evanfeenstra

@fchollet Just like a LLM... Human memory is layers of progressively compressed knowledge - so u can can fit higher level patterns into a single thought, the higher up the stack you go (further forward in your frontal lobes)

The role of memorization and knowledge is to cache & reuse past cognitive work. It should be leveraged as a way to speed up cognition, not as a *replacement* for cognition.

Simply retrieving a reasoning trace looks a lot like human reasoning, until it's time to navigate uncharted territory. If you memorized all reasoning traces of humans from 10,000 BC, you could automate their lives but you could not invent modern civilization.

@akibablade @TFTC21 @nvk Yup everyone has this already... It's called a bunch of AGENTS.md / CLAUDE.md files organized around you dirs

I mean you can vibe code this yourself pretty easily. A lot of people I know, me included, have been doing their own version of this for a while. It's certainly one of my main priorities right now and I'm about 70% through my version. Love reading about other people's approach to the 'second brain' concept regardless. Gonna dig into this now.

.@nvk just released llm-wiki v0.0.10. llm-wiki is an open-source tool inspired by Andrej Karpathy’s idea for building persistent personal knowledge bases with LLMs. Instead of stateless chats that forget everything, it lets AI agents compile raw documents into a structured, interlinked Markdown wiki that keeps improving over time. This update brings better research quality, session persistence, and reliability fixes.

@jack Yup we have a multi-agent cloud coding UI pushing 100s of PRs a day... Almost all of them with goose

people are sleeping on how excellent goose has become under the hood (interface needs some work but team is pushing). it's a superpower. github.com/block/goose

Dang claude 4 is smart github.com/stakwork/stakg…

@karpathy @balajis Like here’s a view of a page on my web app, with all connections to other components, utils, endpoints, and backend data models (the graph crosses repos and services!)

@karpathy @balajis You can index and view your codebase with this project github.com/stakwork/stakg…

Good post from @balajis on the "verification gap". You could see it as there being two modes in creation. Borrowing GAN terminology: 1) generation and 2) discrimination. e.g. painting - you make a brush stroke (1) and then you look for a while to see if you improved the painting (2). these two stages are interspersed in pretty much all creative work. Second point. Discrimination can be computationally very hard. - images are by far the easiest. e.g. image generator teams can create giant grids of results to decide if one image is better than the other. thank you to the giant GPU in your brain built for processing images very fast. - text is much harder. it is skimmable, but you have to read, it is semantic, discrete and precise so you also have to reason (esp in e.g. code). - audio is maybe even harder still imo, because it force a time axis so it's not even skimmable. you're forced to spend serial compute and can't parallelize it at all. You could say that in coding LLMs have collapsed (1) to ~instant, but have done very little to address (2). A person still has to stare at the results and discriminate if they are good. This is my major criticism of LLM coding in that they casually spit out *way* too much code per query at arbitrary complexity, pretending there is no stage 2. Getting that much code is bad and scary. Instead, the LLM has to actively work with you to break down problems into little incremental steps, each more easily verifiable. It has to anticipate the computational work of (2) and reduce it as much as possible. It has to really care. This leads me to probably the biggest misunderstanding non-coders have about coding. They think that coding is about writing the code (1). It's not. It's about staring at the code (2). Loading it all into your working memory. Pacing back and forth. Thinking through all the edge cases. If you catch me at a random point while I'm "programming", I'm probably just staring at the screen and, if interrupted, really mad because it is so computationally strenuous. If we only get much faster 1, but we don't also reduce 2 (which is most of the time!), then clearly the overall speed of coding won't improve (see Amdahl's law).

@karpathy @balajis And then create maps and code “paths” for feeding to an LLM. Its the only way i’ve found to accurately describe large complex repos

@stevewdavens @stakwork Yup once your codebase is indexed go to the `mcp` dir and just run `yarn dev`. The mcp SSE server starts up on localhost:3000

@stakwork @evanfeenstra can I use this as an MCP in Cline.

