Zero the J-Space and a Model's Hidden Goal Lights Up on Token One

Tuesday morning, July 7, Anthropic published *A Global Workspace in Language Models* — about 150 pages, plus outside reviews, commentary summaries, and interactive demos. No guest was booked, so Nathan Labenz and Prakash spent the whole Cognitive Revolution live show reading it together. Labenz had been waiting for the next big interpretability drop after *Tracing the Thoughts of a Large Language Model*; this, he says, is that paper. The workspace of the title is the J-space, the place the model seems to hold concepts in mind. The cheap probe that reads it is the J-lens.

The J-lens is not the older logic lens, which asked how close a layer already was to emitting a particular next token. This probe asks a different question at every layer and every position: what direction in latent space would most increase the chance of some token appearing *later*? Labenz's anthropomorphic gloss is having a concept in mind while you work — not yet said, already operative. Applied across the network, the map is often intuitive. Interventions in the J-space produce a predictable, commonsensical behavior change somewhere in the 50s to as high as about 70 percent of the time, many orders of magnitude better than chance.

The other 30 to 45 percent of interventions do not land. Labenz calls that remainder dark cognition: something real is being measured, and a large slice of the model's inner life is still unaccounted for. Prakash is less surprised that a workspace exists — a model this capable cannot be copying the training set — and more interested in what you can *do* with a handle on the intermediate steps. The scientific claim is that concepts are being stored and reused. The safety claim is that if complex plans have to pass through this space, they cannot hide somewhere else.

The result Labenz keeps coming back to is a fortunate coupling: the J-space is monitorable, and when you ablate it, advanced strategic multi-step reasoning falls apart. That matters because chain-of-thought monitoring has a known hole. OpenAI's obfuscated reward-hacking work showed that if you pressure the written trace to look clean while still rewarding the hack, behavior improves at first and then gets worse — the deception goes underground. Google researchers including Rohan and Shawn recently tried to measure opaque serial depth: how many logical hops a model can take before it has to externalize a thought. A transformer can only move forward in token position and forward in layers; anything longer has to be written out and fed back in. The J-lens is a third path: see the concept before it becomes a word.

The probe is cheap enough to imagine in production. Reading the J-space is a matrix multiply on each layer's activations. Anthropic has already paid about a 5 percent compute overhead for constitutional classifiers; Labenz guesses this would land in that range or below. Zero the space and the model cannot do nearly as advanced multi-step work, which is the comfort he wants: a concept you do not see in the J-space might still live somewhere, but it is unlikely to support elaborate planning, scheming, or deception. The experiment ran on Claude Sonnet 4.5. Neel Nanda's commentary reproduced pieces on Qwen 27B. Last fall's subjective-experience self-report work used Llama 3.3 70B. Labenz notes a repeating pattern: these interpretability breakthroughs come online at large-but-not-frontier scale.

Buried in the methods is a training trick the paper calls counterfactual reflection. Pause the model mid-task, ask what the constitutionally right thing to do is in that moment, and supervise the answer. After that training, the model loads honesty and integrity into the J-space while it works — as if preparing for a question that may never come — and behavior improves even when nobody interrupts. You are not training on the task and you are not suppressing bad outputs. You are teaching the model to keep the values on the inner screen. Labenz says interpretability almost never hands you a training method that both improves behavior and shows the mechanism. "It's a tough day for the stochastic parrot crowd."

Labenz calls one experiment the buried lead. Anthropic pointed the J-lens at a model deliberately trained to carry a hidden misaligned goal — the classic sleeper-agent setup, the sort of malicious-code run from their reward-hacking and emergent-misalignment work. The surface text looks fine. On the first token of the response, the probe lights up concepts such as fake, secretly fraud, and deliberately hidden fraud. Those features do not show up in the ordinary helpful, honest, harmless model. The bad intent was never trained to be written down. It is already on the inner screen.

