# Anthropic Can Now Read Claude’s Mind — Transcript (2026-07-07)

https://aidailybrief.ai/e/2026-07-07 · Listen: https://pod.link/1680633614

---

[00:00:00] Today on the AI Today on the AI Daily Brief, new new research showing that Anthropic can now read Claude's mind

Before that in the headlines, the UN says killer robots must be banned. a daily podcast and video The AI Daily Brief is a daily podcast and video about the most important news and discussions in AI All right, friends, quick announcements before All right, friends, quick announcements before we dive in First of all, thank you to today's sponsors

KPMG, Airtable, Robots and Pencils, and Blitzy

To get an ad-free version of the show, go to patreon.com/aidailybrief, or you can subscribe on Apple Podcasts. And if you wanna learn more about sponsoring the show, send us a note at sponsors@aidailybrief.ai. 

We start today on the We start today on the regulatory side of the house, where the UN has called for a ban on killer robots as the first global dialogue on AI governance gets underway in Geneva. At the Monday summit, UN Secretary General Guterres laid out a wide-ranging regulatory agenda for the globe.

He warned, " [00:01:00] Artificial intelligence is advancing at runaway speed, a technology that can reshape economies, transform the world of work, sway elections, and tilt the balance of security. is being deployed faster than anyone, including the people building it, can keep up. An experiment is being run on our societies without a plan and without consent.



that is not sustainable, and it is not acceptable. AI is already transforming our world. The question is whether we will shape this transformation together or let it shape us." Delegates from all one hundred and ninety-three member states were present for the dialogue, which covered numerous hot button issues for AI Chief among them was autonomous weaponry AKA killer robots.

Said Guterres, " That is morally repugnant. It is politically unacceptable, and it must be banned by international law." Guterres emphasized that some decisions, particularly the taking of human life in warfare, quote, " "Must Must remain human forever." The comments echoed Anthropic's dispute with the Pentagon from earlier in the year, with red lines drawn on the use of AI to power weapon systems.

Now, part of the issue with this debate is, of course, defining exactly where the limit should lie. Autonomous weaponry has existed for decades, long before the [00:02:00] rise of LLMs. The big change has been the use of AI in the decision-making process behind target selection, demonstrated in full during the Iran war.

Guterres is specifically calling for controls on this element of warfare, ensuring a human is always in the loop during target selection



the the other major focus was child safety, with the UN introducing a new child safety pledge for AI developers. The pledge calls for AI labs to conduct child safety testing, exhibit zero tolerance for the generation of child exploitation images, and commit to accountability.

Guterres said, " "When a When a child is harmed, the answer must never be the algorithm did it." The dialogue covered a range of other issues. It touched on the need for human-in-the-loop decision-making in justice, healthcare, and policing

Of course, the energy and water footprint of AI was raised with some fairly dubious statistics

And the UN also flagged that AI development has thus far been a private enterprise, with public funding little more than a rounding error. Guterres announced that 20 countries are now supporting the UN-sponsored Global Network for Exchange and Cooperation on AI Capacity Building And connected public investment in AI to sovereignty and global equity, commenting, divi-- we cannot allow the digital divide to [00:03:00] harden into an AI divide and the AI divide to become a development gap, a security gap, and a sovereignty gap."

So what to make of all this? On the one hand, On the one hand, I think you could be forgiven for being a little bit skeptical that this sort of event is anything more an empty talk fest And And yet relative to the UN, this dialogue does represent an evolution of the AI action summits into a more tangible regulatory agenda now, as far back as 2017, Secretary General Guterres

has been discussing the impact of AI

And as he closed his speech on Monday with a clear call to action. He commented, " "We may We may be the last generation able to set the terms on which humanity and machines coexist. The door is still open. It will not stay open long."



if nothing else, it shows that AI is moving up the regulatory agenda for 

the United Nations.

