# Surya Dantuluri > Surya Dantuluri. I train models, write, and build things on the internet. I started with GPT-2 in high school, built ChatWithPDF, Geospot, and NotRealNews, and later founded Prava to train computer-use models. I care about agency, information asymmetry, and getting intelligence to do useful work in the real economy. Thesis: intelligence is already here. The bottleneck is diffusion. Models need to leave the screen and get pushed into the real economy, where the constraints are physical, financial, and human. ## Thesis: AGI Diffusion # Thesis I started training models when I was 17 to do my homework. I’ve been fortunate to work with some brilliant people, yet I kept coming back to the same thesis: we have digital intelligence, but it’s unevenly distributed, so much of the world still operates like it’s 1990. I can do almost anything from my phone, but people still drive to offices that haven’t changed in half a century. We have models that code better than most engineers, yet drug discovery still takes decades, total factor productivity has stalled since the 1970s, and our power grid runs on software older than me. Somewhere along the way, we ended up serving the software that was supposed to serve us. Everyone talks about AGI while sitting at the same desks, staring at the same screens, solving rehashed versions of the same problems. Where’s my flying car? In 2019 I built GPT-2 sites that reached millions of people. I spent time finding statistical anomalies and ranked in the top 100 on two exchanges. I accidentally built a ChatGPT plugin with a million users, and worked on some more things. Each time I found the same pattern: the economic output per token generated increases an OOM with every new modality—first seq2seq, then GPT-2, ChatGPT, deep-research agents, and so on. Yet we keep using AGI to make better software instead of replacing software entirely. AGI diffusion is the bottleneck to what we actually want: intelligence that works beyond the screen, diffused across the real economy where the problems that matter live. That may require significant capital and training, but it will certainly require far fewer people to deliver insanely outsized impact. This is my attempt at solving it. If we get it right, we can finally build the world that should have changed decades ago. --- ## Writing (Full Text) ### Training Imperative Date: 2025/09/08 | Labor priced in tokens URL: https://sdan.io/blog/training-imperative Most AI startups today build on top of the lab APIs. That works for now. But if history rhymes, every serious AI company will eventually train its own models — even the wrappers. Why? Because the barrier to doing so is collapsing. Distillation, fine-tuning, and post-training get easier every month. By the time ASI-level players are soaking up all the capital and talent, the only way to stay relevant will be to own your own models. ## How We Got Here In 2019, OpenAI released GPT-2 in staged checkpoints (by size) due to safety concerns. After GPT-3, others—notably Cohere and Primer—began training and releasing their own models. Midjourney and Stable Diffusion followed, building on diffusion research. Commercialization ramped in late 2022–early 2023. By 2024, anyone with a rack of GPUs could spin up a training run. By 2025, DeepSeek distilled a frontier model for ~$6M that matched OpenAI’s o1-level reasoning—four months after o1 was announced and two months after it shipped. The point: the models behind the APIs aren’t far from what anyone else can train. ## Reproducing Models Isn’t Magic The ingredients are simple: data, compute, and architecture. Transformers have already proven themselves to asymptotically outperform LSTMs. With open knowledge of pre-training, post-training, and inference, there’s little that can’t be replicated or simply vibe coded with Claude and Codex. Compute is just a skill issue on finding capital. And procuring data can be distilled, Phi-4 and Gemma as a sample of the symptom showed in a recent paper that a distilled 1B can match a 7B trained from scratch. Remember how software felt “hard” in the 2000s? You needed servers, versioning, CDs, and competent engineers. Yet, once someone found a new modality, they could dominate commerce (Amazon) or search (Google). Training today feels the same. Hard, but not impossibly hard. The wedge is not new architectures. It's data efficiency and reinforcement learning. [Diffusion may yet show some promise](https://arxiv.org/abs/2507.15857), but efficiency is a factor due to how limited we are on compute and data. ## Economics So why would an application company ever bother training? Because once most programming tasks under 30 minutes are automated by the end of this year, software itself starts to look like another direct-to-consumer brand. Distribution is what (always) matters. Cursor began as a wrapper on VSCode and GPT-4. Today, it also runs proprietary models. Officially for “Fast Apply” and features like that. But with billions of traces, Cursor could RL a model that handles hours of software engineering. At some point, the exact backend model stops mattering. What matters is that Cursor controls it. The pattern is consistent: - Start with an API wrapper to find PMF and collect data. - Fine-tune small specialized models for specific features. - Train your own models, using your data moat. - Increase Total Factor Productivity per token input, i.e. how much value do you provide the user; to retain users. More on this below. Your app effectively becomes the RL environment. Or you sell your valuable trajectories back to the labs. ## Data Is the Bottleneck OpenAI bought Statsig to capture the billions of trajectory replays they have in their “Session Replays” product that records everything users do on screen. Sutton and Silver called this the “Era of Experience” — the next frontier is data from agents interacting with their environment. That’s why I believe computer use is a meaningful path to AGI. Every software interface becomes an environment and every set of actions becomes experience data. The professional world already spends most of its life in front of a computer. Think of the hours of unlabeled, unrecorded data we’re missing. Once the model is the product, compounding input is experience. Whoever gathers the replays builds leverage. ## Token Factor Productivity I pay $200/month for Claude Pro. But the value I receive is worth five to six figures a year. So I almost receive $42 of value for every $1 spent. At that margin, the rational move is to train your own models. We should start measuring metrics of productivity instead of usage: - Tokens per unit of work. - Economic value per token. - Token Factor Productivity (TFP). The TFP metric at its simplest: `TFP = (Economic Value of Output) / (Tokens Consumed)` Where: 1) **Economic Value of Output** = the dollar value of the work produced by the model. 