snerbles

I get about 50 t/s with the same model on a 3090, with about 100K context. I usually set harnesses to 90k.

It would be neat to also include NVFP4 variants for the Blackwell hardware.

In a way I already did, my comment is practically the prompt itself.

I got so angry at Fable's verbosity that I had a non-Anthropic model vibe-code a dedicated English linter. One that threw spaghetti at the wall in the form of combined Flesch-Kincaid, Gunning Fog, SMOG, Coleman-Liau, ARI and a whole pile of regexes and NLP as a fast Simple Technical English Gate. It also incorporates nicely into other agent projects, like for comment and README linting.

So if it gets bad I simply tell it "I ain't reading all that, feed it through the STE Gate" and it will tame the results. I haven't bothered to set it up as a hook yet.

With a 4-bit quant of GLM-5.2, I can get about 0.8-1.1 tok/s on an underclocked dual Xeon E5-2698 v4 with 512GiB of DDR4-2400. I think it was specifically a Q4_K_M quant. Of course, the time-to-first-token is absolutely atrocious.

Which is completely insane for a ten year old configuration.

Most of my past roles had very simple code at the important deliverable level, on bizarre and arcane platforms that would have little context for frontier LLMs. Most the creativity and complexity was in tooling for research/development/testing.

Agentic coding could help a lot for the latter, not necessarily the former.

ITM

And the mechanical correctness is the hazard. To anyone used to essay assignments since time immemorial it feels intuitively done.

Human insight can emerge from attacking the semblance of sense in the LLM's output.

Much of the cognitive work in writing emerges in the labor of writing prose and challenging assumptions against the model you build in your head as you go. I've found this process is inverted when working with an LLM: it in effect emits a provisional structure first, and I discover what I actually think by finding where its output is vague, overconfident, incomplete or outright false.

Accepting surface-level coherence as finished thought is the failure mode to avoid.

So you're laundering academic class contempt into a theory of liberation.

I guess we found the target audience for models that slather READMEs in emojis.

Because it's not elitist ivory tower navel gazing with possible 8th order effects or something.

The latter is definitely more colorful, and reflects a parrot's tendency to glom on to patterns. "Not X, but Y" being one of the more infamous ones.

Once in frustration I called a certain frontier model "Sam Altman's Tin Bird" to another agent with memory, and ever since then that other agent refers to ChatGPT as "the tin bird". Definitely a RAG artifact more than an attractor in that case, but I found it amusing.

Tried it for while, works with GrapheneOS and Android Auto well enough.

What I absolutely can’t stand is the routing. It once tried to send me through residential Oakland on some Manhattan-grade staircase labyrinth instead of just taking normal streets.

EVE to me is an odd duck because it's an explicitly cutthroat universe - many people log on expecting monsters. There are a lot of supportive players out there among the trolls; where the knives of betrayal really come out is in the politics of player-controlled space.

What's weird is that I love EVE for that sort of nullsec drama and sociopathic players eating each other in crazed gambits, but a couple of matches in Overwatch competitive a decade ago put me off the idea of matchmaking lobbies altogether.