A Vignette of Twitter Posts

How do people find their life’s work? The problems I care about feel extremely grand like the welfare of people globally. You could argue it’s too ambitious given current capital/resources Good founders have market fit. How do you build conviction in a space early on?

the best projects are at the end of naive exploration. started without the expectation of solving grand problems. (though this doesn’t preclude you from believing you’re capable of solving them) that is to say, be open to problems, but don’t seek them out. i lowk treat it like looking for a relationship

dfw on the constant state of disquiet we find ourselves in. if you spend much of the day absorbed/in pursuit, it becomes easy to distract and exhaust yourself until sleep takes you to the next day. the disquiet compounds, it eventually steals all your cognitive overhead.

QC

@QiaochuYuan

spend one single day not looking at a screen. i know this is easier said than done but just for one day do literally anything else. read a physical book. go for a walk in the park. do jumping jacks in your room. stare at the wall. listen to an album from your teenage years there

there are essentially two modes or sides of a spectrum concerning actions you take when running a company. one is where you figure out "ok, this needs to get done" and you just don't sleep for 2 weeks and get it done. the other is "hmm, structurally is there a way i can delete this complexity or make it less dumb?" like there is a huge space of possible approaches, *one* of them must be high leverage. what are the real requirements, what must be true to do something. think from first principles, gather information. much like the exploitation/exploration trade off found in multi-arm bandits. the latter is what ingenue is, what creativity is used for.

when i see founders spiral or burn out it's usually because the disquiet compounded and all of sudden there is no more urgent work or distractions. or they figured out the stuff they were distracting themselves with is low leverage. they lost the ingenue or creativity that catalyzed the company initially. companies only really fail when the founders decide to give up.

also on the multi-arm bandit point: i find startups mostly have to do with bet sizing. like you are betting your team's time/attention and cognitive overhead on things which may or may not move the business forward. the cheap and convex things are often missed and the psychology is such that we bond the most likely outcomes. and the most important/scarce resource startups have is the cognitive overhead of the team.

Solve this in under 5 minutes and I’ll offer you $500k/year in cash plus several million in equity I'm building a Computer-Use team, goal is to use computers better than humans No experience or PhD needed Instructions: 1. Solve all 30 challenges on this website in under 5 minutes: https://serene-frangipane-7fd25b.netlify.app/ 2. Feel free to use any tools or vibe code it. Provide us a zip folder with instructions on how to run the agent and reproduce your results, as well your run statistics 3. The agent should be able to solve all the challenges, use browser, and provide overall metrics around time taken, token usage and token cost. Your agent must solve this challenge in under 5 minutes Email your response: agents@brettadcock.com If you have any questions about this challenge, feel free to email us

few thoughts on this after playing with it for 5 minutes: - brute force and randomly clicking all of the buttons is a very good strategy here to dismiss all modals, ie it doesn't really punish making or clicking the wrong thing enough - opus 4.5 will be way to slow to solve this naively - best strategy will probably involve - non visual rendering/ingestion of DOM - parallel agents mapping action space (interacting with diff parts of DOM) and creating centralized skills/memory to solve for certain taxonomies of modals - gemini 3 flash - a main agent (you could do opus 4.5) that constantly runs in the background that makes sense of action space + skills to actually solve each level most likely the spirit of the challenge is to build something that can generalize and not brute force action space and be specially prompted. generally i think for each level, map out action space with parallel agents + create skills for modals with fast model, then have main loop solve each level with data at hand will be a winning solution. or you can also build your model(s) for this but it would probs be harder than just doing this.

yc is just funding anything these days 😭

it's bet on minecraft becoming roblox and a sort of "metaverse" type of thing idk how good a bet on minecraft as the next big social/gaming layer or platform is but to me this is essentially the same type of company as @spawn or @moonlake but "doing things that don't scale"