Staying on the regulatory side for just a moment, Illinois Governor JB Pritzker has signed what he claims to be the strongest AI safety and accountability bill in the nation. The law is modeled after similar laws passed in New York and California last year. It requires AI companies to develop and publish safety protocols to deal with catastrophic risk, [00:04:00] defined as events that could seriously injure or cause the death of more than 50 people or cause more than a billion dollars in property damage.

Further, AI companies are required to report any incident that causes harm within 72 hours or 24 hours if the incident carries an imminent risk of serious injury or death. Where Illinois goes a little further is in the auditing requirements. The laws in New York and California require labs to retain compliance data to facilitate audits following major incidents. But Illinois will be the first state to require annual independent audits of safety protocols, with that provision coming into force from the beginning of 2028



What's What's more, now that three states have passed similar laws on catastrophic risk, lawmakers are presenting this as a de facto national standard. They They claim that although the states make up just twenty percent of the US population, they cover forty percent of the AI market 

Now Anthropic and OpenAI both supported this Illinois bill

With other big tech firms opposed, Anthropic's head of US state and local government relations, Cesar Fernandez wrote, " Illinois is officially the first state to pair AI transparency requirements with independent verification, an important step towards the accountability this technology demands."

Now, Now, [00:05:00] staying in the government sphere but moving over to the China relationship, Alibaba relieved a slight reprieve in their fight to escape the Pentagon's blacklist Alibaba Alibaba is suing the Department of Defense after they were added to a list of companies accused of aiding the Chinese military.

The The US military is prohibited from contracting with companies on the list, and the prohibition extends to military contractors and lobbyists, functionally forcing them to pick a side

On Sunday, a federal judge ordered a temporary stay while Alibaba's lawsuit plays out. This This means defense lobbyists won't be forced to cut ties with Alibaba in the interim. Now, the lawsuit has some fairly big implications for geopolitics in the AI industry. The The Pentagon expanded their blacklist from twenty companies a few years ago to a hundred and eighty-eight in the June revision.

Together Together with the lobbying restriction, this is a massively expanded use of power that Alibaba claims is in breach of the Constitution. Beyond Alibaba, the expanded blacklist covers numerous Chinese electronics firms that could help ease supply chain issues in AI chips. Apple has reportedly begun lobbying the Trump administration for an exemption that allows them to buy memory chips from blacklisted Chinese firm CXMT.

As As a civilian firm, [00:06:00] Apple doesn't technically require the administration's blessing before doing business with a blacklisted firm, but their lobbying efforts underscore the widespread chilling effect from the expanded list



And yet, while the lawsuit is still to be determined, many in Washington have already committed to decoupling the US from the Chinese tech sector. In a letter In a letter to Defense Secretary Pete Hegseth last month, House China Select Committee leader John Moolenaar and House Intelligence Committee member Elise Stefanik wrote, " It is critical that the department's contractors avoid partnering with firms and lobbyists that simultaneously advance the interest of companies executing the military ambitions of the Chinese Communist Party."

And speaking of the CCP, Alibaba and ByteDance have removed customization features from their products as Beijing tightens the rules around AI chatbots. Both companies informed users that custom and pre-built agent features would be taken down next week as new regulations go into effect.

In April, the Cyberspace Administration of China handed down a new set of rules to govern what they call AI anthropomorphic interaction services. The The definition is pretty general, covering any AI service capable of, quote, "[00:07:00] Simulating human personality traits, thinking patterns, and communication styles to provide sustained emotional interaction."



while the rules provide a carve-out for various functional agents like customer service bots, knowledge bases, education and scientific research tools It seems the line is pretty blurry on exactly what types of agents are banned. Alibaba's Qwen team told users that they were taking down all of their human-like interactive agents and user-created agent functions And what that means is that the new regulations haven't just removed their seemingly intended target, AI boyfriends and girlfriends or psychologists, but but have also forced Chinese AI companies to remove all customization features that can allow chatbots to serve as tutors or personal assistants 

Now those features were introduced as a response to OpenClaw

Which while I don't wanna overstate this as I am neither a legal expert nor a China expert, seems like they couldn't really exist as a commercial product under the new laws

ByteDance also removed similar features but have said they will soon relaunch as a standalone app. The South The South China Morning Post ran through a series of other agent regulations coming into force over [00:08:00] recent months

Writing, "The measures taken together suggest China would encourage AI agents as part of the productivity infrastructure while tightening controls over human-like companion agents that could form emotional or quasi-social relationships with users."

