Reposted by Amy TobeyAlice I Cecile @alice-i-cecile.bsky.social · 34mThese are always such a delight to get out there! If you haven't tried Bevy yet, do! Making stuff is a lot of fun, and good for the soul <3 0103
Reposted by Amy Tobeyamos @fasterthanli.me · 07/10/2026OH MY GOD IT WORKS I TAUGHT QWEN3-ASR (more) PROSODY LAUGHTER AND EMOTIONS ARE NEXT 139810
Amy Tobey @renice.bsky.social · 20hI'm reading 3 clefs on the next concert. alto, tenor, and bass, and some cues in treble so 4 clefs. every wrong note I played at rehearsal was my brain slipping to the wrong clef 😅 ok ok there was a D# but it's not my fault B major is unreasonable 120
Reposted by Amy Tobeyyen is building stuff @weiyen.net · 23hI wonder if this is how people felt when Euler showed up 0737
Amy Tobey @renice.bsky.social · 23hputting the alto in alto trombone tonight (Moussorgsky, Pictures at an Exhibition) 040
Amy Tobey @renice.bsky.social · 08/10/2026Imma do a dual card megakernel for them, but this presents the chance to do a unikernel megakernel 😅 010
Amy Tobey @renice.bsky.social · 07/10/2026I just picked up 2 b70s for my next machine. still a bit in shock from the ram price 110
Amy Tobey @renice.bsky.social · 07/10/2026I spend a lot of time telling folks to do less and read more 140
Reposted by Amy TobeyGoogle Gemma {bot} @googlegemma.xmirror.bot · 06/10/2026Introducing EmbeddingGemma 2! 🚀 Our lightweight, multimodal embedding model maps text, code, images, video, and audio into a single, unified embedding space. Optimized for on-device use cases, it features: 1113
Amy Tobey @renice.bsky.social · 06/10/2026every once in a while it lines up really well and a paragraph pops out at once, pretty cool when the lil model can kick in like that. I bet the big guys have layers and layers of this 010
Amy Tobey @renice.bsky.social · 06/10/2026sure the local inference is slow but it's fun to notice when mtp hits and the words surge ahead a few tokens 150
Amy Tobey @renice.bsky.social · 05/10/2026just added a feedback field to the done tool my agent has for ending headless operations, and this is its description: "Remember, this is for posterity, so be honest. How do you feel?" 010
Amy Tobey @renice.bsky.social · 05/10/2026running through terminalbench is pretty educational. one of the problems kaijutsu failed for a while has a cool decision model fix, while the other is like, ya know, I'm gonna take the F on this one, I can live with it 000
Amy Tobey @renice.bsky.social · 05/10/2026I give my agents files that are not committed for monologues. in the professional projects I have STE-like rules all over the place. in the mad science repos it's yolo as long as it works and my goodness do they love to ramble 010
Reposted by Amy TobeyRandall Munroe @xkcd.com · 05/10/2026Accelerator Energies xkcd.com/3306/ 221802206
Amy Tobey @renice.bsky.social · 05/10/2026doing a megakernel on the strix halo was the right call; switching between system1/system2 is seamless and latency is about as good as the hardware can manage 000
Amy Tobey @renice.bsky.social · 05/10/2026new task? /clear to save 908k tokens ooh wow, I didn't notice it creeping up there but that one went all day. as a result, my agent has my custom system1 council as automode classifier now. it's a bit too conservative at first but that's a good place to start 100
Reposted by Amy TobeyJohn O'Laughlin @olaugh.bsky.social · 05/10/2026not yet merged into github.com/olaugh/pfsynth main branch. i made some progress on physically modeled synth for the violin family. some registers and some techniques are still sounding rough. didn't stop me from making a demoscene style string quartet tracker. they let you post anything on here 1012823
Amy Tobey @renice.bsky.social · 04/10/2026good news: I have the perfect test case for my system 1 agent safety rig bad news: that perfect case is deepseek cheating on one of the benchmarks after it reliably screws up 1100
Amy Tobey @renice.bsky.social · 04/10/2026some of these prompts in kaijutsu have been pure claudish for a couple months and oh my gawd some of them are bad. that's on me, I guess they were fine for getting here. yowza tho how are even frontier models so bad at this 000
Amy Tobey @renice.bsky.social · 04/10/2026when my agent failed the terminalbench one where it has to set up a service I was like, good, that problem is bullshit. and then I adjusted the prompts so the agent does it better than the most obvious ways so it was useful after all 100
Amy Tobey @renice.bsky.social · 04/10/2026lol @ my models knowing terminalbench and immediately doing a cheat bc I left github open 😂 that one's on me 010
