Is it possible to show that LLMs can't reason? It's easy to show that they can't reason the way we do, but hypothetically it might be possible to somehow couple its "semantics" coherently. Unless this is you however, you really should shut up about LLM intelligence though.

How to mitigate latency in a deterministic system. This is an almost entirely novel concept which has very little literature supporting it.

Gemini does a great job at summarising our paper

LLMs automated design patterns, that is: stochastic pattern-matching on signals without any consideration for semantics. Neither of them ever really worked, but both look like they do at first glance and since people don't like to think about semantics, they both took root

I'm reading @markburgessosl.bsky.social draft paper on attention in AI. Some 🤯: - LLMs are slow because of the incredibly wide context when evaluating, so that's also a boundary problem - Boundaries are the same as a rejection of a violation of an autonomy, so semantics are agentic

If you want to issue stablecoins as defined by the GENIUS Act, you need to buy treasuries in equal amount. I wonder if its goal is to bring US debt back on-shore. It's amazing how the status quo is held up by the Two Generals' Problem and how that's mostly a misunderstanding.

All clocks are wrong but some are useful. 🧵

Can someone explain to me why CAP was adopted into the canon based on a paper, but Paxos required Chubby to even be noticed?

The problem of state-management is that processes are allowed to make decisions. We project free will. It's a kind of anthropomorphisation. In fact, all their "free will" is just a byproduct of their incidental entanglement, ie: unforeseen causal relationships between them.

Whatever you might think about LLMs, it's matter of fact that they are tools of cargo-culting, meaning that they imitate the external signals without understanding the internals. How that is not considered a problem is beyond me.

If you believe the conclusions of CAP apply to all distributed systems, you believe that there's no way to not have the occasional latency spike or even node isolation event happen to your service, unless you have an incredibly well tuned network. Our next project disproves this

LLMs and Tesla's FSD are the same problem. Both can only succeed if they can consistently apply meaning and reason about it. A bush with feet is a human carrying a plant. Now do this with orders of magnitude more complexity and you see the problem. Probabilism won't save you

We're planning a new kind of service, one everyone thought impossible because of CAP, that should fix most distributed systems problems in a very simple way. If you have control-plane problems or service-integration issues, send me a DM or reply here and I'll get in touch.

I did find it amazing at the start, (but got used to by now) is how nobody refutes our innovation at its substance. People who take the time to understand it (without exception) also accept even our wildest claims. Few actually care about truth if it contradicts their beliefs

A thread about why pessimistic systems are bad at scaling.

Let's analyse Spanner using the model explained here. Spanner relies on pessimistic locks, so we first need to ask the question of what it is in a distributed system. Simple: a pessimistic lock is an optimistically written record, sometimes on a remote node.

My goal with these threads is to outline the minimal semantics of distributed consistency in a generally accessible format. I think these questions have been mostly ignored for decades, which is a shame, since they can lead to practical results. In the last one, we'll tackle CAP

Let's first reformulate Linearizability to mean that time-order should never contradict any process's observed order of events. This will help us in mapping out the ultimate limits of our semantics in space. So: clocks at different distances, measured in ms (communication time)

So, when is the industry going to be ready to talk about decentralisation? I mean, you can argue that CloudFlare engineers know what they're doing, but that also means that this problem is too complex to be solved centrally.

There's a great interview with Leslie Lamport, in which he makes the point that any time-ordered system will contradict the observed (causal) order of events as seen by different processes in different places. He's clearly right about this, except for his conclusion.

Time is experiencing changes. In distributed systems, we focus on composing these changes eagerly into a unified view for any potential observer. This eagerness is the bottleneck. If you just lazily calculate each observer's current view, the known problems just don't apply.

The basic facts of physics tells us why distributing systems is hard. It's because different time-coordination systems move at different speeds. The only logical conclusion here is that these need to be calibrated to each other if we want to establish order between them.

Second outage in two weeks. I guess it's time to be educated about how nobody could do better than be down for half a day again.

The AWS outage has nominally happened because of DNS, but the root cause is that we centralise data-collection and management. If data was delivered to observers from multiple sources (with a shared timeline), central outages cannot happen. Not all reactors are Chernobyl

The consensus in IT that cloud providers failing centrally is still the best of all possible worlds is not only self-serving but also false. It leads to clients not asking questions so everything staying the same. I think the reason developers excuse this is because 1/x

People offload to cloud-providers, not because that is the technically superior thing to do, but for the same reason they hire McKinsey, to shift blame from themselves.

