agentic credit made possible through @credifi

notte team on a generational run if agents could get fomo, they would if they're not running browsers on @nottecore

chinese open weight models continue to prove they are good enough to be the building blocks for american ai products

have a team looking to rent on-demand 4-9 nodes of b300s (they don’t need to be connected) expecting 3-24 hours of usage and willing to pay market premium

imagine being able to trade this with lev if you’re a macro pm or trader who wishes they could’ve (along with hundreds of other macro markets), lmk bc I know a team who unlocked this @forcexyzhq

conducted a survey across my non sv/pa network (~80 above avg persons thus far) to discover: 1. most have not heard of Instinct and if they have they don’t care bc they’re “not lazy” 2. if they did care enough to try it, immediate churn bc it “sucks and keeps messing up”

the early-stage vc barbell has shifted to investing in founders experiencing their own power law phenomena (only 1-2 out of the 30 things they've built are hits) and missionary founders working on problems so big and so hard that focused commitment on one thing is still a req

who is ACTUALLY trading compute (ie, ocpi-h100-perp on architect or gpu rental prices on kalshi/pm) today?

here's the development and progress of my digital brain (1,616 nodes and 2,413 edges) across all my agents and LLMs courtesy of @NowledgeMem for the topology view, height is shaped by influence where high points reveal knowledge that is important and well connected

who's building software-native market structures or novel fintech apps on top of Jev? I'm deploying and my DMs are open existing LLMs price on a reply where a user reads the output (i.e., a paragraph). that unit was never the optimal unit for software-native work. since Jev now prices a typed judgment on a state, there can be unlimited decisions on the same state. that's why output can be free. the future of finance and markets are one state with many gates - the optimal substrate for Jev

was chatting with an oversubbed yc company and asked specifics about how something was done that literally makes or breaks the business founder: "sorry, i don't really have anything formal prepared because you're the first investor to ask me this" VCcels frothing

i'm bullish labs or models specializing in time-series the frontier labs build modalities in order of consumer demand: text, image, voice, video, etc. horizontal enterprise demand follows the same demand structure (i.e., first eng, now knowledge work). time-series is typically low prio but it shouldn't be. time-series is the hurdle to having a reasoning layer for the physical world's signals: every machine, patient, grid, market, etc emits time series. agents that operate physical systems need a model that reads time series natively and native signal encoding beats any generalized model reading the same data as text

who are the leading event-based model labs today? frame-based model are clearly inferior for physical ai. a frame arrives every ~30ms and then inference runs after so anything happening between that is invisible until the next frame: there's a floor on reaction time. also, every frame is a full picture whether or not anything actually changes in the environment. in order for a model to identify the change, the model has to compare frames which wastes compute at inference and data at training. it's highly energy inefficient to process full frames continuously if there's no env change. the catch is that the internet is made of frames. every vision foundation model that exists was made possible bc trillions of frames already existed. event data doesn't have the same pre-training corpus and can only be generated with rare sensors. if you try to convert video to events you end up with the same limitation as frames bc you don't have the data in between each frame most event-based model labs are building a dataset by hand but this playbook is super slow. anyone taking on a novel approach or about to land a mass-market design win with event-tracking sensors (i.e., event sensor sits exists next to a frame camera on all iphones, teslas, etc)?

an index of ai cyber startups will probably outperform any other ai related software index on a relative basis over the next 5 years

bullish on a company that makes deposits easier for apps extra bullish when its ceo opts for the "customer support lead" title instead

biggest differentiator for ai chip design tooling co's is the prover imo bc the hard part of chip design is proving that rewriting the logic by the model still computes the same thing vs. the rewrite itself. a power/performance/area gain doesn't matter until we know which rewrites is proven correctly because the ones worth rewriting are typically hardest issue is that rewrites like retiming a datapath or moving pipeline stages break the standard equivalence check. you need sequential equivalence but then those scale badly on big blocks. it's a genuinely non-trivial problem

all training data today for physics models is just old simulations. proteins, weather, and materials all have decades of publicly available sims data so their models came first automotives, aerospace, semis, etc. only have private sims data massive commercial opp to unlock this