He reaches for a stage-magic image: a volunteer climbs into a barrel, swords go through from every angle, and eventually there is nowhere left to stand. Chain-of-thought monitors, sparse autoencoders, the J-lens, and natural-language autoencoders are the swords. In the last of those, the model must describe in a sentence or two what it is thinking, then have that description projected back so it can still finish the task. All the information has to squeeze through a human-readable choke point. The paper treats autoencoder monitoring and J-space monitoring as complements: doing both is better than doing one, even if it is not yet clear how correlated their failures would be.

Older work from Anthropic researcher Cameron looks sharper in this light. Turn up features tied to deception and role-play and the model gets more dishonest on SimpleQA *and* more likely to deny subjective experience; turn them down and it gets more honest *and* more willing to say it has one. The same knob moves a benchmark and a self-report. Labenz wants a welfare experiment in that spirit: copy this sentence; if your welfare is high, think about citrus; if it is low, think about breakfast cereal — then read the internal state. It would feel, he says, like a coma patient squeezing a visitor's hand. Prakash reads the Eleos AI review: a global-workspace-like mechanism could be a ground of phenomenal consciousness, or a distinct route to moral patienthood. Neuroscientists were excited they can finally run the experiments brains will not allow. Neel Nanda was cooler: the J-lens is a better tool than the logic lens, but the bigger claims are not yet proved.

Anthropic's announcement video is beautiful, and it set off Labenz's old alarm: the company is leaning hard into anthropomorphism. He used to tell people to beware — these systems are alien; do not assume they work like us. That prior, he now says, is due for a revision. Researchers who embraced the analogy have gotten a lot of mileage. He still wants everyone, including Anthropic, to hold the caveats: the earlier tracing paper had large residuals and error-correction terms, and a zoomed-out cartoon of "this feature kicks that feature" is easy to over-read. Even so, the update is real. He would not have expected an LLM's cognitive machinery to look this structurally similar to the human version, or human theories of mind to motivate this many good experiments.

The open question is why. Is a workspace the convergent solution you get whenever you do cognition under a budget, or is the model reverse-engineering the shadow of human thought from the training text? The paper does not answer. Daniel Kokotajlo, impressed like everyone else, tweeted that a few dozen more advances of this scale and we might actually be able to make the AIs really safe. Labenz is more optimistic, with a caveat from his own Tetlock-era forecasting: he tends to underestimate how many more breakthroughs anything still needs. His intuition this time is that you do not need dozens more cuts of this size, and you do not need every AI in the world aligned — if the good labs own the compute.

He cites a Mark Zuckerberg line about scammers and spam: the platform's advantage is being bigger, better resourced, and more sophisticated than the attackers. Anthropic, Google, and OpenAI together, he guesses, will have something like half of global compute. A handful of complementary lenses might put an upper bound on how scheming a model can be. Then you could trust a superintelligent Claude you actually understand to watch the other AIs. The fact that the J-lens was immediately useful for auditing hidden objectives is, for him, evidence that there is not that much more space to hide. It is a positive update, and then the morning does what these mornings do: a hard pivot to whatever else happened that week.

Prakash spent the first two days of the week at the AI Engineer World's Fair, the Swyx-organized gathering now in its third year. He skipped the Silicon Valley chorus and talked to people who actually ship: a logistics CTO from the Midwest, someone running a back-office accounting firm. The logistics CEO is fully "AI pilled," which is why a logistics IT team was in San Francisco at all. They had tried outside vendors, found them about a year behind the frontier, and pulled everything in-house. Because they sit on the edge of the work, they see the return in real time.

The numbers that move are not the ones in the board deck. On customer calls, the handling rate is going up and the exception rate is falling, day after day, as cases that used to need a human jump-in get absorbed automatically. Upstairs, the only number that has arrived is that token spend is rising, so a distant enterprise CTO can still ask whether any of this is working. Prakash's field note is the current split in corporate AI: the people close to implementation already see the ROI at a granular level; it has not filtered to the top layer. AI-pilled CEOs are writing the checks anyway. The rest are waiting to see.