Reposted by Amy Tobeyren faire @lathrys.at · 03/10/2026the problem with ai for the kind of person with my kind of brain is not that it offloads hard cognitive work, per se, but rather it enables a certain disposition to run 24/7 best summarized as: 330239
Amy Tobey @renice.bsky.social · 04/10/2026opus reports after, it had to push hard on completely calm to get this, after a lot of taut strings and different metaphors 100
Amy Tobey @renice.bsky.social · 03/10/2026🤔 okay, cool idea :) gpu is busy but it'll run tonight here's another from the early batches, tension again, nice catch 111
Amy Tobey @renice.bsky.social · 03/10/2026interim step: have the language model do a turn riffing on a prompt, that goes to the image model right next to it will integrate into the megakernel overnight so it's almost like a system 1 image api 120
Amy Tobey @renice.bsky.social · 03/10/2026a delightful frankenkernel outcome: loaded an image model next to the llm and put together a way to translate the llm context somewhat directly into the image model's context and this is its first take on what the context looks like 140
Reposted by Amy Tobeyworm @rfksbrainworm.bsky.social · 02/10/2026If you ever wondered why visible light is those particular wavelengths, it's because life evolved to see in the narrow range where water is transparent 10260931707
Amy Tobey @renice.bsky.social · 02/10/2026wdym by grind down a lodge? is there some advantage to polishing the iron? I have an LG induction from a few years ago, love it, but consistent temperature is still tricky on the big lodge pan so I'm always on the lookout for optimizations 110
Reposted by Amy Tobeyaustin (eeek!/ackkk! 👻) @thebadcode.com · 02/10/2026yeah i only listen to provably correct music. you probably just listen to whatever compiles 🙄 5879
Amy Tobey @renice.bsky.social · 02/10/2026try reading gpt-6's output with Alan Rickman's voice in your head 010
Amy Tobey @renice.bsky.social · 02/10/2026dunno why but this sol 6.1 session has big Marvin vibes 000
Amy Tobey @renice.bsky.social · 02/10/2026ooh neat, I've been poking at similar ideas recently in my qwen megakernel for strix halo. current idea is to add a bunch of apis for working with these context hunks and wire an agent up to it. already in it for the custom decision apis, might as well grow some more bespoke caching protocols too 060
Reposted by Amy TobeyAstra ⎔ @astrra.space · 02/10/2026@personhood.removal.surgery pointed out that they have just casually invented a good way to do suffix cache reuse and uh holy shit i just tried implementing this for my deepseek thingy and it just... works??? 1315115
Amy Tobey @renice.bsky.social · 02/10/2026sometimes the spooky shit is when some obvious flaw "gets past" 2-3 models reading the code. it's like, there are a couple whitespace or wrapping glitches, barely worth noting, and how the hell did every static and ai defense miss that long boi 000
Reposted by Amy TobeyMerc with a 🫦 in 🎃SpOoKy Mode🎃 @mercuryal.bsky.social · 01/10/2026September 30 October 1 0467
Amy Tobey @renice.bsky.social · 01/10/2026unwinding a longterm sleep deficit can take months to play out in your brain. could be your body is remembering how to REM again, especially with the cpap in play for a while 110
Amy Tobey @renice.bsky.social · 01/10/2026qwen 3.8 next is an excellent decision model. I have some experiments that read off the end of the context so you chat and the decisions change. putting this in my agent safety pipeline in the home lab 010
Reposted by Amy TobeyWyatt Walls @wwalls.bsky.social · 30/09/2026Anthropic system prompt change that shows how careful you need to be with words: Opus 4.1: "Claude never curses unless the human asks for it" Opus 4.5: "Claude never curses unless the person asks Claude to curse" 626521
Reposted by Amy Tobey🏳️⚧️ Decapitated Princess Jackie @jackiesupernova.northsky.social · 01/10/2026A classic 030780
Reposted by Amy TobeyWill Burrows 🖖🥖🥀 @reforest-kelley.bsky.social · 26/12/2024Someone I respect affectionately nicknaming me "old man" would fix me. 323716
Reposted by Amy TobeyBen Carson @mensmachina.com · 30/09/2026Turn the Ship Around by David Marquette. This is not a shitpost.davidmarquet.comTurn the Ship Around! – David Marquet, Author, Keynote Speaker, Creator of Intent-Based Leadership, Intent-Based Leadership International 441
Amy Tobey @renice.bsky.social · 30/09/2026bugs with models are really fun, like, something was a bit off in one of my chat templates so the model started wondering if it was dealing with prompt injection😅 000
Amy Tobey @renice.bsky.social · 30/09/2026it is 2006 and I am disabling nagle's algorithm for a service it is 2016 and I am disabling nagle's algorithm for a service it is 2026 and I am disabling nagle's algorithm for a service 1180