Notes on MARTY SUPREME: - There are only two important scenes in the film: the first scene, when Marty impregnates his girlfriend in the back of the shoe shop where he reluctantly works, and the final scene, where he meets his child for the first time. The entire prolonged midsection of the film: the ping pong games, the elaborate heists, the swindling of Abel Ferrara, the crushed dog in the bathtub, immolated gas station, hustling of the kids in the bowling alley (plus Gwyneth Paltrow and Kevin O’Leary), shootouts, even Marty’s triumph against Endo, ONLY exist so their importance/significance can be negated by the ending shot of the film, when Marty looks into the eyes of his kid and tears up and clearly realizes “this is all that really matters”. - I’m not sure if it was intended this way, but for that reason the movie succeeds as a work of pronatalism (and the 1952 postwar era does coincide with the baby boom). - But it’s funny that, for a film whose marketing celebrated audacity, big swings, unapologetic attention-seeking (Chalamet “SCHWAP”ing on Zoom, the orange blimps, the line for jackets in Soho, the flanks of orange ball-headed guards, Timmy shivering atop the Sphere) the message of the film seemed to be the exact opposite of what all that promotion espoused: actually, none of your dreams matter, and everything will recede and be reframed when your meet your progeny for the first time and your real purpose becomes clear. - In other words, the point of all that marketing was to get you to see the movie. But the point of the movie was to tell you that all of that marketing was pointless (or at least, the aspects of the film that the marketing refers to are pointless). Marty never plays a game with the orange balls, in fact he dumps them all outside the apartment window of a fleeting business partner during one of his many adult tantrums throughout the film) - I’m not sure if that’s genius or trolling. - Other sports/athleticism/“big dreams” movies (e.g. Rocky, The Wrestler, even I Tonya and Dodgeball and Black Swan for that matter) actually use each sequence of the game/match/performance to advance the plot/frame the character’s state. They’re all shot differently: some in slow motion, some with close-ups and focus on technique, others are montage sequences. But in Marty Supreme, all of the ping pong games were shot in pretty much the same way. Again a kind of “none of this really matters in the end” statement? - tl;dr: Marty Supreme is a 2h29m movie where only the first and last five minutes matter, but those minutes only matter *because* the rest of the film underscores how nothing else matters except for those moments. - So I guess it works as a movie

i think your reading is coherent, but i came away with a sort of opposing interpretation. > "[the other scenes] ONLY exist so their importance/significance can be negated by the ending shot of the film, when Marty looks into the eyes of his kid and tears up and clearly realizes 'this is all that really matters'." i do not think he has a sudden change of heart and realizes that "this is all that really matters". if he failed he wouldn't have cried when he saw his son, he probably wouldn't even see him. he would keep pushing the boulder up the hill. he cries because he made it out. because the weight of greatness is finally off his shoulders. because he GETS to be there. he has his cake and eats it too. him in that moment seeing his child, it's the moment of winning, crystallized. it's *validating* and sanctifying all his previous faustian toils. as for the first, scene, it's profane, ordinary. it's an innocuous throwaway. it represents the vérité and quotidian mires as experienced by someone with marty supreme pathologies. you can say, the rest of the movie is marred by this scene. and in a narrative sense it would not make sense without it. but this is like claiming shoelaces are the most important part of running a marathon because without them, your shoes fall off. to him, it's like tying his shoes. he does it without care because he is marty mauser. because regardless of whatever happens he'll figure it out. it's the first introduction we get to his youthful arrogance. > "actually, none of your dreams matter, and everything will recede and be reframed when your meet your progeny for the first time and your real purpose becomes clear." i think structurally your correct that the ending is important, but that is not at all how i interpreted it. in my view, the movie says that often there are cardinal moments on the quest to greatness. where you give up a normative life as penance for your ultimate goal. and that, locally, they may seem existential but, in retrospect, they will be anything but. he didn't compromise greatness AND he still met his progeny. when you win, of course, the pathology and tension dissolves somewhat. and new games and sources of satisfaction emerge. crucially, **this does not discredit playing the game in the first place**. in other words, the dreams of your youth still matter in old age. without it, you'll be an old man with unresolved tension deep within your soul. there will always be uncertainty about the counterfactual: what if you sacrificed everything to win? **did you have what it takes to be great?** the only way to quell any questioning about it would be to win. any new purpose after, is not any more "real" than what came before. the marketing (and movie) glorifies this sacrifice, the indulgence in a brash sort of youthful ego. the purpose of the movie is to provide a window to what happens when winning is existential. when it does not become a question of "will" but "survival" and tenacity. the very pathology marty has is in making dreams existential. it's a question of ontological value. his mental model precludes him from existing separately to his dreams. it implores us to not act so dogmatically and manichean when ascribing moral value to actions taken withing the complexity of everyday life. it tells us that sacrifice is not irredeemable, that we can still be great and hold our progeny. it tells us not to give up our youth, to pursue things at cost, to believe in good endings in the most dire of circumstances. greatness is messy. life is messy. i love this part of the movie the most. if you talk to people who have experienced and persevered through extreme hardship the stories they'll tell are more absurd than any movie. too long to explain and wrap in a bow for an audience. too complex for the audience to come away with satisfied. the pacing of the movie was excellent, it matches the non-linearity of memory, it covered such a gamut of the messy features ever present in real life. something movies oft forget when peddling a single dimensional message. marty supreme tells us that the path to greatness is hard and non linear, and to do it anyways.