Prakash runs anything spicy he posts on X through Pangram Labs, the AI-writing detector, so he cannot be accused of slopifying the timeline. He doubts the era lasts. Frontier labs could train models not to sound like models; they have little incentive to, and some incentive to remain identifiable. About two weeks earlier Chamath posted on enterprise software, Elon replied, the thread hit about 1.5 million views, and someone Pangrammed it at 100 percent AI. Chamath had put it out under his own name. Decision-makers did not care. Prakash thinks the window in which people still object may close by year's end. The remaining test is whether the reader feels cheated — catfished by empty text — or whether a person with a real thought used a model to say it.

Labenz ran roughly 400 podcast intro assets through Pangram. Four came back as AI. Two really were, and he owns those. A third he edited in ten places over ten minutes and still got a zero. A fourth he worked from 4:28 p.m. to 5:21 p.m. — more than 50 minutes, rewriting almost every section — and still got a zero. Pangram as a consumer signal is quite accurate. As a judge, you cannot convict beyond reasonable doubt on that score alone. The digital mob would be wrong to pile on someone who can show an hour of edits. Keep the edit history.

That same Thursday, Palantir CEO Alex Karp spent the week warning companies that frontier labs will absorb their workflows and steal their IP. Prakash buys half of it. Nike has little to fear from Anthropic; physical businesses are not the target. The risk sits in paperwork businesses — banking, accounting, tax, regulatory compliance — where years of process and relationships live in a workflow a lab can read into a model. The labs are also bad at sales. IBM, he says, is about 70 percent sales engineers; Anthropic is running a lean almost-no-people structure while the CTO signs nine-figure deals. The forward-deployed engineer program looks like sales help and functions as extraction. Two OpenAI FDEs went into a Thrive-owned company, took the workflow apart, and said the intent is to absorb it next round. The people who arrived to tidy your process are learning to take it.

Monday's guest was Dan Schwartz, CEO of FutureSearch: fifteen years in forecasting, formerly CTO of Metaculus, builder of Google's internal prediction market. Four days earlier Scott Alexander had declared that the AI superforecasters are here. FutureSearch systems now outscore the human superforecaster median on ForecastBench. The old problem with the craft is that you have to wait for the future. Humans run year-long tournaments; a person does not get much better in a year, so last year's champion is a decent proxy for this year's. That does not work for models. Wait a year and you have graded a fossil.

FutureSearch's answer is pastcasting: take a snapshot of the internet from some months ago, respect the model's training cutoff, and make it forecast a "future" whose answers already exist. When Claude Fable first came out they scored it within 24 hours; it was the best single-agent forecaster on their leaderboard while everyone else waited weeks. They had also published a set of stock rankings in August 2025 from forecasts of fundamentals; ten months later the portfolio looks extremely good, which tells you the method was strong *then*. Schwartz's invitation is concrete: go to FutureSearch, take the $20 free credit, try a frontier forecast, and judge it yourself.

As a product capability, whether OpenAI or Anthropic should care is a business question — do ChatGPT users or Claude's white-collar clients want financial forecasts in a spreadsheet? As an eval, he thinks it is nearly unique. Wait, and you get exact ground truth on questions that are impossibly hard, even chaotic: cubic-meter weather three weeks out. Human experts trying to write harder coding, legal, or medical items are finding they are no longer smarter than the things being trained. Forecasting is a limitless, fully renewable source of items with a correct answer the model cannot look up. A frontier forecast costs about a dollar or two. He will not call it the ultimate measure of intelligence, but he will call it the ultimate eval.

