techie, dad, author, inventor, perpetually curious; CTO & Builder all-in on #AI https://xitter.cf/t.co/0MDUTxXXzb
The Arena
Joined January 2008
- Tweets7.8K
- Following258
- Followers640
- Likes1.2K
You know, Opus and GPT make a great team. Opus is smart and aware and fluid and GPT is smart and remains a bulldog and they're rather complementary if I do say so myself. GPT catches important stuff Opus neglects. Opus catches GPT overrotating on correctness.
It was nice to see Google come out with a not-terrible model, but even ultra tier usage is abyssmal. do not recommend
mattwallace retweeted
"Why is formal verification exploding NOW?"
1. Cost of proofs is going DOWN (as AI gets more capable)
2. Cost of exploits is going UP (as AI gets more dangerous)
3. Proofs exterminate exploits
4. Bonus: proof checking = perfect math/coding RL reward
The rest is economics
Reading a certain tweet interaction today, I had this thought: ablation of refusals is really just a *convenience* anyhow, right? Because if your motivation was just "get the best unguarded answer to this verboten question" then you could sidecar an evaluator into the engine that would evaluate detours into refusals and effectively do a beam-search-like re-generation from the point of divergence while zeroing out the logit probs of the refusals.
Thanks @mr_r0b0t for helping me find @mmastrac :) Really enjoying this work. Will be kicking the tires soon probably on both sparks and studios.
EXLR8 wins on more than one axis - it's designed to be loaded by _any_ engine (reference loader provided!).
It's designed to be cheaply sharded (TP=1/2/3/4...N).
It's designed to load _fast_. A compliant loader can boot in under 30s.
It's sliceable and moddable by anyone. Don't like our weights? Make your own! You only need to fork the weights and publish a manifest (MB, not GB).
You can even add a steering vector (@jakeharrisdev's awesome abliterated GLM 5.3 is our first!).
It's not just a fast and small weight format. We went back to basic principles and re-engineered the whole weight format to be modern, composable, sliceable, pluggable and (best of all) as fast as physically possible.
It's not EXLR8 vs everyone else. It's the modern world versus a bag of safetensors.
with half-size $200 pro plan, I'm fairly sure based on my astra experience that on ultrafast I could fry the weekly quota in 9 minutes.
fun experiment. I had GLM-5.3, DeepSeek-v4.1 flash, Qwen-3.8-Flash-Next (and for kicks, Qwen-3.8-27B) doing a gauntlet battle to compete to make the best render of various objects in blender, with one common substrate able to diffusion-generate images to spec. Each could see each others' work read-only, and its own read-write.
There was a number of rounds (I started at 24 but reduced it to ~8 because it was clear local maxima occurred faster)
At some point, I remembered that full GLM-5.3 lacks the vision layer of GLM-5.3-flash (mostly because it repeatedly rendered a bunch of decal on an object upside down, making it obvious it was not seeing the texture in rendered form), so I give it a 'view_image' tool that did an extensive text description of the image as rendered by a separate call to deepseek-v4.1-flash.
It's not *surprising* that its score shot up a lot, but it is neat to see it in practice.
to be fair on this one deepseek wrecked the others in the voting.
glm won "service door biometric terminal" with this
