Ishaan Lal

biotech - doing my masters in systems biology & working on DL approaches in host pathogen ppi models.

[Ishaan Lal @enryuoe]

@ShiggyShaman @sudoingX mmhm this is what a human response on Twitter would look like exactly!

[Ishaan Lal @enryuoe]

@METR_Evals Isnt it kinda weird you throw around some 10% unmonitored usage numbers and don't have any mention of this in your reports from the 14th, I thought the point of this is to be transparent+ you give a 10% estimate and instantly say well oai may redact stuff, like what's the point🙏

[Ishaan Lal @enryuoe]

@blueemi99 you got voxtral

[Ishaan Lal @enryuoe]

@cmsinvests Bro just know that cancer is a better algorithm currently than compared to our drug development pipelines, even what we consider robust treatments like CAR-T, some methods lead to post treatment issues and generally cancer is better at finding loopholes against our methods..

[Ishaan Lal @enryuoe]

@sudoingX yo dude help out the guy doing it on his 3090 I think he's stuck at like 4tok/s

[Ishaan Lal @enryuoe]

@ItsmeAjayKV hopefully it's the 27B 🙏

[Ishaan Lal @enryuoe]

@michaelcostea @mdrakestapleton @LDT0545 How is a dgx spark dead hardware (blackwell) bruh this would've made sense if I said p40/v100s , even like when openAI starts having agent issues they also load up their dgx's and run glm 5.3 😂
Plus once you start using your brain again bud you won't need frontier AI..

[Ishaan Lal @enryuoe]

@AestherML I had looked into 2 core server style boards, low budget build yk - xeon parts, ecc ram, Intel arcs/nvidia p40 it's too much work doing 2 way cpu + 2 memory channels to multigpu optimisation esp model by model 😭

[Ishaan Lal @enryuoe]

@mdrakestapleton @LDT0545 I mean yeah for like 7200/year you can buy a dgx a year, by the 4th you're set and anything after that is j for fun. And this saves you money 😭 (1200 a year if we assume dgx spark at like 5/6k$ with a respectable config)

[Ishaan Lal @enryuoe]

@daradoescode Contributions to the integer multiplication bound project..

[Ishaan Lal @enryuoe]

@QualiaQuanta This is why we need to be able to audit their detection models and understand the data they trained them on cause it's mostly just cross contamination from the fact models are trained on papers and other knowledge and thereafter ai detectors are trained on ai outputs..

[Ishaan Lal @enryuoe]

@alex_verem that's right but also with open weight, I think you can reverse engineer the data from the layers with some approaches still right,,, and do stuff to change model responses + fine tune,, but ya more imp than the arch/weights is actually the curated train data, which isn't OA..

[Ishaan Lal @enryuoe]

@leecronin I wonder in the coming years whether we would rediscover fundamental stuff for the purpose of more accuracy..

[Ishaan Lal @enryuoe]

@kamilbaranek_ plus no restrictions on task type..

[Ishaan Lal @enryuoe]

@BroadsideCode optimised wow

[Ishaan Lal @enryuoe]

@ronisarkar_exe task quality ig but also that could depend on how people code,, I do find ds v4.1 quite good but I'd say comp to sonnet 5.5 nottt better...

[Ishaan Lal @enryuoe]

@kamilbaranek_ Cause they made a pitch to apple w voxtral and all mistral large 4 was good for was reasoning over the conversation 🤠🤠

[Ishaan Lal @enryuoe]

@BroadsideCode Random, but if I send you hate, how much does it cost you 😭😂😂 (not gonna thoo)

[Ishaan Lal @enryuoe]

@sudoingX Bandwidth...