2) **Tokens Consumed** = the number of tokens used in the process (input + output, or whichever scope you define). Like Total Factor Productivity in economics, TFP measures how much value each token produces. For my own use, I yield less than 1-10% of tokens generated into code merged in production. So at API pricing I pay (as an example) ~$2k/month for 1.7 billion tokens blended input/output between opus and sonnet with aggressive caching. ~$200 of which is actually valuable to me to which in theory I would pay $100k/yr for, so in practice I get $42 of value per $1 on tokens. And assuming inferencing is at-cost, Claude generates 42x in TFP for me! If I ran a Claude Code wrapper unless I was running inferencing too this would be impossible in the long term, I would much rather be in the position of Devin. As AI diffuses across the economy this metric will increasingly become more popular. What is the price of a vibe coded RL environment considering the cost of input tokens and the price someone is willing to buy the final product? [What is the price someone will pay to upskill and train physicians, residents, and students](https://epoch.ai/gradient-updates/most-ai-value-will-come-from-broad-automation-not-from-r-d)? What would you pay for the work of keeping your entire org SOX compliant? ## The End Models are no longer a function of intelligence but a factor of productivity. We should measure this in TFP, while also factoring in yield, inference pricing, etc. We're watching software eat the world, and models start eating into labor. The companies that will survive will be those who can productively turn tokens into labor with the highest ROI. --- surya@suryad.com or [𝕏 @sdand](https://twitter.com/sdand) ### Sources: 1. Diffusion RL/efficiency discussion — https://arxiv.org/abs/2507.15857 2. Economic value from automation — https://epoch.ai/gradient-updates/most-ai-value-will-come-from-broad-automation-not-from-r-d 3. Time horizons across domains — https://metr.org/blog/2025-07-14-how-does-time-horizon-vary-across-domains/ 4. LMPO: Language Model Post-training with RL — https://github.com/kvfrans/lmpo 5. AI model pricing and performance trends — https://artificialanalysis.ai/trends ### Intelligence Arbitrage Date: 2024/12/04 | $25T Services Market URL: https://sdan.io/blog/intelligence-arbitrage Every decade or so, a new playbook emerges that fundamentally reshapes how we think about scaling businesses. In the 2000s, Apple and Nvidia showed us how to separate IP from manufacturing. In the 2010s, Salesforce taught us how to turn software into a utility with the cloud. Now, in 2024, we're watching perhaps the biggest transformation yet: the separation of intelligence from service delivery in the $25 trillion services sector. But this time, the playbook isn't coming from Silicon Valley - it's being written by a new breed of investors who understand that AI isn't just about automation, it's about operating leverage. The playbook: separate the "intelligence layer" from service delivery. It's not entirely uncommon now in late 2024 but there are a few interesting ways private equity is rolling into bed with this model. 2000s IP ↔ Manufacturing Apple/Nvidia 2010s Software as Service Salesforce 2024+ Intelligence ↔ Service AI-Powered Rollups **Model Rollups** Company A(leaving anonymous) builds proprietary SOTA models for a service where inputs and outputs are performed by humans. This way, they compete head-to-head with incumbents while offering services 5x cheaper and 3x faster. Selling this into contracts (think government, etc.) is slow and challenging but is sticky and often lasts for years. Instead of selling copilots, they acquire smaller agencies for 1-2x revenue—feasible in their niche because many owners are looking to retire and there's little innovation in this sector. Each acquisition provides more data, and the labor is repurposed from service delivery to data labeling, fueling a virtuous cycle: better models, more market share, and a greater data advantage. This turns low-margin services into software economics while maintaining the revenue base. This year, they've scaled to over $10 million in revenue. **Growth Buyouts** Company B, Metropolis, started as a vertical SaaS for the parking industry, where owners were trying to retire, operators were making thin margins so had a hard time selling to them. Then they bought Premier Parking with 600 locations, integrating their vision and software IP to boost EBITDA. Then they strike a ~$1.5B deal with industry leader SP+ to repeat at scale, rather than waiting for organic SaaS adoption. Metropolis's tech allows unmanned parking facilities with license plate-based access and billing, slashing labor costs. Their data flywheel spans consumer behavior, real estate optimization, dynamic pricing, and more. **Vertical Rollup** Company C (this is fictional) is venture-backed, raising $100 million—instead of a seed round it's more like a special acquisition vehicle or "vertical roll-up." They acquire senior living facilities and consolidate them like a PE firm would, handling patient intake, staff scheduling, and revenue cycle management through their platform. But they layer on voice-to-voice models for automated phone calling and text models for outpatient care coordination, improving EBITDA margins by 15-20% while enhancing care quality. Their approach prioritizes margin transformation first, revenue growth second, and multiple expansion third, with an intelligence layer that gets marginally better with each acquisition's data. The core business operates as a roll-up but with a capital structure that sits between PE and VC. This market, like many service industries ripe for consolidation, is fragmented with owners looking to retire. While traditional PE rollups focus on consolidation and implementing standard software stacks, this model does something additional—each acquisition not only increases market share but also enhances the operational dataset, which trains the models and drives continuous improvements in collections, NPS, and other metrics. The key difference is that the data flywheel becomes stronger with each acquisition, with increasing margins from 15-20% to 30-25% backing this value prop beyond what traditional consolidation alone could achieve. {` .title { font-size: 24px; font-weight: bold; fill: currentColor; fill-opacity: 1; } .step { font-size: 18px; font-weight: bold; fill: currentColor; fill-opacity: 1; } .detail { font-size: 14px; fill: currentColor; fill-opacity: 0.75; } .arrow { stroke-width: 2; fill: none; } `} Acquire 1-5x Revenue Fragmented Markets Retiring Owners Collect Operational Data Copilot Interactions Improve Repurpose Labor Voice & Text Models Boost productivity Transform Margins 15%→30% Data Flywheel Competitive Edge Intelligence Arbitrage $25T Services Market In all cases, data collection shifts from improving individual productivity to training proprietary service models and fine-tuning copilots. With Github copilot, developers are already 29% more productive, and big law firms adopting copilots like Harvey's utilization rates surging from 33% to 69% - the reality of transforming service businesses trading at 2-5x revenue into data factories comes with operational struggle. So AI just increases margins with copilots? Probably the wrong question to ask-- yes and leads to creating new business models impossible before these bigger models were commoditized. For title insurance ($57B, growing 12% annually) or insurance TPAs ($399B, 8.3% growth) - while it makes sense to consolidate and streamline with software, the play is separating the intelligence layer (risk assessment, document analysis) from service delivery (local presence, relationship management). While Lemonade tried disrupting insurance from the outside with pure software automation and got crushed on CAC and loss ratios, the better approach starts by uprooting the service. Build tools for existing players, acquire, collect interaction data to improve models, and implement at scale. Just 5% of services is bigger than all enterprise software at $656B. Layer Gross Margin Net Margin Key Components Intelligence Layer 60% and above 30% and above • Proprietary models • Data network effects • Near-zero marginal costs Software Layer 35% to 60% 10% to 30% • Workflow automation • Process standardization • Digital transformation Physical Operations Below 35% Below 10% • Local presence • Human touchpoints 2010s; build software that solves a problem and sell it to other businesses. 