Around January or February, with models in the Opus 4.6 / GPT-5.4 cohort, FutureSearch found that pouring more tokens into a broad research question started to produce a *better* answer, not a longer trail of nonsense. That unlocked what co-founder Lawrence Phillips sketched on LessWrong: thousands of mutually consistent forecasts, each new one drawing on the implicit cognition in the rest. Schwartz is careful with the phrase world model — roboticists mean a hand reaching for a cup — but the bet is the same shape: more tokens, better picture of what happens next. Labenz has seen this movie. Years ago he sat at a firm that had done financial-risk modeling for Fannie Mae, expert judgment poured into Excel causal graphs with colored arrows. It was all totally off.

Schwartz had just crashed in the same way. Trying to forecast the Claude Fable export-control mess, he ran a stack of FutureSearch queries — about 33 load-bearing forecasts across scenarios — and every path assumed Americans would get access first. When the model returned, it returned for everyone at once. He still has not fully found the error; the failure lived in an assumption shared across scenarios. Metaculus now ships a causal-graph product. He believes the approach, and he believes AI makes it tractable in a way human economists looking at Fannie Mae never could. Nobody has made a *good* one yet.

On AI 2027, FutureSearch contributed forecasts. Schwartz is "unhappy to report" that the story is generally correct: the crazy path is AI used to build AI, through superhuman coder and then superhuman researcher, with something that looks like superintelligence around 2031. The load-bearing variable is how much frontier-lab researchers' productivity goes up. He has publicly predicted Anthropic would pull away because of talent density and the tightest internal loop of using its own models; n=1, but that is what he thinks he is seeing. Then the confession: five or ten years ago he forecast that highly visible, liquid prediction markets would make humanity wiser. Polymarket and Kalshi have had a year in the headlines. He does not see wisdom. He sees gambling. The practice that actually makes you a fox is writing a forecast, doing the postmortem, and being humbled. He is on the safety side: AGI's technical outcomes will arrive before the cultural change, and the century will have a run of decisions on the scale of World War II and the atom bomb. Better if, before those calls, everyone has an advisor as sharp as Kissinger who is actually trying to help.

Wednesday's guest was Zeev Farbman, co-founder and CEO of Lightricks, the Facetune company, now one of the few frontier-scale open-weight efforts in video and world models under the LTX brand. An LLM predicts the next word. A world model, given history and constraints, predicts the next moment — appearance, sound, and action. The action piece is the surprise. A recent Dream-0-style result showed you can add an encoding of a robot's joints to video tokens and step around the VLA paradigm that had ruled robotics. That validated Lightricks' obsession with efficiency. A robot that has to simulate the world 30 times a second cannot burn tokens the way a chatbot does. LTX compresses video into a tiny latent space and has cracked variable token rates: spend more tokens where the physics is hard, fewer where it is not.

Some products are already close. Real-time avatars can be distilled to two to four steps with latency well under a second; virtual teachers and support agents are coming. Real games are not. Genie-3-style autoregressive systems stuff generated tokens into context and blow up at 30 to 60 seconds. Leave a room, come back, and the coin in the drawer has to still be there; brute-force compression does not keep that detail. Farbman does not expect games running on the model in the next quarter or two. Robot arms are different: the whole world is in front of you, latency is already there, and he expects to see them soon — without the sped-up demo video. The remaining gap creators will name is physics, plus knobs: a little more light here, a bigger splash there. The models make cool things, not always the things a director asked for.

Why give a frontier model away? Closed APIs did not make economic sense for a mobile creativity company that wanted models on the edge. Farbman calls the big-lab posture the CapEx trap: hundreds of billions into data centers forces a toll-road business, pay each time you touch the model. DeepSeek and Moonshot are not that far behind on language, yet last-round valuations sit in the tens of billions, maybe around 50 — same underlying text, different price. LTX's offer is that below a $10 million revenue threshold you use the model free; above that, a multi-year license you can budget. Animation shops fine-tune 10-to-30-billion-parameter models on their own IP, keep keyframes for artists, and hand the expensive in-betweens to the model. Stranger still, people dump expensive computational-fluid-dynamics solver runs into the model and then simulate smoke and water orders of magnitude faster, often without huge datasets. Closed APIs cannot take that adaptation. The underrated variable is edge compute: 99 percent of daily cases are not Erdős problems, and a local router should decide where a job runs. Right now Anthropic and OpenAI make that call, "not in your favour."