Here's our approach to getting AI coding to work with large code bases and multiple repos. AST + LSP + Graph lets you retrieve just the code you need. Open source: github.com/stakwork/stakg… Demo by @evanfeenstra

@daniel_mac8 x.com/stakwork/statu…

@daniel_mac8 Here’s an open source project that does the same, and then serves context endpoints for LLM codegen github.com/stakwork/stakg…

🤯 this genius stores his entire codebase syntax in a graph database and queries it so provide context to an llm

@g_korland @LinkedIn Agreed! But RAG should be #1 on the list … context is king. I’m finding that slightly better context leads to huge improvements, independant of what model u use

Check out my latest article: The Inference Revolution: My Vision for AI in 2025 (It’s never too late 🙂) linkedin.com/pulse/inferenc… via @LinkedIn

@Jason New paradigms are never “normal”

QUICKLY!!! RAISE DEBT AND BUY BITCOIN!!!! EVERYONE MUST DO THIS RN!!! 🚨 PUMP PUMP PUMP!!! 🚨 THIS IS COMPLETELY NORMAL!!!! 😂 NOTE INVESTMENT ADVICE!!!!!

@Anthony_Bonato @mehmehturtle Tress have roots yo

Mathematicians throw shade like no others

@RyanSAdams Sry bro but US or EU is not able to change bitcoin supply. Your meme is only relevant to ETH and under

✨

@RyanSAdams He should do what’s right and say “bitcoin not crypto”. We should not be pushing scam coins on the population. Start with the coin that was distributed in a fair way

Trump came out this week in support of crypto. What should Biden do. How about just say "I support crypto too" Then call off Gary, approve ETH ETF, and let Sab 121 pass w/o veto. "Crypto isn't partisan" You're right. It's shouldn't be. Unless Biden and democrat leads keep making it partisan - it's their choice now.

@nvk @moneyball The actual signing and policy stuff is here github.com/stakwork/sphin…

@nvk @moneyball @nvk we are running it on esp32s3 here github.com/stakwork/sphin…. Runs fine but only a few channels max

Mr @moneyball this is not for low resource embedded, how can we load this into a micro? #L15" target="_blank" rel="nofollow noopener">gitlab.com/lightning-sign…

@hus_qy “Unruh demonstrated theoretically that the notion of vacuum depends on the path of the observer through spacetime. From the viewpoint of the accelerating observer, the vacuum of the inertial observer will look like a state containing many particles in thermal equilibrium”

@hus_qy The misunderstanding comes from ignorance of the “medium” that light waves travel through. Its assumed that electromagnetic waves “need no medium” but that is wrong

