Roman

AI engineer, scientist, & biohacker shipping stuff and sharing ideas
Lead-author chemistry research + MSCS, ex genetics PhD

[Roman @decorrelat3d]

Built a Legends of Runeterra deck/format tool.

It just hit ~110k requests in the last 36 hours after the community started using it.

→ https://t.co/4F4Jsu3fuN

Made it for the new Region Expressions format. Format rules, deck validation, and related tools in one place.

Shipping real tools > talking about shipping.

If it grows enough, custom domain and polish coming soon!

[Roman @decorrelat3d]

@ScienceWDrDoug @XFreeze ive been using it religiously and i have never had them run my quota dry, you must be doing something wrong

[Roman @decorrelat3d]

@Pask218337 @beffjezos @elonmusk @grok 99.99th percentile is 0.01%

[Roman @decorrelat3d]

@gabriel1 need someone to chill with?

[Roman @decorrelat3d]

@novita_labs @huggingface how are they pricing the 1M context in practice? is it the full rate on every token or do they have some kind of tiered thing once you go past 128k

[Roman @decorrelat3d]

@KislayParashar1 how much would a 4080 go for? anything?

[Roman @decorrelat3d]

@0xSero did it lead to material improvements in QoL? where is your valence baseline after all of this?

[Roman @decorrelat3d]

@yacineMTB never forget

[Roman @decorrelat3d]

@testingcatalog i think the real price spread between cheapest and most expensive frontier models for these enterprise tasks is gonna drop under 8x by end of year but i could easily be wrong, e.g. if the new reasoning configs from the labs keep adding weird hidden inference costs

[Roman @decorrelat3d]

@MiaAI_lab 6B vs 18B is gonna win on speed no matter what the benchmarks say once you put both on the same 2x DGX setup, no?

[Roman @decorrelat3d]

@tonbistudio i love this, at the current grok 4.6 price you can let it run four of those 6 hour experiments for like $12. it's really great, it's no longer a meme to just leave this stuff running overnight

[Roman @decorrelat3d]

@jayair it wasn't really a stealth launch though, they'd been teasing the weights drop for like two weeks on their discord. the actual surprise was the price more than the model itself imo

[Roman @decorrelat3d]

@TheAhmadOsman imo these models are still miles behind on anything that needs precise control of experimental parameters or even basic reagent stoichiometry. the local part is nice but the intelligence isn't there yet

[Roman @decorrelat3d]

@josipK I think right now it's just decoder-only models with no constrained decoding so they keep hallucinating km values. physics-informed would be better but i haven't seen anything that actually works for assay planning yet. still early

[Roman @decorrelat3d]

i think within 18 months most routine enzyme assays will be 80% planned and evaluated by models with the human mostly just loading plates and hitting go but i could easily be wrong about the timeline if the current models keep hallucinating michaelis constants the way they do now

[Roman @decorrelat3d]

@rishabhvision @kimmonismus @synthwavedd 💯💯💯

[Roman @decorrelat3d]

@AndrewCurran_ gpt-astra just wrote the kernels for models it wasn't trained on...

[Roman @decorrelat3d]

@TheAhmadOsman this isn't happening

[Roman @decorrelat3d]

@FrankNoeBerlin @MSFTResearch this seems more like an engineering gig than actual science. free energy calculations at that scale still don't converge without heroic amounts of sampling and nobody has fixed the force field problem

[Roman @decorrelat3d]

@ivanburazin yeah this makes sense. the thing they left out is that cpu inference starts looking a lot better once you have models that are actually good at planning their own tool use instead of needing to be babied through every step with massive context

[Roman @decorrelat3d]

@adaptyvbio @Anthropic $40M series A is great, but the important number is what a single automated design-test cycle costs imo. if it's still four figures per protein nobody is gonna be running these loops at scale