rnxrx

I have to ask.. what in the world is a "radical centrist" ?

Very much this... The utility of Claude has grown exponentially with the amount of work I've put into memory/context management, tooling, skills development and CLAUDE.md (and friends). I freely admit I may just be fortunate, but I've yet to hit a bug in a new release that rose above the level of minor annoyance.

The thing that really validated this point was when I recently tried Codex again. Out of the box it was OK, but not great. I swung over the memory apparatus and had it import directives from my existing Claude install and it was almost immediately quite useful in its own right and, especially, as a counterpoint to Claude when dealing with some fairly complex operational issues.

It depends hugely on what "rent usage of this model through one of the many LLM hosting providers" means. If you're asking them to host the model privately then yes, all of that 1.6T of RAM is likely in use holding weights, activations and KV cache by an inference engine that's only getting/answering requests from you alone. When you aren't actively using the model the hosting process is still active and waiting with all of that memory still wired to it.

As background: For the most part VRAM oversubscription/paging/swapping isn't a thing in the same way that RAM for a VM often is. There are some approaches to it, but (to my knowledge) not at that sort of scale.

There are some systemic reasons for this, but very broadly speaking the GPU vendors are building toward the highest bandwidth and lowest latency possible, and the overhead/complexity of something like protected memory modes serves neither of those priorities.

$5K seems like an absolute steal compared to paying contracted security experts to find such bugs. I'm surprised these programs aren't pushed harder, as the potential ROI seems fantastic.

I do something similar - as I approach the context limits I have a pre-compact flush skill that extracts anything useful from the context, updates the MEMORY.md and my Obsidian vaults (set up as a poor man's graph DB) and so forth. Once everything's been stored I run /compact to keep the general session flow intact. Recently I added a small embedder/vector search setup to the same skill, which seems promising so far.

On another environment I've been doing something roughly similar, but have integrated Hindsight as a kind of all-in-one of the above and am still trying to suss out the best compaction strategy.

The cost issue is obviously of prime importance, but I'd also argue that the transparency of the innovations creates a tremendous cross-pollination, and not only within the Chinese communities but in the US/Europe as well. How many of us are learning the practical aspects of actually running and building AI based primarily on open models? As an example - how far would the work of vLLM or SGLang or even NVIDIA itself (all random examples) be without these models and the challenges they pose?

This is great - and another example of how much more efficient CLI tool use ends up being in actual day-to-day use. Claude Code and Hermes took it in and it runs great in my initial tries at it. Thanks for making and sharing it!

Another endorsement - I used Kokoro pretty extensively with an app I was developing over the last year and it's been excellent, both on- and off- GPU. Even with Elevenlabs (long time subscriber) the comparative quality of Kokoro keeps up really well until you get to their larger models with their professional voices.

I do wish there were better support for SSML, as well as deeper documentation of how to influence inflection in-line, but the default does well with standard emphasis (e.g. putting asterisks around text elements). Both asks are getting outside the zone of reasonable asks for this sort of distribution, though, and I remain incredibly grateful for the quality of what hexgrad and nazdridoy have put out in the world.

It's worth being careful here: a lot of the affordable enterprise-class routers from 10+ years ago aren't as fast as cheap consumer hardware - like the OP or just a decent mini PC. The primary arguments for the enterprise-class gear are around feature availability and certain aspects of reliability (redundant power/fans, better heat tolerance, higher quality components). It's also worth remembering that this kind of gear tends to be built for dedicated environments: loud fans, higher power draw/lack of power saving features, etc.

Beside the potential performance and environmental issues the other big downsides tend to include firmware availability - either because download from the site requires a login on the vendor's site or, increasingly commonly, the gear has hit LDOS and images just aren't posted. Obviously there are other "unofficial" places for such images, but the risk/legality are a whole other (potentially serious) question.

There's an additional issue mapping the requirements of home networking to enterprise gear: Ethernet switches are lousy firewalls (little or no NAT, primitive built-in security, DLNA/mDNS and friends aren't really sane options, etc). Finally, even at "1/5" the price the gear may still be quite a bit more expensive than other options. And if it's not expensive, it's usually because nobody wants it any more because of the issues mentioned above (being near- or beyond- LDOS).