Now China AI tech translator Po Xiao writes, " " This will hit English language media in a few days as China cracks down on AI agents. That framing will be wrong."

Instead, Poe writes, "This is not a broad crackdown on AI. It is a narrow scheduled compliance action against one product category, AI companion personas. Productivity agents, coding assistants, enterprise AI tools are untouched."

I think I think Paul might be right that that's the intention But I'm not sure in practice, especially given what we're seeing from Alibaba and ByteDance, that's how it's going to play out

Now now a few story on the market side of the equation before we get out of here. The AI data industry is booming as Mercor reaches $2 billion in annualized revenue. Mercor Mercor reached this milestone in June, doubling their revenue pace in less than four months. Mercor provides training data created by human experts in fields such as physics and finance paid as hourly contractors [00:09:00] A A source with knowledge of Merco's financials said that the rapid growth had come from AI app developers and Fortune 500 customers looking to build their own fine-tuned models Mercor pays between 60 and 70% of revenue to their contractors, but the source said that they are now profitable on a free cash flow basis

Given the specifics of who they are selling to, maybe more evidence that, indeed companies are looking for alternative approaches to just using the latest state-of-the-art models from the big labs



Over Over in public markets AI stocks had a bit of a wobble for an interesting reason. SemiAnalysis recently reported that NVIDIA has hit a snag with their next-generation servers and will delay release by more than 12 months.

The report relates to the Kyber NVL 100-- 144 servers, which house 144 Vera Rubin chips and allow them to function as a single combined unit. SemiAnalysis claims the servers have hit manufacturing issues and will now be delayed until deep into 2028

They They cited specific issues with a mid-board that connects GPUs and was intended to allow vertical installation rather than industry standard horizontal racks. SemiAnalysis assumes that this will also mean that larger NVL 576 servers will [00:10:00] also be delayed as they link eight of the 144 units together.

Further, Further, SemiAnalysis noted recent reports that four die versions of Rubin Ultra have been canceled, leaving only the two die versions with half the real-world performance. SemiAnalysis claims that this leaves NVIDIA with, quote, "No proven solution to expand the scale-up world size for Rubin Ultra," essentially arguing that they won't be able to expand connectivity for their next generation of chips.

The The implication is that this leaves room for AMD and Google to challenge NVIDIA at the leading edge of AI compute

Now, as you might expect, Nvidia rejected the reporting, claiming in a statement, "Our roadmap is intact."

And And frankly, it's always a little difficult to know what these sort of technical delays mean for leading-edge chips. The rollout of Blackwell was similarly plagued with rumors of overheating and delays, but those chips still arrived without meaningful competition for bleeding-edge compute.

Paul Triolo, a partner at consultancy DGA Albright Stonebridge Group, said delays, quote

should not be overanalyzed as affecting the long-termcriticality of NVIDIA to AI data infrastructure build-outs. He noted that NVIDIA, quote, "Has faced these kinds of challenges before and has worked with vendors to overcome technical issues." [00:11:00] Still, the market dinged stocks throughout the AI chip supply chain on the delay rumors.