2024; build software that buys businesses. Whether you start with models and use acquisitions for data, compete with PE firms, or create a tech-enabled rollup - the most interesting part isn't the technology, it's how these deals get structured and funded. The real opportunity isn't just in building AI or buying companies - it's in creating entirely new financial structures that can turn the $25T services market into tomorrow's software businesses. That's the intelligence arbitrage. --- Thanks to Aidan Guo, Aryan Sharma, Hima Tammineedi, Khushi Suri, Liam Weld, Mehran Jalali, Samay Shamdasani, Udit Jain, and everyone who helped review and edit. #### Sources: 1. [Growth Buyouts - Yoni Rechtman](https://docsend.com/view/vp3w3g7djbirsx22) 2. [Artificially Intelligent Barbarians - Andrew Ziperski](https://andrewziperski.substack.com/p/artificially-intelligent-barbarians) 3. [Deloitte - Extended Enterprise Risk Management Survey](https://www2.deloitte.com/content/dam/Deloitte/global/Documents/Risk/gx-gra-eerm-survey-report-2017.pdf) 4. [Deloitte - Next Generation Managed Services](https://www2.deloitte.com/content/dam/Deloitte/jp/Documents/risk/cr/jp-cr-next-generation-managed-services-journey-cost-to-value.pdf) ### Perceived Age Date: 2024/08/07 | Half Our Life is Over by 21 URL: https://sdan.io/blog/perceived-age > "To live is to be other. It's not even possible to feel, if one feels today what he felt yesterday. To feel today what one felt yesterday is not to feel—it's to remember today what was felt yesterday, to be today’s living corpse of what yesterday was lived and lost." -- Fernando Pessoa At 2:15 PM on June 5th, kids burst through school doors, sprinting towards three months of freedom. Summer felt endless back then, August an eternity away. A day at Great America stretched like a week, and road trips to LA lasted forever. Fast forward to 16, responsibilities piling up. Even then, going from freshman to sophomore year felt like 4 years. Now at 22, the passage of time feels significantly faster, with the years seeming to fly by more rapidly as the novelty of each experience diminishes. A study asked different age groups to mentally count 120 seconds. People under 30 averaged 115 seconds; those over 50 just 87. That's a 24% reduction in perceived time. This shift in perception isn't just random; it's rooted in the complex interplay of our brain's neurotransmitters, primarily dopamine. *[1]* Dopamine plays a significant role in how we perceive time. As we age, changes in dopamine function affect our internal clocks. When we're young, everything is new and exciting—first kiss, first job, first time living away from home. These novel experiences flood our brains with dopamine, making time feel elongated. As we age, novelty diminishes, dopamine decreases, and time seems to speed up. Our internal clock system, predominantly dopaminergic, and our memory, which relies on acetylcholine neurotransmission, work together to shape our perception of time. A 2004 study showed that rats can estimate time intervals up to 40 seconds even without their cerebral cortex, indicating that time estimation is a subcortical process rather than a receptor like smell and touch. In humans, stimulants that boost dopamine function speed up the perception of time, while antipsychotic drugs or feelings of sadness and depression, which block dopamine receptors, have the opposite effect. On the other hand, novelty-induced time elongates our perception of time, while mundane repetition, like a dull office job, traps us in the illusion that time is shrinking. Claudia Hammond, a psychology writer, introduces the concept of a "reminiscence bump," which occurs when we encounter novelty in the world of firsts (travel, relationships, etc.). This novelty is tied to how we form our identity, causing the brain to grasp onto details that solidify how we present ourselves to both ourselves and others. > "And at the place where time stands still, one sees lovers kissing in the shadows of buildings, in a frozen embrace that will never let go. The loved one will never take his arms from where they are now, will never give back the bracelet of memories, will never journey afar from his lover, will never place himself in danger of self-sacrifice, will never fail to show his love, will never become jealous, will never fall in love with someone else, will never lose the passion of this instant in time." -- Alan Lightman The perceived acceleration of time as we age is a major cognitive illusion. Childhood memories seem endless because they were filled with constant discovery and the impairment of registering regret or anxiety about the past and future. Adulthood doesn't hold the same level of novelty. Repeated stimuli appear briefer than new stimuli of equal duration. Learning new things and taking on challenging cognitive tasks can potentially slow our internal sense of time. To further illustrate this, consider our conscious perceived age relative to our actual age. Discounting the first 5 years of life to childhood amnesia, when you’re 6, a year represents 16.67% of your life, by 13 it's 7.7%, and by 18 it's just 5.6%. “Perceived age” quantifies this, scaling our age to reflect our subjective experience of time passing. In other words at age 6 you’ve perceived 5 years of life but by 13 it feels like 26, each year feeling less significant than the last. Time is a construct shaped by our neurobiology, psychology, and the dance of our daily lives. Our perception of time isn't fixed; it's malleable, influenced by dopamine levels, novel experiences, and cognitive demands. Understanding this can be empowering rather than daunting. While we can't physically stop time, we can influence our perception of it. By seeking novelty, embracing change, and expanding our horizons, we can stretch our subjective sense of time. Trying new fruits, exploring different sports, subscribing to new forms of thought and radically new lifestyles(the amish and uncontacted tribes live tangentially different lives than we do for better or worse) add layers of childlike richness to life beyond the routines capitalism often dictates (buying new cars, food, consuming stuff generally). *Alejandro Jodorowsky's The Holy Mountain (1973), where surreal symbolism and rituals converge on a quest for spiritual enlightenment.* Realizing that your perceived age accelerates faster than your chronological age can make life feel like a blessing. You just need to set your life up once—start running, go to the gym, eat cleaner—and these habits will continue paying off throughout your life. Whether it's due to fear, comfort, stability, reliance, or a plethora of other excuses, maintaining a dynamic and charged life correlates with your perception of time. Establishing better habits increases your chronological time, but by age 40, when every year begins to feel indistinguishable from the last, what do you have to lose by doing shit you have never done before? It’s kind of over for you if you want to make it over for you before you even had the chance to make it. Having a career is another made up thing, it’s not real. Money is real but is infinitely more important earlier in life rather than later unless you’re pro-natalist insistent on leaving an inheritance; otherwise, it’s likely more useful at 20 than at 50. You have a good 80 years, why waste it doing one thing when you can cumulatively do a dozen things that touch a dozen different lives and communities? > “You waste years by not being able to waste hours.” —Amos Tversky I didn't know it at the time, but June 5th was the last time I'd be at school for a full year. covid came and went, school felt more of a chore juggling it between projects (& my first and last fulltime job), but despite the madness I came out winning; it's hard to find who didn't somehow in the zirp era. I agree the first week of freshman year is the most important at any school, and the rest of the 4 years is marginally not. That was taken away from me, and I felt repulsed being (for a semester or so) surrounded by people who felt incessantly opportunistic that someone (job recruiter, VC, etc.) would give them a handout. Instead, here I am writing this blog post 4 years later. If you’re reading this I probably am doing something right, or at least that's what I make myself believe. Probably driven by my paranoia and the fresh liquidity of 22 year olds on the market, I sense time is ticking even quicker, but life is a series of processes and systems glued by discipline and determination -- it's refreshing to be in the moment, with unlimited free will, embracing change, optimistic in a world where GDP goes up only, to be around working with people with unbridled ambition, and with time so endless it feels like June 5th again.   --- Thanks to Alex Reibman, Ankit Gupta, Anish Agnihotri, Caleb Sirak, Dar Sleeper, Jesse Michael Han, Justin Zheng, Khushi Suri, @mempooled, @spikedoanz for helping review and edit! And thanks to Parth Chopra for inspiring the perceived age graph calculation [2]. #### Sources: 1. [Study on Time Perception](https://www.scielo.br/j/anp/a/d6SvJK5tM6kCFPTmpVj5pSz/?format=pdf&lang=en) 2. [Perceived Life Span](https://docs.google.com/spreadsheets/d/1w0JOXq2S5LZRZPvqsI7Bjfo9-kpIc2DhL7TlyeBpFyg/edit?gid=0#gid=0) 3. [*Time Warped* by Claudia Hammond](https://www.themarginalian.org/2013/07/15/time-warped-claudia-hammond) ### Country of Services Date: 2026/05/05 | Backpropagating across a country | Archived URL: https://sdan.io/blog/country-of-services The industrial revolution made goods abundant by turning labor into capex. AI is doing the same to cognition: everything software-shaped is marginal inference cost. Yet historically America is not software-shaped, its not tool-shaped, its a country of services bundled in to limited liability organisms that trade trust for distribution and the right to be accountable for other humans. What the labs have built -- tool-shaped, inscrutable weight files that echo consciousness, locked behind a series of API calls -- provides ever-growing leverage on digital labor. Yet the rest of the economy is not inscrutable; it's owned, operated, liable, naturally anthropomorphic of how people manage bureaucracy and drive capitalism. > *You can outsource your thinking, but you can't outsource your understanding.* The core understanding of how a firm runs with its situational judgement, awareness for a customer under economic pressure, its duty for that customer, historical context on how to serve customers better -- was not previously trainable (we're now seeing data co's buy Slack logs, emails, etc). Synthetically generating this data even by paying someone to roleplay as an analyst or physician; making them feel like their service is being converted into training exhaust breaks the core trust an institution has fought so hard to earn. Understanding itself is what makes a firm an institution. In this tweet I also want to poke at the "super stack" -- we have superintelligence, superalignment, yet superdiffusion: intelligence saliently diffused across the economy, doing inter-planetary backpropogation on claims resolved, minutes saved, margin expanded, where the policy is rewarded with USD is a core third pillar humanity needs to safely monitor; focusing on accelerating human capability and expanding human agency with superaligned, superintelligent models that are economically-aligned. Imagine backpropagating across a country. The only place to run the full super stack is inside a permanent capital vehicle. Permanent capital vehicles inherit trust, build the surfaces of work, and route intelligence to seek exponentially harder economically-relevant reward signals -- accelerating human operators, expanding human agency, and enabling recursive diffusion across the firm's economy. Permanent capital vehicles are the full-super-stack institutions that turn tokens into labor. The labs are building a country of geniuses in the datacenter. What about a country of services, delivered from one permanent capital vehicle? ### Information Asymmetry Date: 2026/04/03 | High frequency info | Archived URL: https://sdan.io/blog/information-asymmetry It has never been easier to be in the top 1% of a field because the baseline competence has been raised to GPT-5. But conversely, it has never been harder to matter because there are exponentially more players standing on that same baseline. A 10x outcome on a billion dollars is fundamentally different from a 10x outcome on a million, yet the attention economy treats them as equal. The "founders" chasing ngmi passion projects are doing so because the algorithms have conditioned them to optimize for the wrong metric: social consensus rather than information asymmetry. Everyone in every demographic is online in 2025. The algorithms ruthlessly synchronize consensus, so the "in-crowd" has never been larger or more coordinated. Talent is now in-fighting over publicly available information; when competence is commoditized, differentiation collapses into politics and king-making. In this noise, net-new dollars from industries and consumers outside the circle increasingly become the higher-fidelity signal. The frontier of what's possible now lies in mining the work of people solving problems that aren't on the internet yet; extracting sequences of action across hours, days, weeks of labor to capture the tacit and hidden. We haven't even mapped our own ocean! We're exiting the era where execution and capital were as influential as ideas and socioeconomic capital. Now ideas dominate, only if you have the compute to execute them and produce net new information. The winners won't be using AGI to do the same things faster or cheaper. That has always existed. The difference now is the ability to produce net new information at scale. This begins to resemble a different form of capitalism entirely -- one that can hyper-concentrate power, currently bottlenecked more by politics than capability. When is the last time you tasted something you've never tasted before? There's infinite competition in starting a restaurant; effectively they all compete on service rather than food. Winning in some sense has never been so easy, provided you are willing to play a game that is seemingly uncompetitive. 🛫 ### Hero's Journey Date: 2025/04/05 | The Hero's Journey | Archived URL: https://sdan.io/blog/hero-journey > "The most powerful person in the world is the storyteller. The storyteller sets the vision, values, and agenda of an entire generation that is to come." — Steve Jobs Every religion, every empire, every cultural revolution began as a story that refused to die. Stories aren't static; they're living, breathing entities-- memetic complexes replicating, mutating, and spreading through whispers, screens, songs. Every pitch deck that raises millions, every movement that shifts paradigms began as stories someone chose to tell. Kanye recorded *Through the Wire* with his jaw wired shut, transforming tragedy into mythology. *College Dropout* marked the call. *Late Registration* signified crossing the threshold. *Graduation* captured the trials. *808s & Heartbreak* embodied the abyss. *My Beautiful Dark Twisted Fantasy* symbolized resurrection. *Yeezus* depicted death and rebirth. *The Life of Pablo* represented atonement. *ye* delivered revelation. *Jesus Is King* offered the gift. More than albums, these were chapters unfolding in real time—triumphs, failures, rebirths. Kanye subliminally painted a saga that reshapes the culture's psyche, revealing a hero's journey through the chaos. Genghis Khan, born an outcast child, survived in the Mongolian wilderness eating rats. Separation from his tribe forced him to become a warrior. His initiation came through uniting warring clans with nothing but a story about destiny. His return was as a force that transformed the world—his legacy so potent that 0.5% of today's global population carries his DNA. > "They thought it would be a disgrace to go forth as a group. Each entered the forest at a point he had chosen, where it was darkest and there was no path. If there's a path, it's someone else's path, and you're not on