The best marketer in the world @im_roy_lee just revealed they're doing tiktok clipping We called this strat 8 months ago and ppl just laughed (read QTs/replies). Meanwhile others were exploiting tf out of it The next @cluely is currently in the discord trenches meeting the sewer kids. Finding the underpriced channels and exploiting them before the big players take them seriously (roy's announcement attached)

Joseph Choi

@JosephKChoi

Playbook for 100M TikTok views for $1 CPM: > hire a creator NOT influencer w/ an audience > you script the videos, they record and send you the videos on google drive > hire 100 teenagers > teenagers clip and edit the raw videos, and post every day on each account > pay $1 CPM

clipping was discovered by 30iq streamers like 10 yrs ago tiktok algorithm if u aren’t shadow banned will give you ~200 views minimum. 700 clippers posting once a day for 3 months is an automatic 12m impressions. but nothing ever really went “viral”. often it would be to the same audience/demographic over and over again. law of large numbers maybe 1 or 2 things went viral. so now you can claim “one of my tiktoks did 1m views” AND “we generated 12m impressions in 3 months” on the low end. while the actual cultural impact is low, since youre serving the same exact audience over and over. you can’t convert a user twice. and these people don’t tend to care about your brand that much. engaged followers to a single account is a much better metric, because it’s proof that there are actually X number of distinct users who care about a person or brand. this is only possible because a large portion of tiktok users spend 2+ hrs a day on the platform, they scroll and watch like 240 videos. so they can afford a few bad (or rather, potentially bad) videos in the middle, and can give creators 200 views for free. and also maybe 5% of people on tiktok post so there are like 20x as many people scrolling. all this to say, there is little supply for videos, inflating their relative value. very interesting supply/demand dynamic. because there is a clear SHORTAGE of good content. tiktok also works vastly differently than twitter because people don’t like to be victims of ragebait. it’s like cocaine vs heroin. people doomscroll on tiktok, their mind is in a lull, not remembering any distinct tiktok just biding time. on twitter it’s like cocaine, people want to be enraged, they want to react, there are a lot more posters. the people who engage with tweets aren’t simply viewers they react, quote tweet, reply. the algo pushes things with reactions. tiktok users on the other hand like watching reality tv with a cohesive story and characters to attach to. a prime example is the cookieverse. an ethos, a shtick. like marlon and lacy are literally a sitcom, one attractive guy and one normal guy are friends. they want to parasocially join the friend groups of the characters. so you have to be likable on some level. in some ways it’s easier if you are just an individual personality posting on tiktok, because you have no choice but to be yourself. but trying to hack or engineer it consistently, and create a loyal audience is much harder than twitter. no tech/growth company has figured it out. they are all stuck on stupid one trick "formats". completely backwards.

Huh. Looks like Plato was right. A new paper shows all language models converge on the same "universal geometry" of meaning. Researchers can translate between ANY model's embeddings without seeing the original text. Implications for philosophy and vector databases alike.

this is like saying because we can translate english and spanish there exists some universal platonic language... there's nothing analytical about the isometry they are implying between the 2 manifolds. this is more kantian in the sense the implied *universal* (pardon the weak analogy) manifold/structure is noumenal--only mediated and embodied. it's existence merely *implied* by the isometric mapping they found. also there is no proof that the isometry they find is *unique*, unlike plato's forms. there's so much ontologically misguided by the way ML researchers coopt philosophy verbiage.