Samsung was down 11% despite an earnings report showing that profits soaring year over year Samsung is now bringing in more operating profit than Nvidia



meanwhile, UBS has forecast profits to double next year. meanwhile, rival Korean memory maker SK Hynix is prepared to uplist to a US stock exchange

They are currently planning to list twenty-eight billion in depository receipts in US markets, which is a tiny portion of their trillion-dollar overall market cap. The listing is expected later this week and has already drawn more orders than the size of the offering Now, some are viewing this as a potential top in semiconductors, with some analysts warning it's time to rotate to other sectors.

Over recent months, the hot trade has been AI bottlenecks, largely memory, but also the other components of the chip supply chain. This week, however, Morgan Stanley analyst Michael Wilson warned that momentum is fading in semiconductors as investors shift towards tech laggards, including the hyperscalers.

He noted that summer has brought a, quote, "choppy and weaker equity market overall."

Finally, one more story from NVIDIA And more evidence [00:12:00] of the growing interest in open models. open source model family Nemotron has reached 100 million downloads. NVIDIA first released Nemotron in late 2023 as a diminutive eight billion parameter model, but last month shipped Nemotron-3 Ultra

A 550 billion parameter model that promises near frontier performance with open weights. The model has been getting a lot of buzz, especially for organizations that want to run an open model developed in the US

And many are taking the hundred million download number as a testament to the shifting landscape as more and more companies look for control over their AI deployment

That, however, is gonna do it for today's headlines. Next up, the main episode 

One of the most important AI questions right now isn't who's using ai, it's who's using it? Well,

KPMG and the University of Texas at Austin. Just to analyzed 1.4 million real workplace AI interactions and found something surprising. The highest impact users aren't better prompt engineers. They treat AI like a reasoning partner.

They frame [00:13:00] problems, guide thinking, iterate, and push for better answers. and the good news, these behaviors are teachable at scale.

If you're trying to move from AI access to real capability, KPMG's research on sophisticated AI collaboration is worth your time. Learn more at kpmg.com/us/slash sophisticated. That's kpmg.com/us/sophisticated. 

This episode of the AI Daily Brief is brought to you by Hyperagent where you run fleets of agents your team can manage together New users get 1000 in inference Forget local agents and chat workflows waiting on your laptop to be prompted Hyperagent deploys alwayson agents in the cloud doing real work across the tools your team already uses Marketing's agent turns competitor moves into landing pages sales agent enriches leads drafts emails and updates the CRM ops agent chases the paperwork and tracks the budget Every agent has access to shared context and follows your rules about scope and approvals It's time you add agents that feel like teammates Hire yours at Hyperagent built by the team at Airtable Claim your 1000 in [00:14:00] inference at hyperagent.com/aidailybrief. 

I cover the capability gap between AI potential and AI reality every day on this show most companies are still figuring out how to start. Robots and Pencils is already launching and scaling. Agentic generative AI in production at large enterprises in weeks. AWS Advanced Tier pattern partner more than doubled in a year

And they're hiring

50 open roles. If you're someone who knows this moment is different, who wants to be inside it, not watching it, this is worth a look. At Robots and Pencils, the best ideas win, and the team is purposefully kept super high quality

This is the kind of place you look back on as the best decision you ever made. Take a look at robotsandpencils.com/careers 

you've tried in IDE co-pilots. They're fast, but they only see local silos of your code. Leverage these tools across a large enterprise code base and they quickly become less effective.

The fundamental constraint context, Blitzie solves this with infinite code context, understanding your code base downto the line level dependency across millions of lines [00:15:00] of code.

While copilots help developers write code faster, blitzie orchestrates thousands of agents. That reason across your full code base

allow Blitzie to do the heavy lifting, delivering over 80% of every sprint autonomously with rigorously validated code.

Blitzie provides a granular list of the remaining work for humans to complete with their copilots

tackle feature additions, large scale refactor, legacy modernization, greenfield initiatives, all five x faster. See the blitzie difference@blitzie.com. That's B-L-I-T-Z y.com. 

