this is a great deck by @a16z 1) token prices are down down down 2) COGS (memory, gpus, electricity) all up up up. biggest takeaway not clearly highlighted... this implies: 3) frontier lab gross margins are getting crushed!?

and and what if we used jev as judge to downselect the set even further! NEURIPS SPOTLIGHT HERE WE COME!!!!!

this is my wife taylor chaubard. she is not chinese.. does not speak chinese... and is not looking to "meet up" with you. (i hope) also she does not know how to tweet. is anyone else tired of china hacking everything without any repercussions? can we end this please? @nikitabier @X

Sorry for delay! Here is my work on zero order (ZO) optimization. We achieve SOTA for ZO methods on pretraining controlling for compute and parameters. ZO methods struggle to improve loss as model size grows because relative gradient variance increases linearly with the number of perturbed parameters, inhibiting large model training. All previous methods innovate on the optimizer, but adopt architectures designed for backprop. Instead, we design an architecture with ZO in mind, that caps the gradient variance as you increase model size. Introducing SOMA (Sharded Optimization Mixture of Assemblies). SOMA shards the model into tiny experts, and trains each expert independently, on disaggregated GPUs, without communicating gradients, activations or optimizer state during training. Inspired by the thalamus and cortical columns, each expert specializes on a subset of the train set based on a fixed router. At isocompute and isoparams, SOMA achieves lower loss vs. all monolithic ZO methods tested (e.g. EGGROLL, more perturbations w Vanilla SPSA, etc). ZO w SOMA continues to improve at larger model sizes. Additional inference benefits of this architecture, we can select top-k active experts over N trained experts to reduce inference compute by O(N/k), allowing us to tradeoff inference flops for accuracy wo retraining.

Next YC Paper Club alert: Topic: AI Safety When: Wednesday, 10.7, 5-9pm, MV Description: Since Coxon's very public resignation, AI safety has been the central discussion point now in public discourse. Conversations have been unscientific and have thus invoked fear without thorough analysis, structure or study. Many outside of the EA community have studied AI Safety for decades, carefully studying how AI can drive cars, guide airline pilots, etc. safely to ensure the AI community adopts and deploys automation safely and with standards, without government intervention. We plan to invite these stewards of the original AI Safety community to ground the conversation in scientific rigor once more. Please join us. To attend: events.ycombinator.com/yc-pa… Please DM if you have good ideas for speakers. @ycombinator

ya

astra down night of iclr submission. def not correlated.

trying to convince my professor of strange results before ICLR submission..

guys! for the last time... china does NOT steal IP from every western country, like google, ericsson, anth, oai, at&t, motorola, lockheed, apple, pfizer, merck, etc. all frivolous claims w no evidence. they dont have many spies in our government like swalwell's honeypot or mitch mcconnell's wife. they dont bribe influencers on social media to spin narrative. they dont influence elections with millions of fake accounts. their is no chinese backdoor in espressif or hauwei chips that send all your data to ccp. they didnt develop then leak COVID and then lie about it and pay WHO to also lie about it all classic conspiracy nonsense. they have never hacked all of our critical infra or US Siprnet, and never would. thats something a foe would do that hates america. why keep pushing these obviously false narratives. china's rise is pure china innovation and genius against western democracy laziness. ccp is our friend and its racist to say otherwise. they do not violate WTO regulation by subsidizing exports. they play fair. thats not the china i know. they believe in freedom of the press, freedom of speech, freedom of religion, and IP rights. this is why ALL of the world's smartest want to start their companies in china, not the US. chinese engineers have no wish to ever go to the US bc they already have such freedom in china. no need for exit bans. ask the manus founders. they can travel to US whenever they want. they just choose not to bc china is everything they need. BUT (on the other hand) we ALL know of the US government's coordinated IP theft of chinese companies IP for the benefit of US companies, and exit bans for dario, sama, ilya, and schulman and its disgusting. Therefore, we should continue giving ccp more power, more US GDP, share more IP, and more partnership. We def should not retaliate in any way bc they have been nothing but great partners. what we SHOULD do now that i think about it is ban american datacenters and AI development and american nuclear development, and just use china's AI, power, datacenters, and chips. give them all US data! (the CCP approves of this message)

dont listen to me. listen to bill gurley @bgurley piped.video/F9cO3-MLHOM?is=MM1y…

buy your shirt..

wtf.. launched a research idea last night w my autoresearcher.. woke up this AM to Astra hillclimbing the crap out of it... claude gave up after naive baseline, and said "its theoretically not possible, lets publish this negative result." product idea: claude should tool-call to astra for advice...