FWIW this is from someone who literally built a commercial-class machine room in his house with dedicated AC, subpanels, commercial UPS, etc for data center class Ethernet, Fibre Channel and Infiniband switching as well as carrier-grade routers and still runs "enterprise" grade WLAN and switching and can lay hands on as much as I could reasonably want without too much drama or cost... Going down this road can absolutely be amazing if you either have a.) the background to properly source and run the hardware/software or b.) have a driving desire to learn how to do so or c.) have some very atypical requirements for home networking. Otherwise it tends to not be something to be done lightly.

It depends on what's meant by "fully utilized" but fp8 quants of Nemotron 3 Super, the latest Minimax, Cohere A+ and the Mistral small and (especially) medium variants all sit in that 128-256 category, especially with full context or even moderate concurrency. In fact, in a 192GB environment I work with (Hopper GPUs, fwiw) I was pushed into using 4-bit quants with a couple of those to get the model working with a reasonable context window (..but 256 would have rocked out).

There are also nvfp4 quants of Qwen 3.6 27/35 floating around. I've done benchmarks of both and the quality difference vs fp8/bf16 was barely notable. Honestly the nvfp4 capability is the most interesting feature of the Spark (at least for me).

There is at least one MCP server in Obsidian's community plugins, plus the REST API access capability which is already addressed in several open source MCP plugins.

I use Obsidian as a persistent context store and knowledge graph (..loosely defined, i.e. link/back-link) for both Claude Code and Hermes, while also using it to generate live Wiki pages for working documentation. The native replication and the Git integrations work well keeping it all synchronized across multiple harnesses, as well. I use the native MCP server mentioned above, plus just letting the agent work with the markdown files directly.

That said, having built out all of this manually I'm excited to try out something that addresses much of this out of the box. I'd also be curious about the integration with Hermes/OpenClaw/etc.

It *is* hard to make memory, especially HBM (...which is what the AI market wants, and is what the manufacturers are focusing on) and bringing on new capacity takes years. There's the additional wrinkle that the manufacturers we have left are the ones who survived periods where the market was glutted with oversupply in the wake of previous shortages.

These decisions play out on the order of trillions of dollars and 3+ year horizons. They're also incredibly sensitive to other geopolitical issues (Taiwan, issues with Chinese tech capability vs export/import controls, etc).

There are a lot of valid discussions to be had about how we got to this state of oligopoly: Taiwan's consistent sponsorship of its semiconductor capabilities and the subsequent concentration of technology (expertise, capacity, etc), the lack of investment/support (and ceding of technical leadership) in Western countries, the various rivalries with China and the implications of it becoming a first-class producer of semiconductors at scale, etc. None of those discussions and none of their potential outcomes can substantively change that we're going to continue in this situation (massive price increases, spotty availability, etc) for at least the next 18-24 months.

It’s usually open loop - closed loop, so closed loop goes through CRACs or liquid cooled equipment manifolds. That heated water circulates through an heat exchanger on the roof that uses open loop cooling to shed the heat to the surrounding environment.

In terms of platform variants the Taurus outsold the Crown Vic by a fair amount. The latter was certainly hugely popular as fleet vehicles, but the Taurus was Ford's best selling sedan for quite a while. The Taurus was never rated for anything more than fairly light-duty towing (<= 1500lbs).

More to the point - at least for automatic transmission variants in the early 90s - even the base F150 could certainly out-tow Crown Vics of the time, unless the latter had a tow package, which would push the rating up to 5000 lbs (..same as the lower end F150). Later on the delta was a lot greater (in the truck's favor) as the tow packages for the cars were phased out.

Nowadays the minimum ratings of F150s (even with shorter beds) is 5000lbs, or more. The smaller CUV/SUV platforms are usually rated ~1500lbs *max*, with some of the larger truck-based models obviously running in the same range as the trucks.

TL;DR - Even back in the day the "typical" sedan was still only rated for towing small loads, unless specially ordered with towing packages. Now..all that said, the frequency with which typical F150 owners ever actually tow anything is a whole other question.