mondrian

Good point. In the scenario I described, I'm assuming Stripe will launch their own stablecoin. I tend to think all major tech companies are incentivized to launch stablecoins and give you discounts and perks when you transact using their stablecoin in their own ecosystem. The more of their stablecoin they issue out, the more money they make on interest.

Yeah and this is codified in the GENIUS act which passed recently. It enables tech companies to act like banks in certain dimensions, without being regulated like banks.

Yeah. Stablecoins create demand for Treasuries which drives the price of Treasuries up and interest rate down. So this pressure lowers debt servicing cost for the US government, and Stripe is the holder of those Treasuries and gets paid interest.

This would also serve to counter the drop in global Treasury demand due to recent tariff stuff where presumably our traditional debt holders are losing appetite for US debt...

It also creates a kind of strange situation where stablecoins are basically spendable "Treasury tokens". So you give 1 USD to Uncle Sam (via a middle man like Stripe), get back 1 stablecoin. Then you go and spend the stablecoin, and Uncle Sam goes and spends the USD. It's like a weird double spend situation. Prior to stablecoins, you buy a treasury bill with USD, you hold this unspendable treasury bill while Uncle Sam gets USD to spend.

The GENIUS act enables tech companies to become reserve holders -- buy US Treasuries with customers' money. Stripe offers a "transactional ecosystem" to the customer in stablecoins, the customer gives USD to Stripe in exchange for stablecoins, Stripe buys short-term Treasuries and makes a shitload of money on interest.

Part of the very high level play is the US Govt seeks to diversify away from depending on nation states for borrowing, and to promote tech companies to the status of reserve holders.

This doesn't add much to the consumer however. I think in fact we are looking at a "fragmented currency" future where you hold like 36 different stablecoins in your wallet because certain platforms accept certain stablecoins. The GENIUS act doesn't offer strict guarantees for getting out of a stablecoin into USD, so I predict dark patterns and "incentives" to make it hard to get out of a stablecoin.

Trying to make AIs more factually-accurate with more training is probably hopeless. Current events and encyclopedic knowledge will be provided by tools. The LLM's core job is to choose the right tools for the job and synthesize their outputs.

Also in the vicinity: https://www.anthropic.com/research/tracing-thoughts-language...

There's also distillation, where you can drastically improve a small model by training it on chains of thoughts of larger models. You can't achieve the same performance by training on original human texts. This suggests that those chains of thoughts reliably contain "densely packed reasoning", meaning the LLM probably has developed internal clusters of "reasoning circuitry", loosely speaking.

Related to this: https://x.com/karpathy/status/1835561952258723930

The "core reasoning" part of AI may be increasingly important to improve, and its "database of factual knowledge" aspects may be less and less important, maybe increasingly a hindrance. So more focused and specialized training may take over toward increasing reasoning precision, and not this never-ending stream of new data.

So maybe we'll get better reasoning and therefore better generated data/content in the wild, without this negative feedback loop everyone is worried about.

Home building is interesting because I think a major blocker to monopoly-forming is the vastly heterogenous and complicated regulatory landscape, with building codes varying wildly from place to place. So you get a bunch of locally-specialized builders.

Regulation can increase concentration in a high corruption/cronyism environment — regulatory capture and regulatory moats. There is plenty of that happening.

In building, I think we have local-concentration, due to both regulatory heterogeneity and then local cronyism - Bob has decades of connections to the city and gets permits easily, whereas Bob’s competitor Steve is stuck in a loop of rejection due to a never ending list of pesky reasons.

It's a read/write protocol for making external data/services available to a LLM. You can write a tool/endpoint to the MCP protocol and plug it into Claude Desktop, for example. Claude Desktop has MCP support built-in and automatically queries your MCP endpoint to discover its functionality, and makes those functions available to Claude by including their descriptions in the prompt. Claude can then instruct Claude Desktop to call those functions as it sees fit. Claude Desktop will call the functions and then include the results in the prompt, allowing Claude to generate with relevant data in context.

Since Claude Desktop has MCP support built-in, you can just plug off the shelf MCP endpoints into it. Like you could plug your Gmail account, and your Discord, and your Reddit into Claude Desktop provided that MCP integrations exist for those services. So you can tell Claude "look up my recent activity on reddit and send a summary email to my friend Bob about it" or whatever, and Claude will accomplish that task using the available MCPs. There's like a proliferation of MCP tools and marketplaces being built.

Concentration is the default in an unregulated environment. Sure pure monopolies with 100% market control are rare but concentration is rampant. A handful of companies dominating tech, airlines, banks, media.

