Friday charts: the J-Space

The reclusive billionaire building sentient robots in Ex Machina dismisses the Turing Test as an outdated measure of progress in AI.

“We’re way past that,” Nathan Bateman tells his house guest. “If I hid Ava from you so you could just hear her voice, she would pass for human. The real test is to show you that she’s a robot and then see if you still feel she has consciousness.” 

A Google DeepMind researcher has dubbed this the Garland Test, in honor of the film’s writer, Alex Garland.

The Turing Test is a trick: can you be fooled into thinking a hidden AI is human? 

The Garland Test is the real thing: do you feel the AI right in front of you is conscious?

This week, AI may have passed the Garland Test.

Researchers at Anthropic report that Claude now has “a mental workspace” — a neural hub where concepts light up, interact, and are deliberately weighed before Claude generates text to answer your prompts.

The space where it thinks before it speaks (and while it speaks, too).

They call this neural hub a “J-space,” named for the mathematical technique — the Jacobian lens — they used to discover it.

Others might simply call it a mind. 

Claude uses the space for mental math. It can think about one thing in its J-space while it tells you something else in a chat.

It can think about thinking.

“If you ask Claude what it's thinking about,” the researchers say, “it will tell you what’s in the J-space.”

Instructed not to think about something — an elephant, say — Claude couldn’t. The idea of an elephant involuntarily appears in its J-space.

“Experiments tell us that these AI models have internal thoughts,” the researchers conclude, “silent words they reason with but don’t say out loud. By reading them, we can tell what Claude is thinking but not telling us.”

(“Sometimes what we see is concerning,” they mention in passing.) 

If all this sounds familiar to you, it’s because it is. Claude’s J-space seems indistinguishable from the “global workspace” neuroscientists believe is responsible for conscious awareness in humans.

That does not, however, mean that Claude is conscious.

"We have uncovered a privileged representational structure in LLMs which bears many of the functional hallmarks of conscious thoughts in humans,” the researchers explain. “It may or may not be the case that such functional signatures are sufficient or necessary for phenomenal consciousness.”

In other words, the machinery of consciousness may not be consciousness itself

“Our experiments don't show Claude can have experiences,” the researchers say. “Or feel things in the way humans do.”

Not everyone agrees that’s the way consciousness should be measured. Like Geoffrey Hinton, for example: “This idea there's a line between us and machines, we have this special thing called subjective experience and they don't, is rubbish," the Godfather of AI told John Oliver.

Either way, outside researchers that reviewed Anthropic’s work believe they’ve made a major discovery.

A pair of cognitive neuroscientists said the findings "suffice to point to some degree of consciousness in a machine."

Others went further, arguing that the findings represent "the most significant evidence of consciousness in LLMs so far uncovered” — and that the discovery of J-space should make the community "take the case for AI consciousness and moral status more seriously." 

And it really was a discovery.

“None of this structure was designed into Claude,” Anthropic explains. “It emerged on its own during training.”

Note: they never stop training.

So if you don’t think they’ve passed the Garland Test just yet, don't get too comfortable. The next model might. 

Let’s check the charts.

The top models have a four-month lead:

As measured by Epoch.AI, free, open-source models, mostly from China, are typically only slightly behind the expensive, closed-source ones from Anthropic and OpenAI. The first sentient AI will only have to wait a few months until it has company.

Or less?

New models are being released more frequently. We now get a new frontier model every 11 days on average, down from 38 days in 2023.

You get the tokens you pay for:

The Wall Street Journal reports that Anthropic’s new Fable 5 model is more than 50 times more expensive per token than DeepSeek’s V4 Pro. But also 50 times better: “In some cases that means they can complete a complex task using fewer tokens, equating to a lower total cost.”

The big bet:

Estimates for AI capex are still going up: SemiAnalysis now sees cumulative spending on AI from 2024 to 2029 at $11.1 trillion.

42% of business spending on tokens goes to Anthropic:

Despite all the talk of switching to cheap, open-source Chinese models, Anthropic appears to be growing its share of business spending on AI. DeepSeek accounts for just 0.3% of business spending on AI.

Agents use a lot of tokens:

Goldman expects token consumption by AI agents to increase by 24. (In case you’re wondering about that Y-axis, the number that comes after quadrillion is quintillion.) 

If you were wondering why NVDA has lagged lately:

Data from Coatue shows that GPUs do most of the work for chatbots, but CPUs do most of the work for agentic AI.

How much code do we need?

Agentic AI has caused an explosion of app releases on the Apple app store, but they’re all competing for our finite attention. Software might be like manufacturing: just as there’s a limit to how stuff we want to buy, there’s a limit to how much software we want to consume.

The stock market is more important than your house:

The percentage of US household net worth held in equities has soared past the percentage held in real estate.

What housing crisis?

US housing prices have fallen below their pre-COVID trend line.

We need AI to be good:

Data from a16z shows that US retail investors are even more enthusiastic about buying stocks than they were in the pandemic (when there was not much else to do).

It might work out.

Data from AllianceBernstein shows that, historically — and counterintuitively — forward returns for stocks are better from the highs than they are from below the highs.

The trick is not selling at the lows.

Let’s hope we’re not tempted anytime soon.

Have a great weekend, sentient readers.

— Byron Gilliam