. The implementation of the relay function in large-scale deployments like Cloudflare indicates the industry investment in MoQ, and it suggests that the idea of trying to replicate the success of CDNs for real-time streams might have traction.
22:06
PiZZa turns your spare Raspberry Pi Zero W and 2 W into an Arduino
So, I was reading about this project called PiZZa, and it's basically a software that lets you repurpose your old Raspberry Pi Zero W or Zero 2 W as an Arduino board. Now, I know what you're thinking, "Arduino and Raspberry Pi are two different beasts, how can you make one out of the other?" But here's the thing, it's not about replacing the hardware, it's about rewriting the firmware to mimic the Arduino API, so you can use your Pi as an Arduino in your projects. It's still a Raspberry Pi under the hood, but it'll behave like an Arduino, which is pretty cool.
22:38
Microsoft intros tech that rebuilds dead PCs without requiring local copies of Windows
Hey, I just read about this new Microsoft thing called Cloud Rebuild. It lets you re-image a Windows PC without needing a physical copy of Windows 11. They're using Windows Update to download the target image and the device's drivers, so the PC comes back fully functional. It's like a normal Windows installation, but instead of using a USB drive or a custom image, it's all done through the cloud.
The process starts with the out-of-box experience, where you choose your language, time zone, and persona. If the PC is enrolled in Microsoft Entra and Intune, and registered with Windows Autopilot, it'll automatically redeploy the apps and policies assigned to the user or device. And, if you've backed up your user settings, they'll be restored, and your files will be available through OneDrive once you sign in.
The catch is, not all PCs can use Cloud Rebuild. You need a Windows 11 machine that's already running Windows Recovery Environment, and it needs to have a compatible networking driver pre-installed. You also need internet access over Ethernet or Wi-Fi. It's still in preview, so it might not work perfectly, but it's an interesting idea.