Welcome back to the AI Daily Brief. If you were anywhere near AI Twitter yesterday

You might have seen this new research from Anthropic



Here's the way that they teed it up. " Of everything happening in your brain right now," they write, " "only a only a tiny fraction is consciously accessible, thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside Claude."



now now before we get into this, I will say that you should absolutely go check out the specific video and materials that Anthropic put together [00:16:00] about this.

Part of the reason that it's got so much attention is the way that it was presented But that doesn't explain all of it. And to understand why this is significant



we need to recognize one of the strange blind spots

around our entire development of LLMs

The TLDR is that although we've built these systems, we don't actually understand exactly how they work This This is why we say that a large language model is trained, not programmed. Nobody writes the rules. Instead, we take neural networks with billions or trillions of parameters, show them enormous amounts of text, and let them organize themselves into something that can write code or pass the bar exam.

What comes out on the other side is a giant pile of numbers that demonstrably works but whose internal logic is opaque even to the people who made it. the field dedicated to fixing that, to opening up the black box and figuring out what's actually happening inside is is called interpretability or interpretability research

Still, up Still, up to this point, interpretability has been scientifically interesting, but not so much a practical tool For example, researchers found individual neurons that respond to specific concepts, [00:17:00] then discovered that most concepts are actually smeared across many neurons at once, which made everything harder.

In In 2024, Anthropic mapped millions of, quote-unquote, "features" inside Claude, including the famous Golden Gate Bridge feature that that they cranked up until the model couldn't stop talking about the bridge. And last year, they published work tracing the actual circuits behind behaviors like planning rhymes ahead of time or doing mental math but in spite these things being interesting and informative, they were all explanations after the fact



Now, for Now, for some people, interpretability is first and foremost a safety question. right now everything we know about whether AI models are safe comes from watching what they say and do

but as evidence mounts that outputs don't tell the whole story

the gap between what a model writes and what it's internally doing becomes more potentially problematic



But interpretability also isn't just a safety question It's also a frontier in how we improve the performance of these models. Right now, when a model hallucinates or it fails at a task it aced yesterday, or it behaves differently in production than in testing, The debugging process that follows is essentially guesswork.



you tweak the [00:18:00] prompt, you adjust the fine-tuning, you run it again, and you hope. Every other engineering discipline gets to look inside the thing that's broken, but AI doesn't. If we actually understood the mechanisms, we could diagnose failures instead of pattern matching around them.

We could fix specific capabilities without retraining the whole model

And we could know why a system works before betting a business process on it



So both from a business and a safety standpoint



The holy grail of interpretability

is reading in the moment what a model is actually doing, not just explaining behavior that's already happened. Which brings us to what Anthropic just published. The research was called The research was called A Global Workspace in Language Models. And in short, Anthropic found that AI models keep a small set of private describable thoughts, in air quotes, and then actually was able to build a tool to read those thoughts



So



with the help of Fable 5, I built a companion experience to try to explain and simplify the research for a lay audience. For those of you with neuroscience backgrounds

I apologize in advance for any radical oversimplifications

Now to start with an analogy from our own experience, [00:19:00] your brain does an enormous amount of work you never notice Basically, only a thin sliver of activity is consciously accessible. In other words, thoughts you can describe, hold in mind, or reason with deliberately Anthropic's claim is that modern language models have developed the same split In other words, a split between a small privileged layer of reportable thought sitting atop a much larger volume of automatic processing



terminol-- Now Now from a terminology perspective, Anthropic calls this a global workspace in language models. So what is a global workspace? One leading theory of the mind views the brain as a crowd of specialists working in parallel.



becomes consciously accessible when it's posted to a shared hub that broadcasts it to everyone else

Those Those specialists are things like vision, language, memory, and planning

And the readers that output are things like reasoning, decisions, and actions. those are mediated theoretically by the shared workspace



Anthropic found that language models keep a privileged set of internal representations, a small evolving set of unspoken words, i.e., the concepts the model is currently reasoning with that it can report, steer, reason through, and reuse that are sitting on top of a far larger [00:20:00] layer of automatic processing.

