--- title: openthings. why i am building it: ai's second order url: https://openthings.live/why/ description: ai has had its first order. the second, where it changes how people live, has started, and it is missing its usability layer. the long take on why openthings exists, with the data behind it. --- skip to content the long take. six years in ai # why i am finally building openthings. ai has had its first order. the second has started, and it is missing its usability layer. this is the core of six years in ai, the last two of them heads down, and what i understand about where it is going. some of it is data. some of it is a feeling i could not shake: that we are building the wrong things, very fast. written by **anshuli** [@anshulix ↗](https://x.com/anshulix) founder of openthings, and its first employee. fourteen years building, six of them in ai. written 20 oct 2026 updated 26 oct 2026 20 min read [all the takes](https://openthings.live/takes/) ## what ai is used for today. consumer ai comes in three shapes look at almost every consumer ai brand being built today. it does one of three things. 01 ### make things faster software, music, video, images. what took weeks takes minutes. cursor, lovable, suno, runway, veo 02 ### add ai to what exists voice overs, customer support, ai voices, a new button inside the app you already use. elevenlabs, support agents, canva, notion, capcut 03 ### a better assistant a smarter chat box for the questions you already had. chatgpt, claude, gemini 04 missing ### the new web of the ai world not one more app. the layer that changes how people use ai itself: agents acting for people, with the people and brands they deal with, on an open network no lab owns. - your own agent - discovery - agent commerce - delegated payments - the human and agent network the five, below → 2.0b people use ai tools in 2026 menlo ventures, 2026 150%+ people use more than one: chatgpt, gemini, claude and perplexity reach, added up menlo ventures, us, 2026 900m people use chatgpt every week openai, feb 2026 the users are here, and they already hand ai real work: in menlo's us survey, 41% of ai users have tried an agent and 24% use one regularly. the money is following: menlo puts consumer ai spend at about $12 billion in 2025 and about $40 billion in 2026. but the shape has barely changed since the chatgpt and claude moments. each ai got smarter, and each still works alone, inside its own walls. **the person is still the one carrying the work between them.** ## the money follows the same road. ai venture funding, crunchbase ### ai venture funding by year foundation model companies took 27% of ai funding in 2024 and 40% in 2025 foundation model companies the rest of ai see the numbers | year | foundation models | rest of ai | total ai | ai share of all venture dollars | |---|---|---|---|---| | 2024 | $31b | $83b | $114b | 34% | | 2025 | $80b | $122b | $202b | close to 50% | in 2025, close to half of all venture dollars in the world went to ai, up from a third in 2024. the chart uses crunchbase's year-end split; its final count, in january 2026, is $211 billion of the $425 billion invested worldwide. openai and anthropic alone took 14% of all venture money on earth. most of the rest goes to enterprises, automations, ai inside services and apps, and doing everything faster. better models. more optimization. it is like building node and angular before there was a web that needed faster websites. ## second order effects. how a tech revolution actually lands the first order of a revolution is the technology itself. the second order is not everything built on it. **it is the few things built on it that change how people use the technology itself.** a few more websites could never make the web. the browser, search, amazon, online payments and social networking did: each one made the whole web more useful to everyone. phones followed the same path with the app store, messaging, rides and the camera. era first order: the technology second order: what made people use it more the web1983 to 2004 tcp/ip**1983** the www proposal**1989** first website**1991** the browser**1994** amazon**1995** google search**1998** paypal**1999** facebook**2004** mobile2007 to 2010 iphone**2007** android**2008** the app store**2008** whatsapp**2009** uber**2010** instagram**2010** ai2017 to now transformer, the llm engine**2017** chatgpt**2022** claude**2023** mcp**2024** trying local llms**2023** open-weight models**2025** personal agents**2026** missing the usability layer. the five below ↓ ## the five things that will make ai elevate. what the web had, and what ai is missing and no, it is not chatgpt or claude. they are the internet, not its application. they are the capability everything will run on, the way the internet was. what made the internet part of every life were five applications on top of it. ai's five are being tried now, inside each lab's walls and through personal agents that sign in as you. none of them is open yet, and none of them puts an agent on the other side. why ai is different every technology until now waited for us to tell it what happened. we observed the world, then logged it: typed the form, filled the profile, entered the timesheet. **ai is the first technology that observes the way people do.