here is raw video of me asking him.. and his well thought out response:

what?? translation... dont be in the US and good at AI at the same time, or you will go to jail for 2 decades.. but.. if you are in china, training the same models, and releasing weights PUBLICLY.. (and funding my campaign)... that is completely fine.. won't comment on that at all. am I taking crazy pills? or is it just the socialists.

golf sucks.

anti-datacenter but mandate telework as default option... cant.. literally cant...

its quite amazing getting called far right AND far left on the same day.. i guess it wasnt a spectrum after all.. we have completed the circle!

iykyk

nice! ive been on this for awhile. piped.video/7xyarz2-LZQ?is=OJD-… ambien wishes it was this effective

we need to start a new revolution. why stop at datacenters. #bangolf buying 1M of these shirts. please DM me if you want one.

for details

Signup below for next YC Paper Club - "Harness Edition" Killer Lineup: - @sethkarten - Prime Agent @PrimeIntellect - @beyang / @sqs - @AmpCode - @josh__france - QM @ycombinator - @JonSaadFalcon - OpenJarvis - ? One more slot for speaker. Who should I ask? 8.26 5pm-9pm MV Also, fresh look for paper club! 100% credit goes to @eve_bouff! Aiming for: ... 2am, full team, in lab, tryna meet May 6, 2026 AoE deadline, paper not compiling-kinda vibe.. events.ycombinator.com/yc-pa…

small gift for yc paper club members for wednesday. stamped myself! ty @eve_bouff and @mattkang

best form of flattery.. great to see. hope one of these pops up in every major city! piped.video/watch?v=1TROrlh_…

FYI.. just did this analysis.. Every LLM is worse than random guessing on predicting 7 day returns of individual equities (last 12 mos).

always go in with the attitude "how hard can it be" - Jensen #YCStartupSchool

something big happening tomorrow. @ycombinator

must be.. flamingocompute.com is avail

FLOTILLA COMPUTE! SOMEONE DO THIS STARTUP! I WILL WRITE YOU A CHECK AND ADVISE YOU!

sure.... then post this for me when you get a chance. and say that it looks like someone in china.

lol what?? freedom of the press.. how does china feel about images like this? can you post this for me obvious ccp member?

maybe high has it fixed? but look dude. not fake. also maybe youre engagement farming

hmm no its not.

misguided? why does it think its claude.

agreed.. magnus carlson level psyops to get ppl to believe this is how you cool a gpu and write about it..

at frontend. not "coding". and it thinks its claude.

no it didnt beat fable in coding.. in frontend.. and they mass distilled claude. so much so it thinks its claude. this post is horribly misleading

winning? why does it think its claude?

why does it think its claude?

why does it think its claude?

dude it thinks its claude. why?

why this happening?

lol dude. it thinks its claude. why? this is an IQ test.

how do yu think it got good??

lol it thinks its claude dude. wonder why?

it thinks its claude. why?

why does kimi think its claude? mass distill attack?

it thinks its claude. mass distillation attacks are not "leading the AI race".

it thinks its claude... mass distillation isnt that innovative.

it thinks its claude dude.

kimi thinks its claude. ??

kimi doesnt know his name. actually thinks its claude. thats how hard they distilled it from ant.

Great YC Paper Club Kernels/Chips edition last night. Papers:     1    Stuart Sul (Stanford/Cursor) - ParallelKittens (arxiv.org/pdf/2511.13940)     2    Avanika Narayan and Jon Saad-Falcon (Stanford) - Intelligence Per Watt: Measuring Intelligence Efficiency of Local AI (arxiv.org/pdf/2511.07885)     3    Mark Saroufim (Core Automation) - To automate research we must automate systems (no paper)     4    Misha Smelyanskiy (Nvidia/NewCo) - Why AI Inference Needs Heterogeneous Hardware (no paper)     5    Brennan Shacklett (Stanford) - An Extensible, Data-Oriented Architecture for High-Performance, Many-World Simulation (madrona-engine.github.io/sha…) pic of @marksaroufim teaching CUDA! next one on robotics.. who should we invite to present?

explaining whats happening in AI these days to my non-AI friends... @HazyResearch

idk about everyone else, but fable sucks now.. this plot is 100% true.. 60% drop in debugging..

New paper coming soon.. teaser.. no transformer, no backprop, no problem! Zero Order CAN pretrain! very exciting.. stay tuned!

#s26 is underway! @ycombinator

Where is "frontier research" still happening? my 2c..

Looks like I got a new job! @Flock_Safety Cant wait to work with you on the transition @glangley