@wildcatfi fine, we'll do it ourselves
Joined November 2021
- Tweets14.5K
- Following690
- Followers1.2K
- Likes64.8K
the entire tax software industry is an extractive middle management cabal that protects each other.
Awaken has been amazing since day one and it keeps getting better. you should use it.
and you should fire tax advisors that refuse it.
i'm usually just busy, and genuinely surprised when my kids show up with a sparkler stuffed horizontally through a cupcake
i really want to like Liquid models. i'll take them for a spin, maybe tune a couple, but the license requirements really put a damper on using it for anything that might scale
Introducing 𝙻𝙵𝙼𝟹: our new generation of Liquid Foundation Models. With a new tri-hybrid architecture and new multimodal encoders, the LFM3 family is the ideal model suite for specialized intelligence.
> 6 model instances, across text, audio, vision, and omni modalities, generative and decision models.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱: text, 256k context window.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱: vision-language model with video-input support for spatial understanding
> 𝙻𝙵𝙼𝟹-𝙰𝚞𝚍𝚒𝚘-𝟸𝙱: tool calls, voice chat, reasoning in 18 languages.
> 𝙻𝙵𝙼𝟹-𝙾𝚖𝚗𝚒-𝟸𝙱: an omni model combines vision, audio, and language.
> 𝚍𝟷-𝟸𝙱 and 𝚍𝟷-𝙾𝚖𝚗𝚒-𝟸𝙱: decision models with typed outputs and calibrated probabilities in a single pass. Omni adds audio input.
> fine-tune on 1 GPU.
> a new tri-hybrid architecture
> architecture optimized for 14 processor classes across GPUs, CPUs, and NPUs.
> 𝙻𝙵𝙼𝟹-𝚅𝙻-𝟸𝙱 is a beast. It performs competitively with Gemma 4 12B (6x larger), and Gemini 3.5 Flash Lite, and GPT-5.6 Luna, with 1.9B parameters.
> 𝙻𝙵𝙼𝟹-𝟷.𝟽𝙱 is #1 on the Pareto frontier of device intelligence 𝚙𝚒𝚙𝚎𝚝𝚝𝚎.𝚕𝚒𝚚𝚞𝚒𝚍.𝚊𝚒
Every modality for specialized intelligence. Pick the one that fits your use case, customize it, and run it on-device.
1/ 🧵
genuinely irritated my prefill is still so bad. taking a break from the competition. can't compete with people spending 6 digits on compute
today is proofreading and frontend work, and then deep-diving @RareSkills_io zk lessons while my rig autotunes ds 4.1f 1017 on itself.
people want to hate ai. so they will hate it, regardless of the result. then they will get lost in their private subculture while the rest of ius move on
they are attempting to add more ram to a system. the old chips were removed, and they've connected wires to an external set of chips.
ram is fast. really fast. the electrical signal travel times are measures in nanonseconds. 1ns can generally travel 6-9 inches on copper.
ram needs the signals to arrive very near the same time, consistently, across every parallel wire.
pcbs go to great lengths to make sure this happens. handcut and handsoldered wires dont have the same precision.
so this person made the timing less precise by doing this, and they also added actual travel distance, which makes the error margins much smaller. it's unlikely to work, but it's possible.
then there's the sheer persistence of this; they had to cut and solder every piece of wire here. ddr5 is 288 pins, so that's 288 individual wires they cut and soldered on both ends. impressive effort.
I'm not going to migrate just because one is actively maintained and one is not. I don't need any new features, and gitea works well with outline and infisical, so no reason for me to create unnecessary work.
sorry, it's late and my brain was tickled. there's probably a more fun direction to go with this but I was doing dishes