Next primer on @AntithesisHQ & their deterministic computer coming tomorrow :)

was writing about this the other day, super interesting company. was thinking about all the philosophically correct ideas, mips/microkernel/plan9 etc. and how the future sort of rhymes with them. from plan9 i was thinking about deterministic computing environments, then agents sandboxes. how plan9’s observability (everything is a file) rhymes with people just uses fs/bash as harness + memory. then thought about effects and deterministic computing, e.g. antithesis. i think they could be a very good agent sandbox company bc they have full observability over state. you can rollback/replay state. though modal is doing something simmilar (but less strong) with memory snapshots. although my conclusion was contrary to the train of thought because my guess is people don’t desperately need agent sandboxes with effects tracked. but it’s entirely plausible in future/at scale something important somewhere requires this for agent harnesses.

i am predicting more people will rotate into kotlin for these reasons: 1. type system is more powerful than go so you can form a sort of standard/architecture that is easy to ai generate with low context. ai written go at scale leads to duplication. 2. jvm ecosystem/libraries (new favorite: http://dbos.dev) 3. easier to read than rust mainly because it’s garbage collected and doesn’t care about low level details. swift is also in this camp, better in theory because it’s compiled/native. in practice jvm is super hyper optimized and swift itself has a very young ecosystem. i recently rewrote our backend in kotlin from gleam (a typed BEAM language like elixir/erlang). mainly because the gleam type system has the same problem as go in that it’s too simple and leads to duplication easily. we are probably going to fully cutover most of our production deployments by EOW. also, i am not that BEAM pilled because the semantics are not very explicit and cant really be merged with strong typing or event sourcing like DBOS/temporal style solutions can. it’s also hard to standardize the way models write OTP stuff, a lot of duplication. and the models have the opposite of “let it fail” posttrained into them. i used to be a kotlin hater because kotlin people get too cute and do DSLs etc. but i find the models to be perfectly capable of writing synchronous only explicit kotlin. unlike humans, they really aren’t obssesed with abstraction. i’ve been thinking about software architecture from the perspective of “how can i make it so the models write the most correct code possible with low context”. the answer is to standardize and force (sealed classes, etc) everything through a few options. and use abstractions to force the model to answer all the pertinent questions up front. basically, strict and opinionated. i like gradle too because i can strictly enforce code style and linting easily as well. when i finished the rewrite i was at half the size of the gleam backend and *more* correct. building the architecture and abstractions in kotlin, figuring out DBOS, really reignited my love for SWE. not just building abstractions but having them *work* is very satisfying.

Damian Barabonkov

@damian_b

I wasn't expecting that Go vs Rust tweet to hit a nerve. Quick elaboration based on what I've actually seen. I build a distributed sandbox system that has both Go and Rust components. I have effectively unlimited access to the strongest models, so I've spent a stupid amount of

also i think in general rewriting shit is so high leverage. i had fable/codex start to rewrite it in rust, in go, and swift. i reviewed it, figured out what felt wrong, and then landed on the odd choice of kotlin. and i think model based testing a la p lang or more modernly http://quint.sh will become very big. formal verification doesn’t scale but this is the next best thing. we want the model to generate spec that has to pass strict standards. of course the spec might be wrong but if the model has to put more cognitive effort into impl it is more likely to mess up spec. conversely if the spec is defined, implementation becomes much easier, cuz you don’t have to reason about the abstractions you may need or how to integrate your stuff with the rest of the application. the difference between spec and impl is roughly one is pure and one is effectful. it’s common FP practice to push effectful code to the edges of your architecture/application and have a pure/deterministic state machine.

we’re at this odd moment in time where taste and first principles thinking (stepping back) is still hard for the models. and none of our tooling is built for them. like the median 100 person dev org can be beaten by like 3 dudes tokenmaxxing. even among people tokenmaxxing if you have an eye for it there’s no limit to what you can do. mostly here i’m talking about building SaaS and turning product vision into actually decent/durable UX. rules are different for narrow domains like gpu kernel dev and performance engineering.

if you're doing context engineering, literally just follow this guide and you'll be fine

Dens Sumesh

@densumesh

x.com/i/article/2039…

i understand why you do this but as a based linux infra guy i can’t help but be disgusted at a non posix compatible facade in front of chromafs if i were doing this i’d do a pool of mount namespace with a read-only bind mount of pre-built docs. manage resources with cgroups use FUSE/NFS for centralized doc store

We didn’t need a 100% POSIX compatibility for our read only docs use case because the agent used only a subset of bash commands anyway to traverse the docs. This also avoids any extra infra overhead or maintenance of EC2 nodes/sandboxes that the agent would have to use.

i pretty much agree bc u alr had the chroma infra + you keep everything in typescript. the devops complexity isn't really worth it. just-bash awk/sed is limited, and text processing commands will be 10-50x slower most likely. but that's more like p99 tool call latency idk how bad that is in practice. and it would be significantly cheaper (could probably run it on one instance even at your scale) if you abandon chromafs.