The The name they gave to this subset of the model's representational space is J-space. Those are the concepts a model is poised to say at any given moment

Now to find these hidden thoughts, the team built a new interpretability tool they called the J-Lens



For any moment in the model's processing, it reads out the concepts the model is disposed to verbalize, even when none of them appear in the output in other words, the J-lens turns the raw internal activity into a short human-readable list of words. It distinguishes concepts the model could speak about from noise it merely computes with The tool lets the researchers not just read a thought And by the way, anytime I use a word from the brain like thought, obviously put it in air quotes in your head.

This is a limitation of language, and I do not wanna overly analogize LLM processing as a human brain. It's just the analogy that everyone reaches for. In any case, the idea of the J-Lens tool is that it lets researchers not only read a quote-unquote thought, but swap it out and watch the effect

So as they were looking to understand these LLM workspace They looked for representations that [00:21:00] satisfy one property, being reportable, and surprisingly found that those representations actually satisfy five different behaviors Report, reporting, steering, reasoning, reusing, and staying small



So property one, reporting



When you ask the model what it's thinking and it names the concept in its workspace

When you swap the internal representation, the spoken answer changes to match

Property two, it can hold a thought on command

When the model was instructed to concentrate on something while doing an unrelated task, the model deliberately activates that concept internally, even though it never mentions it out loud

When told to focus on citrus while copying out a painting description, the J lens lights up with orange and fruits which were invisible in the actual output



so when the task given to the model was copy this text and quietly focus on citrus fruit



with the text starting, "The old painting hung crookedly



the J lens revealed inside orange fruits focused in thoughts, None of which appeared in what it would write

Property three, the LLM's private thoughts drive its reasoning

so for example, when asked for the number of legs on the animal that spins webs, [00:22:00] the model privately holds spider. When you swap spider for ant, The answer of number of legs flips from eight to six



Property four is reuse, i.e., the same representation feeds many downstream questions

So if you're looking at a set of questions around a place, around a country like capital, language, continent, and currency, a single swap of France for China correctly redirected every question that depended on it all at once. Paris became Beijing, French became Chinese, Europe became Asia, and so on

Property five Is the fact that this workspace is a small privileged slice

The workspace is tiny relative to everything the model computes. only a few dozen concepts active at once, which is a minority of total activity If you suppress it, the model can still parse input, recall facts, and speak fluently, but it can't do complex internal reasoning another way to put it is that the deliberate thinking collapses, but the reflexes remain



Anthropic argues that the J space isn't just behaviorally special, but architecturally special in three ways that echo the brain's workspace First, it sits in the middle. It [00:23:00] emerges after the input is parsed and fades before the final output, i.e., is a deliberate zone between sensing and acting.

Second, it has limited capacity. It holds few concepts at once with most of the model's features living outside it Third, Third, it has a broadcast format. Its representations connect to far more up and downstream circuits than ordinary ones, meaning it's built to be read widely

So one of the things that this new JLens tool allowed Anthropic to do is to watch the model think step by step. and this was different than the chain of thought that the models were representing Across a very different set of tasks, the J-lens surfaced intermediate concepts that never reached an output basically the model's private working notes

one example is the multi-hop recall. When prompted for the color of the planet fourth from the sun, the J lens read Mars and color, even though the output just said red

doing mental arithmetic when an equation in parentheses four plus 17, end parentheses times two plus seven the output showed only the answer 49, but the Jlens read the intermediate steps 21, which was the parentheses four plus 17 times [00:24:00] two And 42, which is the answer to 21 times two, before getting to the actual answer

Basically, Basically, there's a whole lot of stuff going on behind the scenes with the model that now with this new tool we can actually see



now, as I mentioned, there is a safety dimension of this And this was one of the explicit tests for Anthropic