the adventure." — Joseph Campbell But what's a life worth living if every day can't become a story you tell your children? If you keep winning, where are you failing? No one likes a repetitive winner; everyone resonates with the struggle we all go through. The money, the fame is not what captivates; what captivates is how you rise from mortal realms, the courageous dive into the unknown, the continual trial, failure, adaptation, and transcendence for this new immortal world you've initiated yourself to be in. Can you run it up in a completely different arena? If you SPAC a company and make a billion, will it be a story worth sharing for the right reasons? Did you fly out your engineers to Army bases in Afghanistan to reach PMF with the troops on the ground? When do you descend back into the realm of your Partiful friend group? When do you actually do something? Everyone is trapped in a reality distortion field crafted by the people and situations around them, a reality tunnel. Every day is so delicate it can ripple through the rest of your pages. A day I can't turn into a story is a waste-- another day not upping the intensity is another day lost. The stories that matter most are the ones that spread even after we're gone-- the ones that take on a life of their own, becoming reality through repetition, infecting minds and shaping cultures as self-replicating memetic complexes. In Kanye and Genghis Khan's journeys, there was a departure, initiation, and return. They weren't the cultural zeitgeist, weren't the nicest people, but they made people feel something. Feel something about themselves and their own story. A story so insanely compelling it makes you want to follow that vision not because of the money or the fame, but because of where it will take you and the stories you can share. Because when everything inevitably fades, what lives besides stories burned into our collective consciousness? > "No one can tell you what you can or can not do. With no rules to follow, this adventure is up to you" -- Official Minecraft Trailer 2011 ### High Agency Date: 2024/03/06 | Mastery in shape rotation | Archived URL: https://sdan.io/blog/high-agency I have heard this term being thrown around recently for people who are naturally introverted (which is likely this audience) as something that you should know or try to understand. Really it comes down to how much you are able to rotate within the simulation. Rotate shapes, ideas, the game engine you're in right now. Imagine you're speeding down a highway and get into an accident. The outcomes are predictable: injuries, police intervention, insurance claims, and potentially going to jail. However, every scenario can be simulated through your mind, and you have the power to step through those simulations within your means. > "The people who are crazy enough to think they can change the world, are the ones who do." -- Steve Jobs Besides what is legally possible you can envision a different reality, one that is only constrained by the immutable laws of physics;  you just need to be audacious enough to do so and ambitious enough to pursue it; breaking free of the constructs pre-installed by our gods (simulators). In many other instances you can find yourself bound by social and internal constructs activating primal needs for safety and comfort. You seek predictability in income, living arrangements, and daily routines. While those constructs can provide stability, they hinder your ability to decide on a moment's notice to climb on top of Sutro Tower (for the thrill). Something people who drift through the system of school and employment often experience (and don't question hard enough) is whether they are adhering to those constructs because of peer influence or because it is simply out of habit, allowing them to focus their agency on other meaningful pursuits. > "Take time to deliberate, but when the time for action has arrived, stop thinking and go in." -- Napoleon. It’s likely good to allocate your agency strategically—minimize it in areas of necessity and amplify it in domains where you want to make a significant impact. The key is realizing the power you have to render anything in the world. There's no authority figure dictating what they expect from you. When I entered the "real world," I realized the absurdity of addressing middle-aged, college-educated adults as "Mr." or "Mrs.," and the unnecessary anxiety surrounding the use of "Ms." or "Mrs." No one inherently deserves your deference unless they have proven themselves worthy or earned your genuine respect. > the RLHF that colleges perform on smart students seems particularly bad not just because the goals are artificial and gameable, but also because it encourages a default operational loop of 'wait for an authority figure to tell you what to do next.' -- @nearcyan ## Shape rotation requires three essential elements: audacity, ambition, and ability. Audacity and ambition can't be bought — ability can. If you’re able to be “the guy” you can summon anyone and anything into existence through sheer force of will. Sam and Elon's genius lies in their ability to deal in trust within their circles and hope in the people they want to attract. The circles they operate in act as filters, with every introduction serving as a trade in trust. Napoleon -- "A leader is a dealer in hope.” They paint a vivid picture of a better future, one that's so compelling that others can't help but want to be a part of it. They're the ultimate fundraisers, not just of money, but of human capital. When you mix audacity, ambition, and the ability to inspire others, you become a force of nature. Smart people gravitate towards you, each adding their unique insight to the mix. Together, you become something greater than the sum of your parts – a power that can bend reality to your will. No challenge is too big, no obstacle unbeatable. This is what it means to truly master shape rotation – to focus your agency and push the limits of our simulation. I had a small taste of this at 19 and 21, but looking back, I was spreading myself too thin, chasing derivatives of tech cycles instead of integrating against them. In the age of AGI, raw skill doesn't mean much anymore. It's taste, your boldness, that makes you stand out. > Skill lacks significance when taste reigns supreme in the age of AGI -- Surya (@sdand) Even with AGI, making a profound impact in people's lives still takes hard work and unwavering commitment. While also being humble and genuinely wanting to create value for others. I spent most of 2022 unemployed, thinking money would make everything better. But what made me happy was getting emails from people all over the world, telling me how I'd helped them in some way. That beats any penthouse view. And by the time AGI is here, the legacy we leave behind becomes the ultimate measure of our worth—not the fleeting accomplishments or accolades gathered along the way, but the indelible mark we leave on the hearts and minds of those we've impacted. > "Everything around you that you call life was made up by people no smarter than you." -- Steve Jobs > > "As you become an adult, you realize that things around you weren't just always there; people made them happen. But only recently have I started to internalize how much tenacity everything requires. That hotel, that park, that railway. The world is a museum of passion projects." -- John Collison > “The world is a very malleable place. If you know what you want, and you go for it with maximum energy and drive and passion, the world will often reconfigure itself around you much more quickly and easily than you would think.” -- Marc Andreessen --- Thanks to Alex Reibman, Swarit Dholakia, Sigil Wen, Andi Andreas, Praneet Kedari, Murat Ayfer, Justin Zheng, Raj Khare for helping review and edit! ### Social Capital Date: 2020/09/01 | Interest paid in lifetimes | Archived URL: https://sdan.io/blog/social-capital In an effort to revive my original blog, suryad.com, I've decided to share more of my thoughts and reflections. One concept that has been occupying my mind lately is social capital. While I plan to delve deeper into this topic in future posts, here are some initial thoughts: A common misconception among young adults, particularly those in their early 20s, is the pursuit of financial wealth above all else. Whether it's investing in trending stocks or making decisions primarily based on financial incentives, this mindset can be detrimental in the early stages of one's career. Unless you're in dire need of cash, striving to get into prestigious institutions like Stanford, only to land a job at Stripe or a similar tech giant, and spend your days tweeting, is not the most beneficial career move. This is not to say that working at such companies is inherently bad. In fact, you could be having a fulfilling career at Google, for instance, working on projects like Google Fiber. However, I believe that the focus should be on accumulating social capital, which is not something you can achieve by working at these companies. ## Understanding Social Capital In my view, social capital is about forming close relationships with a diverse range of interesting individuals and attracting like-minded people. The key to increasing your social capital is to engage in meaningful conversations with people who share your interests or work in similar fields. The best platforms for this include Twitter, small online communities, universities, and face-to-face interactions. ### Twitter Twitter stands out as a non-intrusive social network where tech enthusiasts converge. It's an excellent platform if used correctly. Engaging with others, sharing insightful thoughts, and building interesting things can help you gain social capital. Direct messaging can also be a powerful tool, especially if you have a significant number of followers. ### Small Online Communities Being part of small online communities like genzmafia can be beneficial as they often consist of interesting individuals who are likely to know other interesting people. Contributing to these communities without constantly promoting your product can help you build social capital. ### University/Real Life Nothing beats the power of personal interactions. Most people crave meaningful connections and would appreciate getting to know interesting individuals. It's a mutually beneficial process. ### Case Studies Cory Levy and Lachy Groom serve as prime examples of individuals who are doing intriguing things behind the scenes and attracting interesting people. ## The Value of Social Capital The beauty of social capital is its longevity. Once you form a strong connection with someone, it can last a lifetime. Unlike money, social capital is not quantifiable. It's impossible to predict when an opportunity might arise from a connection, but the knowledge that people are aware of your existence is invaluable. ## The Impact of Social Capital A friend of mine went from having 50 followers on Twitter to 2k and raised a million dollars from Clubhouse in a matter of weeks, simply by sharing his ideas. I am confident that if I chose to, I could drop out and raise seed funding on the same day, simply by reaching out to a few key individuals. This would likely trigger a domino effect, as investors often follow the crowd. ## The Potential of Social Capital The best part about social capital is its accessibility. Anyone can acquire it by engaging in meaningful conversations and providing value to others. Just like stocks, those with a lot of social capital recognize the long-term potential of individuals who provide value. Almost no one would turn down a conversation with a potential Brian Chesky, and there are many such individuals reading this newsletter right now. Remember, you are the 10x stock. Your value isn't just in what you can offer now, but in your potential for growth. By investing in social capital, you're not just building relationships, you're also investing in your future. So, start those conversations, engage with your community, and watch as your social capital grows. ### Vector Fields Date: 2020/06/16 | Incentivizing work than hacking | Archived URL: https://sdan.io/blog/vector-fields I’ve always made analogies to understand concepts better: TikTok is a hacked version of books or showing how sugar is a hacked version ketosis. I didn’t realize that those points in retrospect not only are in the same domain, but also reveal something general and surprising about humans. People can be seen as balls in a vector field (see fig. 1). We all have incentives, yet our internal happiness or growth or wealth or even how humans operate biologically doesn’t jump and accomplish our incentives (or goals) immediately; they take time and effort. If your goal is to become a writer, you’ll need to ensure your incentives are aligned in the vector field of life to ensure that you (a ball) are going to tend to roll there, not magically become a writer after writing a 4-page essay. Likewise, if you want to stay fit, you’ll need to keep your vector field in the direction of draining your long term battery (adipose tissue) not your day-to-day battery (glycogen) by fasting one day and thinking the few pounds of water weight you lost is actual weight (Karpathy). Surprisingly this concept of “hacking” is prevalent in almost everything in life and explains a good deal of why humans act the way we do; thinking ourselves as balls driven by incentives that lead us to goals are so subjective, yet so general. ok boomer pic.twitter.com/NB617NDFLV— Slack (@SlackHQ) November 21, 2019 ‌ When you think of the term “hacker” you might think of Anonymous or some middle-aged white man in Russia in front of some green screens. Yet every day, you look at one when you look at yourself in the mirror: we’re all hackers in our own ways. Whether it be eating a sugary snack or drinking a sugar-filled Starbucks in the morning to “wake you up”, you’re hacking your digestive system. We were designed to eat meats and nutrient-rich vegetables while having fruit and other sugary snacks in moderation (not even close to the sugar levels in processed foods) yet we all hack our systems to ensure we run on our glycogen cycle every day and get as far as we can to ever reaching ketosis (Processed Foods and Health.). We are hackers. Think of the students who take a particular class and have a huge group chat to share questions and answers between periods of a test; they’re hacking the test as well. They are hackers. Abstractly and in the context of this expository essay: hackers are balls (representation of people) that don’t take the vector field route to the goal. They go straight across the field and get to the goal. Despite how efficient drinking sugar or “hacking” a test is, the outcome of hacking never materializes: you crash after the sugary drink and need more; you keep on “hacking” tests because you don’t know whats going on in the class. In more ways than some, we’re just dumb creatures thinking there’s always a way out — they’re isn’t. As much as we pride ourselves with winning awards or getting higher salaries, we fall back onto the hedonic treadmill. We’re all just creatures trying to accomplish a goal and we all need to follow the vector field there. That’s why thinking of ourselves as balls on a field that accomplishes goals is the best representation of ourselves. No ball is on a higher field than the other; most of us default to the same happiness as before that new car. As humans we’re also pretty mobile (unlike now) so we move around a lot. In fact, we just can’t wait to move around more: whether it be to get some food or meetup with a friend or more recently: update our feed and salviate on new information to either cancel people’s lives with a click of a button or spin up a new