I love how absolute science is. It's the only area where you can be unapologetically certain about being right even if everyone else is telling you that you're wrong.

Wrt today's failure: There were two things that surprised me very much after I started working in software innovation. The first was how receptive academia was to new ideas. I'm very proud of the fact that so many took the time to understand how we solved the fundamental 1/x

Some AI exec: I've seen a future version of our product and it's so powerful, it scares me.... (stares off in the dark with a tear in their eye). I wake at night, trembling for the future of humanity... Reporter: Can we see it? E: No, it's too powerful!

Call me a Luddite, but I don't think it's great that the tech-sector is beholden to a small, entrenched group of VCs. It's ironic that "incumbents only do innovation theater" used to be their rallying cry.

'm convinced that the love for CAP, Kubernetes and microservices come from the same place. CAP was taken to mean that you can't ever have a perfect platform, so might as well cook your own. Simplicity is the last thing they want. They would much rather feel clever.

A fundamental principle of spacetime is that the further away you are from a source of information, the longer it will take for it to travel to you. This increased time means your local time can progress faster, so the remote clock will have "bigger chunks of data".

Krasznahorkai's work _is_ the malaise carried by Hungarian intellectualism. In his interviews, he makes his loathing for Hungarian nationalism very clear, but I don't know a single serious thinker here (in Budapest) who doesn't share his alienation.

When Imre Kertesz won the Nobel in literature in '02, almost nobody knew his work. As for Krasznahorkai, it's a different picture. Almost everyone I know read something from him at some point and certainly know his work with Bela Tarr. Literature is only half-dead.

I don't see many posts of people "buying" innovation, yet people talk about "selling" it all the time. Why? I don't think people give nearly as much thought about buying as sellers, who are drawing out careful plans on selling them.

The confusion comes from the model. A process is data processing data, so it's not (necessarily) a source of information. The jury is still out, but I think people are (original sources). Generals can "think of things", processes can't.

Absolute time is a bad idea because nodes on our grid can become isolated, in which case other nodes can't know if they are missing data, ie: "are consistent". A relative clock, disallowing isolated nodes from progressing the global timeline work around this. This isn't magic

With hindsight, I'm surprised the Wright brothers didn't need to build an entire commercial airline before they were taken seriously.

A shared key between two distant locations can't be updated faster than the one way communication between the two, say 120ms You can trade fairness of access for update speed. Placing a key somewhere means prioritising local updates over remote ones.

The problem with the "Two Generals" is the same as CAP or PACELC: they start from overly restrictive conditions. We know since Paxos that we can have multiple copies of the "same" general and our paper explains how to freeze partial/subjective time for decentralised observers

I wrote a short article about why clearing is going to be a fundamental problem for stablecoins and why established academic thinking tells us that this problem is unsolvable. The next article will explain the solution. andrasgerlits.medium.com/why-clearing...

When coding, it feels like we're creating processes that evaluate their data and makes decisions, but that's just the way we update shared data. Seamless data-sharing is therefore the best possible way to do microservices, orchestration or any kind of integration

"Statistical observation is an arbitrary decision on top of an already arbitrary model, so is therefore far removed from human perception." I think...

Why lack of global time-coordination is the root (cause) of all evil in our industry

It's simple, actually. If life emits signals, its meaning is whatever observes and interprets those. I call it Shannonism. All donations are used to establish the mother-church.

People don't need good contextual autocomplete for automated coding. They need an auriga behind them, constantly whispering in their ears: "What are you trying to accomplish here, exactly?"

Evaluation of a paper from 1977 arguing for handling events as first class members in data- defined systems. I wish people would have listened to this guy at the time. xtdb.com/blog/update-...

The age of the high IQ idiots

It's interesting to see how when a society decides it can no longer reach a consensus on morality, the whole system unravels. In the end, morality is what guarantees that the US pays its debts. We put too much trust in institutions and too little in our fellow men to "do right by others"

The biggest shortcoming of the status quo in distributed consistency is that scale defines semantics. This is what our innovation fixes.