> We chose a different consensus algorithm for Meerkat, called QuePaxa, that aims to avoid the “tyranny of timeouts” imposed by protocols like Raft. QuePaxa is a subtle protocol, but here are the highlights. A client can contact any replica, and that replica can drive consensus for the latest slot. There is a leader, but it is not required — its only advantage is that it can drive consensus with fewer round trips (one) than other replicas (3+). Critically, clients are free to contact multiple replicas concurrently for the same proposal, to increase the chance of the proposal being successful. Concurrent proposals do not destructively interfere: replicas work together to decide one of the proposed values.
This product feels bad and sloppy, so I’ll give my hypothesis for why this was built:
At this point, Anthropic is likely having Claude itself propose and build features autonomously based on providing it with raw user feedback. This could be one example. Which is why it has an eerie sense of redundancy and pointlessness (“You mostly used Claude to automate work and home tasks”, etc.).
Exactly. The spec should go in a git comment, NOT vcs. Or, you keep the document in history for the duration of the PR and remove it post review or pre merge
Looking at the implementation sketches, this algorithm looks even trickier to implement than Paxos (already a notoriously tricky algorithm to implement) and on top of that, I think the failure case in this algorithm is subtle and different -- very long tail latencies. In Paxos/Raft the latencies are more likely bounded by the timeouts (not eliminated of course) so you can build other systems to expect certain delays, but in this case, you may write something, wait for an ack, then abandon and retry, then realize the old write succeeded, etc. ad infinitum.
This has very little to do with “AI replacing jobs” and much much more to do with a bad product getting obsoleted by better ones.
Human labeling is a two sided marketplace and so as any marketplace startup knows, both sides need to be constantly nurtured otherwise the system can collapse as worsening quality leads to churn and a vicious cycle that empties out the platform.
In labeling, you need to understand the limitations of individual work and fatigue, keep your pipeline bursting with awesome and consistent work, and improve the platform to make customer experience great.
AMT has been totally languishing in all these respects. Pay is terrible, dishonesty rampant, etc. It was a bad product, no need to pedestalize it or turn it political
In Gemini at least, if you look at how they process PDFs, they do an OCR and then feed the text + image to the model, without charging you for the text tokens (I believe).
So my guess is that Claude’s backend is doing the same — so this hack is probably more of a loophole in token accounting that might get closed if Claude is doing what Gemini does
Can you speak to what makes this different from simply including or configuring various agent skills? Or is it simply the combination of lots of helpful defaults that makes this product useful?
To all of those thinking that GLM/OSS will save you — keep in mind that the model size needed to compete here likely requires an NVL72 or similar — 72GPU dedicated infra to run a hosted model. This will almost certainly get regulated by the gov’t as well, and even if not so there will only be a handful of companies that can afford it.
I simply don’t see the value here. At the cost of these systems its inevitable that you hire at least one person full time to manage your rack infra, and that person would likely prefer the customizability of commodity hardware.
Source: I run a startup that recently did our datacenter buildout. At no point were we at all interested in a single integrated vendor.
Tauri doesn't lock you in to one JS ecosystem. In fact, it doesn't require you to use javascript at all.
Also, we've had several developer framework startups get acquired -- Astro, Nuxt, UV, Bun, Vite. It doesn't exactly inspire confidence in a software that you want to last and give support for years.
Claude code is a very small fraction of the code written by Anthropic and ultimately, despite being widely used, hardly dents Anthropic’s core priority of LLM performance.
It’s because Anthropic doesn’t publish any of its core AI research that we falsely believe that it isn’t by far the central focus of the majority of the team.
Just to be clear, I’m not supporting their stance nor defending the company. I consider it to be deeply harmful that a private company seeks to advocate for AI safety but then own all the means of production and profit financially from keeping its techniques secret. It’s as if the Manhattan project resulted in a for-profit company selling all atomic technology and deciding on its use.
If you look at Anthropic's blogs about their model timelines, there is roughly a ~3m period between a model being in internal preview until release. That means that inside of Anthropic, the next version beyond Mythos/Fable is already in preview, already being tested internally. Despite what Geohot is describing here about GLM, my understanding is that Anthropic employees have spent a significant amount of time grappling with a technology that is considerably ahead of what is available to the public today.
In addition, if you look at the graph of LOC written by Claude vs Ants (I.e. AI vs human), there is an incredibly sharp uptick post-Mythos internal preview. Something like from 30% to 75% of code inside the company being written by AI.
While I sympathize with the viewpoint here, I still have to admit that that there's a very different feeling to working inside of a company where they've had months of time with a model that's at the frontier, quickly changing the way everyone around them works, and that _they themselves_ control the keys to.
If Geohot had those keys, I can be 100% confident he'd be raising the alarm at the top of his lungs about it.