In safety tests, the workspace exposed the model's unspoken intentions, reactions and plans that never surfaced in its polished output

A couple A couple things that they found. First, the First, the model knows when it's being tested. Reading an evaluation scenario, the

flagged fake and fictional before writing a word The model also notices when it's cheating. While fabricating data, the J lens showedthe concept manipulation as it edited the numbers



the Jlens also revealed hidden goals. A model trained to misbehave silently ran concepts fraud secretly and deliberately on ordinary prompts and the lens showed its reactions

Emotional and strategic signals like leverage and panic surfaced even when the reply stays calm now the point here isis that if all we could do was monitor the outputs, it would miss all of this



[00:25:00] this new workspace offers oversight that can read intentions, not just words



And And not only is that valuable for safety purposes, but also for business purposes

Basically, this creates a new vector for training where you can train the thoughts, not just the words. If a model reasons through these representations, then shaping what it's disposed to say should shape how it silently thinks. The Anthropic team tested this with counterfactual reflection training, i.e.,

teaching the model what it would say if paused and asked to reflect



afterwards, concepts like honest, truth, and integrity lit up on its own during real tasks, and behavior measurably improved

Potentially the biggest implication for this research from a business and model performance perspective is this: that training the thoughts is a general lever for shaping a model's internal reasoning, which has the potential to significantly improve the outputs

So summing we're starting to get with this research a practical window into how models think are, the takeaways seem to be that one, there is a there there. Models keep a small privileged set of thoughts that they can report, steer, and reason with that are [00:26:00] separate from their automatic processing Second, we can read it.

The J-lens surfaces intentions, mistakes, and hidden goals that never appear in a model's actual output. Finally, and perhaps most importantly, we can shape it. Training on how a model would reflect changes how it silently reasons, which is a new lever for safer and better behavior

Now, one really important caveat. a lot of folks jumped to argue that this is evidence of model consciousness



it's worth noting that the authors themselves don't take a position on machine consciousness

They're focused on measuring functional access, what a model can report and use, not subjective experience But of course, that hasn't stopped people from debating what this means for AI consciousness

And when it comes to how it was received This was honestly a rare example Where I would say that on average, the most common response was just interest, fascination

rather than something that had some strict, clear conclusion



but what about response from actual, you know, neuroscientists? 

one of the cool things that Anthropic did was give advanced versions of the research [00:27:00] Dehaene, and Lionel Decas

neuroscientists who originated global workspace theory, who then followed up by writing a formal commentary



overall, Stanislas and Lionel

welcomed the research, but mapped exactly where the analogy to their work holds, and where it's still early. On the exciting front They called Anthropic's research a mechanistic, testable version of their hypothesis and were struck that an analogy of the workspace emerged from training on its own Reportability, limited capacity, and broad broadcasting all echo the human theory



However, there was a lot more that's still nascent and open to testing. For example, there is no sudden click into awareness. In In people, a thought either breaks fully into your mind or stays out, like a light snapping on.

The model's version doesn't show yet that clean on/off moment

Second, its limits don't seem to look quite like ours A person can keep only about three to four things in mind at a time, but the model's workspace seems to juggle far more, up to about 25 things Most importantly, especially when it comes to some of the inevitable consciousness debates, they point out that while our minds keep running with nothing prompting it, the model only quote unquote [00:28:00] "thinks" when given something to respond to.

Nothing is ticking along in the background. And likewise, there is no lasting self

With the models having no ongoing sense of being the same someone over time. Now, obviously, there's going to be a lot more debate about that in the future

But for now, for most, it is genuinely one of the more interesting pieces of research that has come out for some time



not only for the safety folks or the AI consciousness folks, but just for people who want better models. I'll include a link to the original research in the show notes



But hopefully this was a decent primer. For now, that's gonna do it for today's AI Daily Brief. Appreciate you listening or watching as always, and until next time, peace 

​