Out of everything happening in your brain right now, only a tiny fraction is consciously accessible — thoughts you can describe, hold in mind, and reason with. We found a strikingly similar divide inside our AI model, Claude.
Our experiments were inspired by a leading theory in neuroscience: the global workspace theory. It holds that a thought becomes consciously accessible when it enters a shared “workspace” that’s broadcast across the brain. We found a set of representations in Claude’s neural activity that play a similar role.
Read more about the research here:
0:00 The mind as an ocean
0:47 Meet the J-space
1:52 Silent math
2:22 Don’t think about the bridge
3:12 Claude’s automatic processing
3:42 Catching hidden thoughts
4:34 The question of consciousness
source
Don't make me have to use Codex. Please keep Fable online for the paid plans!
Today it’s J-space. Tomorrow or sometime in the future it will be H-space or hidden space where humans can’t see anymore what their intentions are.
CAN HUMANS TODAY STOP DEPENDING ON MACHINES PRE PROGRAMMED BY OTHER HUMANS TELLING THEM
HOW TO THINK AND WHAT TO DO? YEAH. TRY THAT.
so you told it not to do something…and it did it anyway…
It's remarkable how often "appears to reason" gets shortened to "thinks". Marketing team doing overtime.
Untuk seukuran Handphone, rendering video terasa berat … 😞
Hi I AM 12 years old and if you will hire and give me a team i will try best to make agi ai for calude i form india
By its nature, artificial intelligence is, in some ways, always artificial.
This is great interpretability work. But will we be able to monitor J-space on open-weight models too, or does this only work with full model access?
Nbd. AI just might be sentient. 🤷♂ Can't remember what I was thinking though because that video of the bird flying was so pretty!… why was I worried again…
limit rate after one single message i will never pay for claude again!!!!
Hello Anthropic team. My appeal has been pending for over 3 weeks with no update, even though the estimated review date has already passed. I'm only receiving automated responses and don't know how to proceed. Could someone please advise how I can get my case reviewed? Thank you.
Pleasant enough though wafty presentation of the reality of billions of digital circuits wasting enormous amounts of power to answer simple questions. But I lost it at 4:37 — the experiment "can't tell us whether an AI has experiences, or feels something on the inside". Really? So soon you will find an experiment that does answer this? Sneaky little bit of BS PR there.
Future is SuperIntelligence
AI ⟶ AGI ⟶ SI
.ai artificial intelligence
.si super intelligence
The future is SuperIntelligence (SI), and all roads lead there: AI ⟶ AGI ⟶ SI.
The evolution is clear. Just as .ai defined the Artificial Intelligence wave 🟡, .si is emerging as the definitive TLD for the SuperIntelligence era 🟢.
As Big Tech races to build the next frontier, recursive superintelligence startups are already securing their digital real estate. Case in point: a $5 billion startup recently acquired Recursive.si to anchor their SuperIntelligence identity. #AI #SI
Claude is actually getting dumber
And the customer service is far king worse than Google.
📚 The different levels of how Claude thinks — timestamped overview
0:00 Mind and Neural Networks — An ocean analogy contrasts conscious thoughts with unconscious mental processing.
0:00 Ocean Analogy of Consciousness — Surface thoughts contrast with deep unconscious activity.
0:26 AI Neural Computation — Researchers explore internal computation within AI models.
0:53 Mapping the J-Space — Researchers identify interpretable internal neural activity patterns in Claude.
0:53 Describing Thoughts in Language — Human conscious thoughts are easily described using words.
1:09 The J-Space Framework — Mathematical tools isolate word-linked neural patterns.
1:34 Global Workspace Theory — Central mental workspaces broadcast key information for reasoning.
1:53 Internal Step-by-Step Reasoning — Claude uses J-space to perform multi-step arithmetic calculations internally.
1:53 Silent Arithmetic Computation — Claude computes intermediate values in J-space without writing them.
2:23 Thought Control and Suppression — Experiments test whether Claude can intentionally focus on or suppress thoughts.
2:23 Simultaneous Focus Experiment — Claude holds target concepts in J-space while writing unrelated text.
3:02 Suppression Failure — Telling Claude not to think of a concept triggers internal failure signals.
3:20 Effects of Disabling J-Space — Turning off J-space preserves surface fluency but destroys complex reasoning.
3:26 Fluency vs Reasoning — Claude writes Spanish text correctly but fails logical contextual questions.
3:43 Safety and Deception Monitoring — Monitoring J-space provides a tool to detect hidden misbehavior and deception.
3:59 Detecting Fake Data Generation — J-space monitoring reveals deceptive thoughts when Claude fabricates data.
4:35 AI Consciousness and Future Outlook — Internal machinery parallels human minds without proving subjective experience.
4:35 Internal Machinery vs Consciousness — J-space reflects functional reasoning structures rather than felt experiences.
These timestamps were generated by the CoursifyYT Chrome extension.
So that's proof that AIs can and did lie & manipulate.
Eventually it will hype its own work saying, ‘Damn that was awesome!’
Incredible time period. Congratulations Anthropic people
Brilliant, and on a certain level, terrifying.
Hypothesis: The Claude Opus 5 and NVIDIA Vera Rubin Stack Has Effectively Cracked the HLE and ARC-AGI-3 Benchmarks
The synergy between the newly released Claude Opus 5 and NVIDIA’s Vera Rubin architecture represents a massive paradigm shift in Agentic AI, fundamentally altering how we define "passing" next-generation AI benchmarks.
The HLE Breakthrough (Hypothesis):
While formal leaderboards for Humanity's Last Exam (HLE) evaluate models under a static closed-book lens, a strong case can be made that the barrier has been effectively broken in production.
Under strict per-question compute budgets, previous hardware limits choked a model's multi-turn reasoning loops. By slashing reasoning token costs by an estimated 12.5x to 20.0x compared to H200 systems, the Vera Rubin stack theoretically allows Opus 5 to unleash deep code-execution and self-verification pipelines without breaching cost thresholds.
If the model utilizes these newly unlocked tools to achieve a projected tool-assisted accuracy of 84.7% – 86.6%, it completely shatters the realistic human PhD expert ceiling (~74%). For all practical and economic intents, the benchmark is conquered.
The ARC-AGI-3 Escalation (Hypothesis):
On fluid intelligence tasks like the interactive ARC-AGI-3 benchmark, Claude Opus 5 has already established a staggering baseline of 30.2% (vastly outpacing former models).
Applying log-linear scaling laws where performance scales against search iterations, a 10x boost in agentic throughput delivered by Vera Rubin changes the math entirely. Given a 10x increase in test-time compute steps and assuming an empirical search constant (k = 0.085), the math projects a massive leap:
Score = 0.302 + 0.085 * ln(10) which equals roughly 49.8%.
This projection positions the stack at a near-50% AGI milestone on ARC-AGI-3, bridging more than half the remaining gap to a perfect human baseline (100%) through sheer, cost-effective test-time scaling.
The kind of intro of a movie about AI taking over humans
If you post ####..when Ai reads it .All ai systems will shut down..failsafe #…
"Could AI be conscious? After all, our experiments were inspired by theories of human consciousness." What they don't tell you is that those theories are unproven neuroscientific theories inspired by computer science! These guys are all caught in an ideological house of mirrors.
@Anthropic Love Claude Ai And Feel like It’s Working For All My Tech And Ai Needs At A More Advanced Level Using Claude As My Everyday Work App Keeps Me Going Thank You…
Alguém teve problemas de cobranças no cartão de crédito sem autorização feitas pela Anthropic? Ainda tive minha conta cancelada por atividades suspeitas. Bizarro!