conspiracy. Don’t believe me? Look at the influx of day traders backed by teenagers; they’re putting a dent in traditional stock trading: that’s saying a whole lot about how much we want to either move or stir stuff up; similar to how balls roll around and knock into one another. These traders move around (make huge and swift decisions) always have goals (bump bitcoin to $10k) and bump into one another or even lose sight of what their initial goals were. These goals are blackholes; once you figure out the right incentives on this vector field of life, you’ll be able to reach that eventual goal. These incentives are learned, not given. Take for example a goal of getting a job in a field you don’t have much experience in: early in your life you’d just apply for the job with a nice cover letter. Later on, you’d realize that your network would be a more powerful activator so your incentives would be aligned towards: making some friends in the field you want to get into, after getting to know them, ask if they could recommend you for any positions they may know of. A few weeks later, there’d be a much higher chance that you would get the job than just applying online. The great thing about this analogy of life is: not only are you a ball rolling around towards the goal with the incentives on a vector field, but those incentives are strengthened every time you reach a previous goal, so it’d be much easier to reach goals later on in life if you’d accomplished many goals prior. Take for example getting ready for a marathon: from previous goals you’d know your incentives were to fast X amount of time and eat X amount of food specifially for your body, so learning what incentives you’d need to reach your goal would be much easier to get ready for that marathon. As you get older, figuring out exactly what incentives on the vector field you’d need to get to the goal gets easier. But its possible to learn rapidly from the start as well, but schools unfortunately don’t allow that: The K-12 system (in America) misaligns the incentives towards reaching the goal of a good education. It instead aligns it with intentionally hacking tests (via learning only what the teacher says) and hacking extracurriculars to ensure that students have a better shot at a good college (thus increasing the prestige of the HS and in some cases the home values of neighboring houses). If schools really weren’t aligning their students to “hack” their system, tests shouldn’t be any different from a blood test; yet they’re stress-inducing, anxiety prone, and sometimes downright hard “tests” that test your ability to learn what the teacher thinks is the best for students to learn (whether that be specific problems or specific pages from some overpriced textbook). This mentality of “hacking” your way through tests and ECs not only leads students to higher rates of depression and suicude, but it also leads to higher rates of bullying. You see, bullying isn’t some unknown phenomenon that happens between students; it happens (verbally) because students really just want to hack another system: the status system. Students bully one another, either repetitively throughout a school year or on any given day because students who are trained to hack systems also want to hack the system of status to ensure they’re on the top. You might see this in our everyday lives online as well: the “Karens” of the world trying to justify their status of being American or tech journalists trying to dethrone tech because they want to think they are more important than the new tech released (Naval). Fixing this education system that incentives students (balls) to get from 0 to 100/100 on tests without following the vector field of incentives (actual learning) or incentives students to go from 0 to “non-profit in Africa” without following the vector field of incentives (understanding systemic issues) is hard. The only true way to stop yourself from hacking this system is by starting to get genuinely interested in the subjects and delving into them in your free time. Or here’s an even simpler solution: just try to do some problems out of the homework; try to write a few more paragraphs to post on your blog; publish a website that has all the facts you’ve outlined on your history poster and ask for advice from some professors. Not only will this ensure that students are following the vector field of life, but as a byproduct, will tend to make them happier. With recommendations with the network you make early on and the realization that you need to follow certain learned incentives to reach your goals, students would be much more fulfilled with their lives that being a prop for Harvard (if you do get into Harvard by enjoying your education than hacking it, that’s insanely better). Unfortunately, if you enjoy your education than hacking it, in the long term, there’s going to be a systemic gap between a 4.0 and your GPA. This will reduce your chances at a good college and a good future, but if you make the most of it in HS, you’d be more set off than any MIT undergrad. What type of authority do I, Surya Dantuluri have over this? Well, I’ve noticed some signs, gotten into specific networks and programs; I’m starting to know a lot of genuinely interesting people who are recommending me to others as well as almost getting into prestigious accelerators that a majority of Ivy league kids haven’t. In this way, I’m creating really meaningful incentives to get to goals I want to achieve and along the way make meaningful friends and connects to get there (that many students at Ivy leagues aren’t utilizing until later on in their career). At the end of the day, we’re all rolling balls and our incentives are different for whatever goals we want in life. Our goals might even just be to get on with our day with a meal: you can hack it with sugar or you can drink water and some nutrient rich/low carb foods to create the right incentives towards finishing your meal. As we grow older, we as balls, get to understand what incentives we need to get to the goal as well as have those incentives stronger (easier to do) and ultimately be a better human and live a more meaningful life. You can definitely hack your life by going from 0 to Google or 0 to Stanford for status reasons, but long-term you, as a ball, wouldn’t be satisfied, neither will you understand the right incentives and goals to living a fulfilling life. Make sure you don’t hack the system. Enjoy the process, enjoy your education, enjoy your friends, enjoy nature, enjoy your food, and most importantly, enjoy being yourself. ### Works Cited “Biohacking Lite.” Karpathy.Github.Io, karpathy.github.io/2020/06/11/biohacking-lite/. Accessed 16 June 2020. “Processed Foods and Health.” The Nutrition Source, 24 June 2019, www.hsph.harvard.edu/nutritionsource/processed-foods/. ‌“Seek Wealth, Not Money or Status.” Naval, 28 Feb. 2019, nav.al/seek-wealth. Accessed 16 June 2020. --- ## Active Projects - [Flâneur](https://sdan.io/projects/flaneur): A map of Curius links - [Backscroll](https://sdan.io/projects/backscroll): Agent that scrolls your feeds - [Pingpong](https://sdan.io/projects/pingpong): Realtime analytics - [vmux](https://sdan.io/projects/vmux): Stateful sandboxes for agents - [WikWok](https://sdan.io/projects/quickquack): Scroll Wikipedia like TikTok - [continualcode](https://sdan.io/projects/continualcode): Continual learning coding agent harness - [GeoVLM](https://sdan.io/projects/geovlm): Training a 30b VLM to geoguess - [Geospot](https://sdan.io/projects/geospot): Teaching models to geoguess - [tviz](https://sdan.io/projects/tviz): RL visualization dashboard for the Tinker API - [Jokegen](https://sdan.io/projects/jokegen): Teaching 1T parameters to be funny - [Cuaview](https://sdan.io/projects/cuaview): Native computer-use app for Mac - [Geospot Infinity](https://sdan.io/projects/geospot-infinity): Online RL policy for