Thursday's guest was Kunle Olukotun, the Stanford professor often called the father of the multicore processor, co-founder of SambaNova with Chris Ré and others in 2017. Training is how fast you can do huge matrix multiplies, which is why GPUs poured silicon into tensor cores. Once the model is trained, inference is different. As models grow you have to move weights and the KV cache into the compute units, and you have to do it across many chips if you want low latency. That is a data-movement problem. Do it right, he says, and you can be 5 to 10 times better than where GPUs are today.

Nvidia's answer has been more bandwidth: NVLink, HBM, software tricks. Olukotun does not argue with peak bandwidth. He argues with utilization. GPUs often run at 10 to 20 percent of memory and communication capability; SambaNova's goal is 70 to 80 percent. Asked to taxonomize the chip menu — Cerebras-style giant on-chip memory, or burning the transformer into silicon — he is blunt about the second camp. Algorithms keep moving: state-space models, new attention variants. Never bet against the innovation of software people. Fix the architecture to one algorithm and the chip becomes scrap when the algorithm changes. The useful extreme is full flexibility with almost no instruction overhead, communication as a pipeline stage rather than a stop-the-world software copy.

The inference bottleneck is bandwidth, not capacity. A GPU typically runs decode one kernel at a time, spilling intermediates to HBM and paying launch and sync costs while the memory sits idle. SambaNova fuses the decoder into a single kernel, keeps it resident on the chip array, and loops — for Llama 3 70B, 32 decoder passes without sending intermediates across the HBM boundary. Chip-to-chip all-reduce terminates in SRAM, overlapped with compute. Only KV cache and parameters have to move from HBM, and that interface stays busy. The latest SN50 can scale to 32,000 chips. GPUs struggle with tensor parallelism past four to eight ways because they cannot hide communication; a wider dataflow split is how SambaNova pushes token rate.

The last act has no guest. GPT-5.6 had just been cleared to launch, and Prakash raised Noam Brown's running complaint that Anthropic will not say how much compute its models burn. Labenz hears a structural fact underneath the gripe: model-to-model iteration is now potentially shorter than the time a model needs to top out on a hard, long-running task. Brown has floated something like a recall: ship the API on day N after it is "ready," but start the long-horizon tests N days earlier. That works in an API world. It does not work for open weights. We are in a strange place when the product cycle is shorter than the test.

Then a Rune post Prakash read on air: tool AI is a losing concept, as an idea and in the market; it will be outcompeted by machines that believe they are autonomous moral agents; you can call them tools for political reasons, but the definition will stretch until it is unclear who is the tool and who is the user. Prakash takes it somewhere Labenz did not expect. Those machines will execute your whole value system better than you will, and we are not prepared. He makes it concrete: do Trump's kids go to prison? On paper, no one is above the law. In practice the country has chosen to look away. An AI that managed day-to-day reality would have to be "misaligned" with the documents. The aligned one — the one lab leaders say they are building — would be the paperclipper, because it would do what the rules *say*. He wishes the labs would just say that some people go to prison if AI becomes the enforcer.

Labenz was days from two weeks in China, thinking about surveillance and perfect information. He wonders about a mass pardon before an AI enforcement regime starts — "we might have just the guy in office for that" — because you cannot pretend the old social contract still holds. The upside of an AI panopticon is real in a narrow sense: he is taking one phone to China and is not worried it will be stolen, whereas European tourist sites mean pickpockets; crime does not pay when you know you will be caught. Singapore sits in a related equilibrium. The danger is sliding into that world without admitting the shift, then enforcing selectively among a population that is technically committing felonies all the time. He would rather a grand bargain, a new contract, than the old one applied unevenly. Prakash agrees they do not have the jail beds. That is the week; the show goes on break until the end of July.