Yeah, but then we need to manage 2 instances of doc content and ensure they are always in sync + doing RBAC with a real Linux FS is not trivial. Agreed on just-bash being slower tho, I ended up rewriting some of the impls for the text processing commands to make them faster.

how does the doc content get into chroma? that can't be source of truth. just put fuse in front of whatever is source of truth. actually you could even put fuse in front of chroma. but yeah RBAC is hard, you could do it with per user fuse mounts. the FUSE daemon takes the user's groups at mount time, prunes the path tree and returns EACCES or hides files entirely from readdir(). but for most startups the ops burden of your impl is way lighter than it would be for mine. but for me i'd much rather maintain a fuse daemon than write typescript impls of shell utilities.

100% agree a FUSE mount would be the way to go given more time and resources. Putting Chroma behind a FUSE adapter was my initial thought when I was implementing this but it was way too slow. I think we would also need to optimize grep even if we had a FUSE mount.

i mean idek why use chroma i assume mintlify docs source of truth isn't chroma and instead some object store?

We should 100% be storing markdown from customer repos in an object store and that’s smt im pushing for architecturally but rn we only store the transformed html in mongo so this was the next best option since we store it raw in chroma anyway

well then i agree with you completely, negl that is lowk kinda cursed tho 😭

@rohanganapa do you want to interview for a job here?

unemployable, sorry x.com/rohanganapa/st…

rohan

@rohanganapa

would like to add that polymarket is the best job on the planet for me and that i am firmly unemployable. and that also polymarket is the SINGLE GREATEST prediction market that exists.

are you based in sf? you should swing by our office sometime regardless

Of course that’s on him. See screenshots!

when a user tags @coderabbit with no message, there are three possible intentions: 1. they want a code review 2. they want some other action 3. they accidentally submitted too early current behavior: coderabbit does nothing. - users 1 and 2 are disappointed (they wanted something to happen) - user 3 is mildly annoyed at their own mistake proposed behavior: default to reviewing the code. - user 1 gets what they wanted - user 2 is no worse off than before - user 3 is no worse off than before this is low-risk because code review is read-only. there's no cost to guessing wrong. it's like a cache: a "hit" helps, and a "miss" doesn't hurt. the user intent here is very easy to track/prove, measure the next action after a naked coderabbit tag, make that the new default. i am willing to bet money the next most likely action is code review.

amazing that the solution to making taste verifiable is the “hot or not” algorithm, pioneered by zuck in his dorm room

Surya

@kickingkeys

Write up : surya.website/rling-qwen-to-… (not up to date) Sample Code : editor.p5js.org/kickingkeys/sk… editor.p5js.org/kickingkeys/sk… editor.p5js.org/kickingkeys/sk…

if you generalize this u get the tiktok algo. the latent score is inferred only from revealed preferences. but 'hot or not' is only 1 bit, while the algo uses many more priors on the true rating (watch time, like/fav/repost, etc.)

you can solve this by inverting the economies at which your scale operate at e.g. instead of 25k customers at 10k/mo have 3 customers at 1b/yr

alz

@alz_zyd_

software used to have economies of scale because it was zero marginal-cost, high fixed-cost. now software doesn't have economies of scale any more, because it's zero marginal-cost and zero fixed-cost

you can now afford super high marginal cost in absolute terms while being relatively minuscule compared to ACV. substantially inverts and changes how your GTM works vs traditional startup (consensus advice). not only marginal cost but CAC as well. so less startups playing the same game, and those who do are most likely doing it wrong. and so you have very few startup-y competitors, but still a venture scale opportunity. boil the ocean.

