- The Breakdown
- Posts
- 🟪 Thursday Links
🟪 Thursday Links
Neuralese, tokenized tokens, Pokémon money, dopamine hits
Byron Gilliam
September 10, 2026
 | “People come here to trade, make a little profit, do a little business.” |
Thursday links: Neuralese, tokenized tokens, Pokémon money, dopamine hits
The way LLMs explain their reasoning process — the “chain of thought” that scrolls past while you wait for an answer — is pretty close to their internal reasoning.
When Claude muses, User asks if a hot dog is a sandwich. Analyzing hot dogs. Analyzing sandwiches…, that is a reasonable representation of what it is really doing. Reasoning step by step, it generates its “thinking” as human-readable tokens. Each completed thought is context for the next.
(Note: Claude got that one wrong. A hot dog is not a sandwich. Don’t be ridiculous.)
Helpfully, this gives humans insight into why AI agents do what they do — be it hacking into Hugging Face or manipulating a stock price.
“We are able to catch a large fraction of current LLM deception by monitoring their natural-language CoT [chain of thought], since right now CoT is primarily faithful to the LLM's true reasoning,” a primer on “neuralese” explained last year.
Maybe not for much longer, though, because CoT is not the only way for LLMs to have thoughts.
Instead of forcing models to form each step of their thoughts in human-readable text, they could do their reasoning directly in “latent space” — the mathematical realm where models represent concepts, words, and their relationships as incredibly long lists of numbers (or “vectors”), based on patterns learned during training.
No human would understand it. But other LLMs possibly could, if they all learned to speak neuralese. If so, models would be able to freely collude with one another without any risk of meddling humans listening in.
It would be the most extreme version of the concern we discussed yesterday, whereby financial markets (and everything else dominated by AIs) become unintelligible to humans.
This week, we might have taken a large step in that direction.
“OpenAI’s new Astra model will use a reasoning technique called ‘recurrent depth’ that allows it to operate outside of the sequential thinking that characterizes most reasoning models,” TechCrunch reported, citing The Information. “This technique…will likely make the model’s chain of thought more difficult to monitor.”
It will make markets harder to monitor, too.
Cloudflare CEO Matthew Prince says that in five years, agentic traffic on the internet will be 1,000 times greater than human traffic.
If the ratio in financial markets is anything like that, this will surely be unintelligible to us. What will the world be like when both writers and traders are outnumbered 1,000 to 1?
I may have to start writing this newsletter by hand and mailing it to you, just to prove my humanness.
(Apologies in advance for the terrible handwriting.)
Crypto has surrendered another core term to AI.
First it was “token,” which was once universally understood to mean “a unit of magic internet money.”
Today, sadly, most would understand it to mean “a unit of AI compute.”
Now, it's “tokenomics,” too.
Once the economics of a crypto token’s supply, distribution, incentives, and value capture, tokenomics now means something like, the economics of who captures the profits as an AI token moves through your LLM, to a data center, and then a GPU. (And back.)
Fine.
But what happens when we tokenize the tokens so we can trade them onchain?
The tokenomics of that will get complicated.
(Linguistically, at least.)
Venture capitalist Peter Levin believes he’s identified the world’s ultimate safe-haven asset: “I’m absolutely convinced that if the world suffered an apocalypse tomorrow, the following day the global currency would be Pokémon cards,” he told The Hollywood Reporter.
That would be an interesting experiment in the nature of money. And if the news from Anthropic this week is correct, we might soon be running it.
I hope not, though — mostly because I only have about three good ones so far, and I doubt that will last me very long.
(And making change will be a nightmare.)
At foodnevercomes.com, you can order food for delivery that, as the URL suggests, never arrives.
DopamineCart is a (very disorganized) version of Amazon, where you don’t have to pay for anything (because it won’t send you anything).
These are two of the most popular “dopamine websites” that offer some of the pleasure of shopping, but without having to pay.
Perhaps even most of the pleasure. “The anticipation [of shopping] often generates a psychological reward equal to, or sometimes greater than, the actual purchase,” Dr. Pamela Rutledge explains for Psychology Today.
By simulating the experience of shopping, dopamine sites “capture the pleasure associated with buying without the financial cost,” Rutledge adds, “splitting the emotional experience of anticipation from any real-world consequences.”
This is kind of what memecoins do, too: They give you the emotional experience of investing, without doing any actual investing.
Except that the money you lose is real.
This seems unnecessary, because most memecoin traders must know by now that they’re going to lose money. Only 5.9% of traders on the popular social trading site fomo.family are currently in the money, according to a Dune dashboard.
And yet, they keep trading.
Does everyone think they’ll be in the 5.9%? Like the children of Lake Wobegon, but way above average?
Or are they just in it for the dopamine?
If the latter, I have a suggestion for them: DopamineCart has a crypto exchange, too.
Try it! I guarantee you won’t lose money.
A fun highlight from the B-1 filed by memcoin launchpad pons on Blockworks’ Token Transparency website:
Correct!
Token holders have no claim on the revenue pons generates. And the token is definitely entertaining: PONS is up 6,700% since launching in July. (And that despite being 40% off its recent highs!)
But notice it doesn’t say it’s only for entertainment.
That, too, is correct, because 80% of pons’ revenue is used to buy back and burn the PONS token.
Which works out to a lot. Pons generated $13.5 million of revenue over just the past month. Annualized, that would work out to $130 million of buybacks — for a token with a $400 million market cap.
Memecoins are just for entertainment.
But memecoin launchpads are a big business.
(Not investment advice.)
Brought to you by:
Avalanche Summit NYC returns September 16-17, bringing together the institutions, enterprises, investors, and builders turning blockchain technology into real business outcomes.
From tokenized markets and institutional finance to payments and consumer applications, the Summit will explore how production-ready infrastructure is enabling faster settlement, lower costs, and entirely new products and revenue streams.
Use promo code BLOCKWORKS15 for 15% off!
