Tuf Ted

[Tuf Ted @srvtest]

@VictorTaelin @lookingforgames @ThePrimeagen I mean it helped me understand it as reward-hacking behavior and not just "LLMs are bad at understanding intent."

[Tuf Ted @srvtest]

Thanks, this helped me understand the reward-hackish behavior I've been seeing (as @ThePrimeagen put it: "I'm tired of getting everything I ask for and nothing I want"). My intuition is that the result of training is a gradient: on one end, a model with perfect general problem solving ability; on the other, a model overfit to solve only its training problems and nothing else. It seems a model can land somewhere within that gradient.

[Possum Reviews @ReviewsPossum]

Last night, I read the entirety of C.S. Lewis' The Screwtape Letters. It's a novel told in the form of letters written by a demon to another demon instructing him on ways to manipulate his "patient" to do evil.

This one quote sounded familiar.

[Tuf Ted @srvtest]

@ArtemisConsort I've always felt that this kind of "bad" thinking is a heuristic that protects against sophistry. Like we're supposed to get inoculated with our beliefs when we're young (with the fact that those beliefs bore us successfully serving as the "proof").

[Tuf Ted @srvtest]

@EpoptesG2_7 @ArtemisConsort People are relatively good at figuring stuff out when they have the benefit of a short feedback loop including a hard check from reality. Watch out when there is a long feedback loop (long enough to exclude iterative refinement) or the feedback comes from only one's peers.

[Tuf Ted @srvtest]

@UnderSneege @edwest The progressive mind seeks to convert everyone into a fellow comrade who always does the "right thing." Hence, there is no need to consider 2nd+ order effects as they will be mitigated with more progressivism, more people doing the "right thing," more "if everyone would just..."

[Tuf Ted @srvtest]

@dystopiangf A lot of our important social patterns are the products of extensive trial and error processes a la natural selection. That means they work, but people don't necessarily know why. But even the worst ideas people produce always come with "knowing" how they work due to their origin

[Tuf Ted @srvtest]

@uubzu The truth is itself ultimately just a heuristic if you consider the purpose of a belief to be to cause a beneficial change in one's behavior.

[Tuf Ted @srvtest]

@uubzu The truth is itself ultimately just a heuristic if you consider the purpose of a belief to be to cause a beneficial change in one's behavior.

[AIPAC 🇺🇸🇮🇱 @AIPAC]

Ed Gallrein’s victory in KY and Clay Fuller’s win in GA ensures two outspoken pro-Israel voices are positioned to fill seats previously held by outspoken detractors, Thomas Massie and Marjorie Taylor Greene.

Our community was proud to help pro-Israel candidates win these races.

[Tuf Ted @srvtest]

@ArtemisConsort @johnbutinfosec I actually encountered this reality when I needed the underlying hashes to double as u128 (or bigger) uuid's (i.e. so they could be generated without collisions). That necessitated extra cpu work compared to the u64 hashes normally used in Rust.

[Tuf Ted @srvtest]

@realtj @LundukeJournal bootstrapped builds address that concern by keeping the entire chain human-readable, starting with a tiny hex-only but still human-readable compiler called hex0 (of course you still have to worry about the build environment I suppose...)

[Tuf Ted @srvtest]

@KCG3D @theo Defense in depth. We're in a "catching-up" period that will be temporary

[Tuf Ted @srvtest]

I've been running mine with the larger UD XL model, more precise q8 context, ngram speculation, lower 170K context:

llama-server --model ~/ai/models/Qwen3.6-27B-UD-Q4_K_XL.gguf --jinja -np 1 -ngl 99 --temp 0.6 --top-p 0.95 --top-k 20 --min-p 0.05 --ctx-size 170000 --cache-type-k q8_0 --cache-type-v q8_0 --host 0.0.0.0 --reasoning on -b 2048 --spec-type ngram-mod --spec-ngram-mod-n-match 4 --spec-ngram-mod-n-min 4 --spec-ngram-mod-n-max 16

[Tuf Ted @srvtest]

I agree, and the right extensions make it notably more capable. I'll have to look into the ones you're using to see if there's any good ones. Looks like you have a web search, but do you have something to ask you questions and interactively plan?
pi install npm:@juicesharp/rpiv-ask-user-question
pi install npm:@plannotator/pi-extension (use /plannotator to enable)

Also have you tried the mixed-precision gguf K quants (presumably slightly slower but better)

[Tuf Ted @srvtest]

@grok The github connector is a great feature. I wish we could specify one repo through the UI.

[Tuf Ted @srvtest]

@bubblegumbword This is a book about punctuation. It is highly entertaining. My eyes were glued to the pages when I read it. I recommend it to everyone, especially you.

[Tuf Ted @srvtest]

Technology has advanced faster than our culture has had time to adapt to it. In fact, we live in what could be viewed as a toxic environment, subject to a constant and constantly-evolving barrage of products that are designed to exploit our brain's reward mechanisms. No matter who you are, you can be assured that there is a product that is just the right match for exploiting your unique brain. On the other hand, we also have access to incredible resources and technology such that more possibilities than ever can be realized. Success in this environment requires someone to have a strong guide to what is valuable, lest they get lost in the ever-more-abundant temptations of the day or leverage great means towards something that is not fruitful.

[wanye @xwanyex]

I’m sorry, I know this bums a lot of you out, because you’ve built your personality on being to pro-market, pro-technology guy, but technology is just very clearly making us less happy

[Tuf Ted @srvtest]

@OkamaKichigai @esrtweet https://t.co/0S8zEXyOHY

[Tuf Ted @srvtest]

@OkamaKichigai @esrtweet The May 2026 Nixpkgs GitHub Issue #516544 suggests otherwise. A trusted contributor unilaterally inserted a magic string in maintainer-list.nix intended to sabotage AI agents. Some other members are defending him