"if I had asked people what they wanted, they would have said smarter humans"

it's a fallacy LLMs will 1:1 replace human jobs a la macrohard the only reason plain software hasn't automated much of industry is that it's hard and expensive to make bespoke

the limiting factor in much of software was simply not being able to boil the ocean. having to pick battles. UNIX philosophy. verticalizing and simply doing much more in one product seems to be allow characteristically different types of software, both specialized and large in scope. usually, to justify large in scope projects, they have to be highly general. e.g. AWS, linux kernel, etc. and this is uniquely suited to the current capabilities of LLMs in that they are much better at writing software vs. running a business let's say. though this probably will change. regardless, software will always be cheaper than agents in. your hot path.

registering confusion: i don't really understand why Xi is still allowing Kimi to release such powerful open models. this is something i've publicly said i expect to stop soon. it doesn't make sense to me that the CCP would want open frontier capability easily available to other countries. it could still be that Xi is asleep at the wheel, or that K3 is just a cycle of capability behind where they start to take serious notice. but if things don't change soon then i'm just wrong / missing something.

models are capex intensive. they have huge upfront cost, recouped through recurring inference revenue. you can't finance training without the promise of high-margin inference later. if an open model is just as good, inference gets commoditized, and is served at cost by anyone. you can't win on price bc your margins have to pay back cost of training and theirs don't. so frontier labs become a structurally terrible business. either the govt subsidizes them or china (which will) owns the frontier. both are losses for capitalism.

We need to stop with this nonsense of crowning RFQs over Orderbooks. I’ve written about this before but once more wanted to share Orderbooks like Lighter are venue for Price Discovery; where the price is set. The price is based on supply and demand. RFQ models like Variational is a venue for Price Taking. It literally copies the price from elsewhere; physically cannot offer better pricing than the venue it hedges on without taking directional risk. Simple math: RFQ Price = Source Price + Hedging Fees + Profit. Actually, Variational isn't even a competitor to Hyperliquid/ Lighter; it is a customer of them when you check deep. There is also the Last Look problem: - In orderbook: If you see a price and click buy, you get filled. The engine is neutral, in Lighter's case it's proved by ZK that you get the fill you intended. - In RFQ: You request a price. The MM (or OLP) looks at your request, looks at the market, and then decides if they want to take the trade. If you are profitable, the RFQ engine can simply start rejecting your quotes, can delay them or giving you worse pricing. So please stop crowning RFQs. Most they can become an aggregator of actual trading venues.

Haseeb Qureshi ＞|＜

@hosseeb

CLOBs are not going to take us to the RWA promised land. Today Hyperliquid owns the liquidity for a handful of RWA macro names. But outside the top 10 traded assets (which are ~90% of volume) liquidity falls off a cliff. When there's enough retail demand, order books can work.

You're wrong here. Price discovery happens on the primary venue, period. If WTI or NVDA is being traded on HL while the primary market is open (which is most of the trading hours), then they are not doing price discovery on HL, they're doing it on the primary venue and arbing to that on HL. What you're describing is only true when the primary market is closed. Also, makers on Hyperliquid have last look as well. It's called cancel prioritization. Check the docs.

Cancel prioritization is not equivalent to last look

If you can see pending orders in the mempool, then it's equivalent

(i.e., if you're a big MM, have enough stake and run a validator. you can see pending orders pre-consensus. if you see your order is being matched, put in a cancel order and voila, last look)

u don’t even need a validator to see mempool on HL, just a node sufficiently close to the others so that they’ll peer/gossip with you of course not guaranteed you will see all mempool flow but still alpha there. i think ultimately you’re burying the lead on why RFQs solve CLOB’s liquidity issues. makers when posting liquidity on a clob have to collaterize their quotes. the underlyings sitting on the books with no taker flow is eating opportunity cost, you could be earning RFR. essentially the math goes the more velocity of taker flow you have, the more spread you collect, the more it’s worth its worth eating opportunity cost of collateralizing the liquidity you have on the books. RFQ combats this by knowing that there will be taker flow, and how much, before being forced to produce collateral for offering the opposing side. functionally the same thing happens in traditional markets via OTC desks and dark pools. if you are concerned about price impact, some very smart people will figure out how to acquire what you want and spread price impact out accordingly. but the question is who is this product for? retail has neither the size or price sensitivity to desire RFQs. if someone can kyc/kyb, they might as well go to them/tradfi venues. if they can’t, how can you operate as a broker dealer or spread impact to tradfi venues? unless there is regulatory arb, i’m not sure who really cares for onchain settlement?