Because everyone worships Thiel's "competition is for losers" and dreams of being a monopoly. Monopolies being the logical outcome of a deregulated environment, for which these companies lobby.

It’s still unconvincing that the shift to AI is fundamentally different than the shift to compiled languages, the shift to high level languages, the shift to IDEs, etc. In each of those stages something important was presumably lost.

You have a card with a name, when you click it a detail popup opens and the name is in there. There's also a delete button saying "Delete <name>", which pops up a confirm dialog and the name is in there, too. When you update the name you have to update all those places.

The alternative is there's a canonical name variable, and it's rendered in all those places. To update the name, you just update that variable and "re-render", and those places naturally pick up the new value.

The problem the name will have to be updated in 6 places in the UI.

You really think so? How about the alternative being he's fully pardoned of anything that might be brought against him and coasts to billionaire status by launching companies and having his funding rounds and paths to exits secured now that he's an absolute NRx legend with the full backing of Thiel and Musk?

I mean it seems you’re dunking on a strawman yourself. Like if I said Denmark virtually denies no life saving claims, and when it happens it’s due to edge cases, you’ll insist just because it occurs it’s indistinguishable from a system where it happens systemically and regularly?

There are systems that cannot deny life saving care, and where everyone is necessarily insured.

It’s facetious to compare those to a system where 30 million have zero coverage and the rest are systematically denied life saving care as a profit making mechanism.

And yeah letting someone die when you could help them live is violence. When it’s baked into the rules it’s called systemic violence.

That’s pretty interesting, thanks. It’s related to abstraction (interfaces) vs representation (implementation) in programming. To my eye, there’s no conflict that 1 \in 3 is true in one representation and not in the other. Trying to use \in like that seems like a violation of an abstract interface, somehow expecting that the various implementations of an abstraction must be identical. It also doesn’t seem to violate platonism, since “numbers” are abstract ideas that are not expressible directly in set theory. Set theory can only encode concrete representations of numbers. Much like any physical chair cannot be identical to the platonic Chair.

Tl;dr (a) program synthesis is provably hard and (b) English is too ambiguous as a specification language. Therefore we’re nowhere close.

Ultimately not a convincing argument. We could use this same argument to argue that a human cannot write a program to spec. That may be strictly true, but not interesting.

JD Vance is on record mentioning Yarvin’s ideas as inspirational for how republicans should run government.

This is a deep insight. It's a reactionary vs. establishment dynamic where the reactionaries get a free boost because they're fundamentally more provocative from a content perspective. I think it's more like "the reactionaries get to hate, the establishment doesn't" and R and D may swap those positions.

The US is currently in proxy war with Russia on two fronts: Ukraine and Israel/Iran. Conceding Ukraine would help Iran, which is probably not what Israel wants. The idea that Trump would be friendly to Putin -- lift sanctions, give him Ukraine -- seems like a strategical contradiction.

Curious about this statement, do you mind expanding?

Relatively speaking, Instagram and Twitter are far more interesting than LinkedIn. LinkedIn is a special kind of horrible.

There's a gradient of risk, though. Suppose someone is sitting on 5 billion dollars unrealized gains by holding the index fund VT. The idea of "risking 100% loss" in VT is ludicrous.

Maybe unrealized gains tax can be formulated as "cushy gains tax" if riskiness can be quantified in a reasonable way based on an asset's age and metrics over time. Then if an asset's risk score is above X, you don't pay tax. If it drops below X, you start to pay tax.

This would probably lead to more innovation and fewer monopolies, as people are incentivized to invest in riskier companies, and companies are incentivized to self cannibalize to maintain a healthy risk score.

Good points regarding test quality. One takeaway for me from this paper is that you can increase code coverage with LLMs without any human checking of LLM outputs, because it’s easy to make a fully automated checker. Pure coverage may not be super-interesting but it’s still fairly interesting and nontrivial. LLM-based applications that run fully autonomously without bubbling hallucinations up to users seem elusive but this is an example.

Recent work from Meta uses AI to automatically increase test coverage with zero human checking of AI outputs. They do this with a strong oracle for AI outputs: whether the AI-generated test compiles, runs, and hits yet-unhit lines of code in the tested codebase.

We probably need a lot more work along this dimension of finding use cases where strong automatic verification of AI outputs is possible.

I 95% agree, but "what people want" is probably not a strong indicator on the thresholds of paradigm shifts, since people don't know what's possible.

I fully agree, and that is my experience. I am just grappling with what looks like quite a bit of misunderstanding around aphantasia. Some people make it sound like it involves imaginary objects in the visual field, causing other people (who probably don't have aphantasia) to start suspecting that they too have aphantasia, because when they close their eyes they can't see vivid images on the backs of their eyelids.