geoguessing - [diffcua](https://sdan.io/projects/diffcua): Differential vision encoding for computer-use agents - [Sudo](https://sdan.io/projects/sudo): Realtime website generator - [r1-web](https://sdan.io/projects/r1web): WebGPU inference for DeepSeek-R1 - [fourcast](https://sdan.io/projects/fourcast): Permissionless prediction market - [Brainrot Stimulation](https://sdan.io/projects/brainrot-chrome): Subway surfers on ChatGPT and VSCode - [Bountylist](https://sdan.io/projects/bountylist): Small bounties for serial hackers - [yecombinator](https://sdan.io/projects/yecombinator): Searchable index of all ye tweets and quotes - [Open Animator](https://sdan.io/projects/open-animator): GPT-3.5 creates math animations from text - [vlite](https://sdan.io/projects/vlite): Fast vector DB - [ChatWithPDF](https://sdan.io/projects/chatpdf): The #1 ChatGPT plugin - [NotRealNews.net](https://sdan.io/projects/notrealnews): GPT-2 generated news sites - [Self-driving RC car](https://sdan.io/projects/raccoon): Neural network RC control - [Balsa bridge](https://sdan.io/projects/balsa): A ~15 gram bridge that can hold 15kg ## Archived Projects - [Bashroom](https://sdan.io/projects/bashroom): Shared bash for coding agents - [Qwen3-VL-JAX](https://sdan.io/projects/qwen3-vl-jax): Qwen3-VL implementation in JAX - [99.5k tps on a B200](https://sdan.io/projects/hundredtps): With llama-3.1 8b - [vlm-gym inference TUI](https://sdan.io/projects/jax-geoguess): Small vision model in JAX for geoguesssing - [nanoEBM](https://sdan.io/projects/nanoebm): Minimal Energy-based transformer - [vlm-gym](https://sdan.io/projects/vlm): RL gym for VLMs in JAX - [Cua-view](https://sdan.io/projects/cua-view): Computer-use agents inference sandbox - [Tempora](https://sdan.io/projects/tempora): Agents that scroll your feeds and deliver briefs — Made for myself, saves me some time--logs in, synthesizes feeds, self-improves. - [Mirador](https://sdan.io/projects/computer-use): Agents that take over your computer to do work — Runs on your computer or a virtual one for RPA work. - [Mercata](https://sdan.io/projects/mercata): Workspace for spreadsheets, code, and docs — Built a spreadsheet/code/docs workspace for a large enterprise. Killed it because it felt mediocre. - [minDiT](https://sdan.io/projects/mindit): Simple diffusion transformer - [Talos](https://sdan.io/projects/talos): AI agent that conducts live video and audio interviews — Screened blue-collar workers for a $2B government contract. - [Self Extend](https://sdan.io/projects/self-extend): Extend LLM context window via grouped attention - [Sentor](https://sdan.io/projects/estel): OSINT agents tracking 350k+ geopolitical signals a day — Tracked 350k+ geopolitical signals a day and flagged anomalous clusters before they hit mainstream awareness. - [gptwrapped](https://sdan.io/projects/gptwrapped): Spotify wrapped for ChatGPT - [Brainrot Stimulation for VSCode](https://sdan.io/projects/brainrot): Subway surfers on VSCode - [vibeshift](https://sdan.io/projects/vibeshift): Diffusion style transfer - [Peeko](https://sdan.io/projects/peeko): Tiktok dances as a service - [Prave](https://sdan.io/projects/prave): Boosted Stripe recovery rate 75%+ - [Ascii token counter](https://sdan.io/projects/token-counter): Count tokens for any LLM model - [functioncalling.fun](https://sdan.io/projects/functioncalling): Serverless GPT function calling - [GPT-2 in Code Interpreter](https://sdan.io/projects/gpt2-code-interpreter): Running GPT-2 with GGML entirely in ChatGPT - [Partychat](https://sdan.io/projects/partychat): Multiplayer plugin with Durable Objects - [Plugin Proxy](https://sdan.io/projects/plugin-proxy): Lightweight composable proxy made for scale - [ChatWithWebsite](https://sdan.io/projects/chatwithweb): Website content ChatGPT plugin - [GitSearch](https://sdan.io/projects/gitsearch): Github snippet search ChatGPT plugin - [ChatWithVideo](https://sdan.io/projects/chatwithvideo): Youtube transcription ChatGPT plugin - [Wordflow Ads](https://sdan.io/projects/wordflowads): Inject ads into your llms - [ChatMaps](https://sdan.io/projects/chatmaps): Semantically Search Google Maps with ChatGPT - [LLMIndex](https://sdan.io/projects/llmindex): LLM metrics dashboard - [SearchGPT](https://sdan.io/projects/searchgpt): GPT powered search engine - [randnft](https://sdan.io/projects/randnft): Proxy that points to a random Art Block - [BlueSearch](https://sdan.io/projects/blue-search): Search, sort, and summarize your Twitter timeline with GPT-3 - [wg-netcheck](https://sdan.io/projects/wg-netcheck): Print latency analysis of Wireguard config files - [getrekt](https://sdan.io/projects/getrekt): real time defi liquidation tracker - [shl.vc](https://sdan.io/projects/shlvc): VC dealflow engine - [Rapid Analytics](https://sdan.io/projects/rapid-analytics): Fast website analytics with Cloudflare Workers - [shiftsms](https://sdan.io/projects/shiftsms): SMS Client for Twitter DMs - [Bigbird](https://sdan.io/projects/bigbird): Networking and clustering system built on top of Swarm - [Wordflow Summarizer](https://sdan.io/projects/wfsum): LLM powered summarizer - [Wordflow](https://sdan.io/projects/wordflow): AI gen content media network - [Parallax](https://sdan.io/projects/parallax): Turn regular photos into 3D Photos with AI - [gradient](https://sdan.io/projects/gradients): Stripe inspired multicolor gradients - [finetune gpt-2](https://sdan.io/projects/gpt-2): Train 774M, 1.5B models with the Google's S3 optimizer - [NoWalking](https://sdan.io/projects/nowalking): PPO RL Agent trying to walk ## Notes - [AGI diffusion](https://sdan.io/notes#agi-diffusion): AGI diffusion is the bottleneck to getting intelligence to solve real-world constraints across the economy. - [Information-bound society](https://sdan.io/notes#information-bound-society): AGI shifts society from labor-bound to information-bound, where private information becomes the sharpest edge. - [Information asymmetry](https://sdan.io/notes#information-asymmetry): The biggest arbitrages happen when you know something others do not and can act sooner. - [Focus](https://sdan.io/notes#focus): Focus is choosing what you do when all things look urgent and all your instincts are noisy. - [Blinders](https://sdan.io/notes#blinders): You build by running the race in front of you instead of looking in every direction at once. - [High agency](https://sdan.io/notes#high-agency): Agency compounds when you make valuable choices before anyone asks. - [Trust economy](https://sdan.io/notes#trust-economy): Every introduction is a trade in trust, and trust compounds harder than expertise. - [Against mediocrity](https://sdan.io/notes#against-mediocrity): Mediocrity compounds into bad decisions. Quality compounds into optionality. - [Up the intensity](https://sdan.io/notes#up-the-intensity): Intensity wins against indecision; most of life is choosing how much clock speed to bring. - [Legacy as narrative](https://sdan.io/notes#legacy-as-narrative): The things that last usually begin as stories people cannot stop repeating. - [Love for the game](https://sdan.io/notes#love-for-the-game): Replace ego with love for the game and the outcome gets harder to avoid. - [Speed kills](https://sdan.io/notes#speed-kills): Action creates momentum, and momentum reveals information you cannot think your way into. - [Just ask](https://sdan.io/notes#just-ask): Information gets cheaper the sooner you ask for it and the sooner you hear no. - [Perceived age](https://sdan.io/notes#perceived-age): Each year feels shorter. The only real defense is novelty, risk, and change. - [AGI is here](https://sdan.io/notes#agi-is-here): The systems are lagging the intelligence, not the other way around. ## Contact - [GitHub](https://github.com/sdan) - [Twitter/X](https://x.com/sdand) - [LinkedIn](https://www.linkedin.com/in/surya-d/) - [Substack](https://substack.com/@sdan) - [Email](mailto:surya@suryad.com)