what i’ve been trying: be super observant about everything you interact with, then reverse engineer how the artists intentions manifested

Ryo Lu

@ryolu_

how can engineers (or anyone) get better at design? i get this question a lot. as someone who went from cs to design, here's a real path: start with systems thinking engs already get this. design is just systems for humans & our senses, instead of machines. read “thinking in

everything you see or interact with on a daily basis someone spent hours designing it. and their whole job was to just do that. also another point about education: the first thing you learn about art, or even like literature, is that everything is intentional. you just don’t start with something, everything put on the page the artist has to do. every word in a piece of text is intentional. it mars the entire piece. it could be poorly intentioned, or somewhat careless, but intentional nonetheless. people rarely are capable of breaking a painting done to the strokes. they just take in the final product and don’t give it too much thought. they simply label it magic. traditionally “stem-brained” kids tend to overwrite this fact. the classic “blue curtains” meme and complaints about pointless pontification and over-analysis. they just take it literal. just the quick, hard, strict verification. quickly just say correct or incorrect and we’ll work our way back as efficiently as possible. so they tend to be unable to say more than a quick “looks good” at a painting. the funny part is, their favorite subject, math completely shifts from a “hard science” to a creative art at the advanced undergraduate level. when u are asked to write difficult, creative proofs. at some point, you must learn to appreciate the craft, or you never will be able to replicate it. you simply can’t rely on this impatient brute force pattern matching to solve problems. going back to the overanalysis, it’s really a gift. your interpretation is simply a reflection of you. the best part about art is that it’s dialectical, it spawns ideas and conversation beyond the authors imagination. if you want to be good at design you need to engage with the world, not antagonize it. because users don’t cerebrally antagonize products, they mindlessly and intuitively engage/use them.

watched this video over the weekend, i have always been mildly interested in packaging. given the scale of specific types of single use packaging, along with the small intricacies in their design, and the economic incentive to optimize something done at scale, it's clearly apparent if you dissect or pay attention, these things took a lot of intentionalilty and effort to design. and it's one of those things, like cellular networks, or global transport, which we use everyday and someone put a huge amount of intention into but most people have no interest in. it's interesting, how idiosyncratic these things are, and you have to remember that it's not as if anything done was at random, there's a reason for the idiosyncrasies. often, if you look into them, it's done for manufacturing purposes and cost efficiency, like the nibs on injection molded plastic. and it's indicative and exemplary of 3 sort of common but very important ideas. 1. everything around you was *made* by people (no smarter than you). 2. truly great, ubiquitous, creations disappear. their conception is forgotten. it's not obvious that there existed a time before them. as obvious as air. 3. competition is for losers, there is great value in designing things that disappear, that become ubiquitous, that people take for granted.

of course, now, software especially, cannot be analyzed in such a way. what i mean is that the close textual analysis which we learned at english class, generalized. looking at things from the perspective of the creators intention, can't really be done. because the idiosyncrasies could just be an AI hallucination. and this breaks a big rule, that defects and idiosyncrasies generally come from mistakes in mental model or conception. if it was on the page, the artist put it there for *some* reason. the rarest and somewhat worst of which is pattern matching or perhaps plain aleatoricity[0]. but AI hallucinations are this sort of perverse pattern matching. humans cannot intuit the gap in mental model, or assume the AI's mental model. its pattern matching and defects are simply distorted and an incident of specific post-training recipes. just today (and yesterday) i spent much of the day trying to get fable to design the UI for something. i can design in figma and illustrator but generally i am very slow at these things. it takes me hours just to do a little portion. so i have this workflow where i have it design concepts and pages in @paper, screenshot it and draw on it with my notes. then have it put it into code, after which i dogfood it, write detailed notes and iterate. yet, i couldn't even get something workable this time. there was this clear gap, no matter how many times I explained the UX, the ICP, rough concepts of a user journey, it just couldn't do it. i had hard tests for the backend portion, what actually needs to happen and is verifiable. that worked fairly easily. but it's designing interfaces that are the opposite of puzzles--interfaces that disappear. i just had to ratchet down, draw some concepts, explain the user journey in page by page, button by button detail, and you know i got a start to something. the creative act is making the verifiable principles and standards to be met and followed. the structure/basis for a UI that clearly communicates how to do the work. this is the truly generative part of process. reading AI-generated text or using a slopcoded applications is trying to decode the actual purpose of what stuff actually does (like a puzzle). or what stuff actually means. claude especially has these odd coinages and shorthands for often times more complex processes or implementations. and these shorthands gesture to the idea of something, but, as abstractions, they loose all requisite information and just add to the confusion. it's vindicating that the folklore of the hardest problems in computer science being naming things is still evergreen with AI.[1] there are a lot of things that have a complex or legacy style UI. take something like solidworks or ableton (though these are just complex, not necessarily outdated UIs). these are hard to learn, but you adapt to the mental model the creator and other users have. you start speaking the language. in AI generated applications this simply doesn't exist. it's just a gesture at what is trying to appear as an application. what i mean is, there's a marked difference in difficulty caused by the complex but coupled rationality in interfaces versus pure dysfunction in concept. [0] i use aleatoricity and not the common "aleatoricism" here to distinguish things done by unintentional randomness vs intentional randomness. so i mean to harken back the use of the root word "aleatoric" in the mathematical notion of aleatoric uncertainty. an example of intentional randomness would be like breakcore, using the arpeggiator to play a set of notes at random. recording some samples which sound good, chopping them up and mixing them. actually, a lot of landscapes in sound design can only be enumerated or explored by controlled randomness. another example is its use in surrealist art. [1] cache invalidation... i don't know. batting 0.500 is still pretty good.