Since I received a few DM's asking if I am fine or having some mental breakdown because I suddenly write about a teleological world view (and even god), I want to try to explain myself 😅. So when I say that the universe is agential / goal directed / intelligent then what do I mean with that? Let's first define what "intelligence" is. I define Intelligence as the ability to successfully pursue some goal "in the future" despite living in a changing and unpredictable environment. Let's use the following situation as an example, where a lifeguard wants to rescue a drowning person: Since he is slower in water than on land, he has multiple options to get to the person e.g.: - take path A (straight line) - take path B (least time in water) - take path C (optimal path minimizing total time) - ... Studies have shown that lifeguards intuitively choose a path that is very close to the optimal one (C) and I think it is pretty obvious that this is a sign of intelligence as it requires a "good understanding" of the world and ones own capabilities to be able to "plan ahead" and run in the right direction. Now let's replace the lifeguard with a dog and the drowning person with a floating ball and ask the same question again, which path will the dog take to fetch the toy? Interestingly, the dog will also choose the same time minimizing path and I think this is actually not very surprising as dogs are usually considered to be pretty smart. So let's go one step further and replace the dog with an ant colony and the drowning person with some food that they want to bring home. Obviously the food can not float in water so let's also replace the water with e.g. gras that has a similar impact on the travelling speed of the ants (they are slower on rough surfaces). Now things start to become interesting as ants are believed to not have an extended "model of the world" that would allow them to peek towards the horizon, spot the food and then plan their trip - but again they happen to travel along the most optimal path. Instead of relying on some internal world model, they use pheromones to encode their knowledge about the world "directly in their environment". So instead of building a meta-understanding like humans and dogs that can intuitively decide these things, they exploit the fact that pheromones loose their smell over time which means that by following the strongest smell, they will arrive at the same optimal path towards the colony. Now people might say, well but these are all biological systems and if they wouldn't be able to pursue goals in the future then they wouldn't survive very long so of course they need to show signs of intelligent behavior. So let's take one last step and leave the realm of biological systems where some form of intelligence is usually assumed to be present and let's go to the lower extreme of our spectrum! What would happen if our lifeguard is an atom sending out a photon and the drowning person is another atom that receives this photon? Since the speed of light is slightly slower in water, this results in a very similar problem as before and interestingly, the photon does not take a straight line (but the same path as the humans, dogs and ants - that minimizes the total travel time). This fact is known as "Fermat's principle" but how does the atom do that? How does the atom send the photon in the correct direction to hit the destination at exactly the minimal amount of time? Does it know its destination or the water it will have to traverse in the future (similar to the lifeguard) and is therefore able to plan ahead and send the photon in the right direction? Contemporary physics doesn't really try to "explain" this phenomenon beyond just describing it in the form of a physical law. Of course people have wondered about the role of "the future" in the corresponding path integrals and how the photon might be able to obtain this knowledge but mostly, people just accept it as one of many "fundamental laws". Now there is one interesting thing about a photon and that is that it sometimes behaves like "a wave" where it's position, angular momentan and so on are not well defined until it "is measured". I believe that this is equivalent to a "search process" where the photon explores multiple different paths "it could take in parallel", to later "select" the path that was actually taken as soon as the first interaction occurs. It doesn't need information from the future because all possible ways to "get to a possible receiver" are explored in parallel and the one that "actually happens" only gets chosen in retrospect as soon as the spreading wave reaches the absorbing receiver. In other words, it is the receiver that "decides what manifests as reality" and the only reason why we observe the photon in a random state is because we can not predict which "version of the photon" will eventually win the race (it is computationally irreducible). Now if this would be true, it would have a few very profound implications: 1. Reality is virtual: We can not simulate multiple versions of reality (and later pick one) if what we call "classical reality" is already the "ground truth". This means that our reality has to be virtual or to some degree "holographic" and in fact, the 2022 nobel prize was awarded for a proof that showed that "the universe is not locally real" (is not just a space of objects bouncing into each other). 2. The universe is agential: Since the universe is running a search process combined with a retrocausal selection process, it has a "goal" (to minimize variational principles or the time between consecutive interactions). 3. The universe "understands" the non-local nature of the branchial space: The only way to stop the search process and collapse the space of possibilities "everywhere at the same time" is by knowing which competing computational processes belong to which emitter (and when they were "resolved" independently of their coordinates), which we know as "spooky action at a distance". This means that the universe is a "computational entity" that pursues the goal of minimizing variational principles while constructing a virtual perception of reality and according to our definition from the beginning it fulfills the classification for agency / intelligence. It is an "agent" that hosts (and guides) all other interactions (the ultimate "markov blanket" around all other states) and I think that humans usually address such an omniscient entity with the name "god" but even if we would prefer to call it a computer - it would still have the "size of the known universe" which seems pretty big for an ordinary computer. Maybe everything that we know as "classical reality" are just "agential programs" that run on the computational substrate of the universe and whose interactions are guided by an underlying optimization algorithm, which we perceive as "quantum mechanics". TL;DR: We seem to exist in the "imagination of some god-like computational entity" 😅