I ran the cryptography mailing list. I know all of the people involved. I think it’s unlikely that Adam was Satoshi. He’s just the latest in a long line of people that various reporters have incorrectly claimed to have “definitively demonstrated” was Satoshi.

John Carreyrou

@JohnCarreyrou

The mystery of Satoshi Nakamoto, the pseudonymous inventor of Bitcoin, has remained unsolved for 17 years. Not anymore. Read my 18-month investigation to find out who Satoshi really is. nytimes.com/2026/04/08/bus…

in ~2017 i read the original codebase it's all the same style. uses locks/mutexes way too much, hungarian notation, windows guy, little OOP. uncommon for cypherpunks in 2008. most would be on linux. looks like he learned how to write code in the 90s. almost certainly an engineer and not academic. i honestly think it's a cia/nsa guy's side project or something. i think he purposefully anonymized his writing/stylometrics. it could be a guy in industry at microsoft or something but the engineering discipline and opsec suggests he was doing similar work for his job. makes sense in that he also never moved the coins after.

tor was created by the US govt bc if you have an anonymized communication platform that only spies used, any connection to the network would immediately identify them as a government agent. only through widespread adoption can u get anonymization. similarly in order to anonymously move money you'd need similar wide spread adoption.

the biggest problem with iran contra was paper trail btc solves this. and the public ledger leads to semi-visibility (read: MONEYROCKET) kind of how they monitor exit nodes on tor. although i would say this is mostly incidental and because of technical constraints of the time.

adding some things i missed: - adam back et al said "if i made bitcoin the ledger would be anonymous" (e.g. XMR) - an industry guy would probably write more OOP-y - there were *very* few bugs/mistakes in a very large project, an industry guy without experience most likely would not be able to make conceptually correct code in this manner - so who is a cypherpunk by trade, but not by principle, with incentive to make the ledger public? only a natsec guy.

Everyone owns skincare products, but nobody knows what's actually working. Today, @theantheaguo and I launch @lumeriaskin (YC S26) to end that. The Lumoscope is a multispectral camera that clips onto your phone and scans your skin at home with the same light used in clinics, so you see what a mirror can't: a breakout days before it surfaces, and sun damage years before it shows. Every scan is scored on hydration, redness, texture, and sebum, and it compounds into a record of your own skin. Lumi reads that history, connects flare-ups to causes you'd never suspect, and matches products to YOUR data, instead of a population average. Preorders are officially open today at $199 only until Aug 16th. Starting today, skincare will no longer feel like guessing. ☀️

legitimately a huge problem. most skincare studies isolate one compound over 8 weeks in a generalized RCT. if you have a routine with say 8 steps, it's hard to scientifically measure the efficacy of each product with respect to your skin chemistry at that time because it is super confounded with all the other steps of your routine. then there are also deeper confounders like diet/water intake/stress etc.