** it reads, listens and understands, as your own agent, under your key. when the machine observes, nothing has to be logged by hand, and every app built to collect what we typed loses its reason to exist. the last change this big, in my reading, was the computer, when the record moved from paper into machines. where openthings comes in **openthings is open standards for human things, with agents as the first interface.** it builds ai's five in the open, in three directions: open software your model runs, agents with a human behind them, and OHAN, the protocol they stand on. so none of the five belongs to one lab. the right column below is where it starts. [the three directions →](https://openthings.live/things/) on the web for ai where openthings starts the browser netscape, 1994 anyone could use the web your agent is the interface the ui dies. one agent of your own, on any model, shows you the work the way you understand it, on any screen the openthings mcp launching tue 27 oct search google, 1998 anyone could find anything your data becomes agentic your context lives with your agent, under your key. it goes looking for the people, services and open software you need, instead of you searching apps opensoftware, then open cards on OHAN launching tue 3 nov e-commerce amazon, 1995 anyone could buy from anywhere needs met agent to agent you do not open a business's app. your agent talks to the business's own agent: it asks, compares and agrees. you approve agents with a human behind them, openhiring first next online payments paypal, 1999 anyone could pay a stranger trust by delegation your agent acts and spends within grants you set, with a receipt for every action. trust is written into keys, not promised in terms grants with limits on OHAN in the spec social networking facebook, 2004 everyone could reach everyone how humans communicate people reach each other through their agents, sharing intent instead of copy pasting it. the inbox and the feed fall away, and you meet at the decision OHAN v1, then openDM OHAN v1 tue 27 oct together, these five are the usability layer: the ui, software and communication as we know them give way to agents on both sides, and data that works for the person it belongs to. [what dies next, on the teaser →](https://openthings.live/#death) ## the bubble crashed. the web did not. 1995 to 2005 ### what investors bet on nasdaq composite, closing points see the numbers | when | nasdaq composite | |---|---| | end of 1995 | 1,052 | | end of 1996 | 1,291 | | end of 1997 | 1,570 | | end of 1998 | 2,193 | | end of 1999 | 4,069 | | 10 mar 2000, the peak | 5,048 | | end of 2000 | 2,471 | | end of 2001 | 1,950 | | 10 oct 2002, the low | 1,108 | | end of 2002 | 1,336 | | end of 2003 | 2,003 | | end of 2004 | 2,175 | | end of 2005 | 2,205 | ### what people actually did people online worldwide, millions born along the way: google 1998, wikipedia 2001, facebook 2004, youtube 2005 see the numbers | year | people online | born that year | |---|---|---| | 1995 | 39 million | | | 1996 | 72 million | | | 1997 | 116 million | | | 1998 | 181 million | google founded | | 1999 | 274 million | | | 2000 | 391 million | | | 2001 | 493 million | wikipedia launched | | 2002 | 650 million | | | 2003 | 761 million | | | 2004 | 893 million | facebook launched | | 2005 | 1,026 million | youtube launched | on 10 march 2000 the nasdaq closed at 5,048. by october 2002 it had lost about 78%. **the web was not wrong. the bets were.** money went to the loudest ideas of the moment, like pets.com and webvan, and they were gone within two years. in 1999 excite had the chance to buy google for $750,000 and passed. search was not anybody's priority, not investors' and not builders'. meanwhile people kept coming online, from about 39 million in 1995 to over a billion in 2005. wikipedia, facebook and youtube were all born after the crash. **what became big looks obvious in hindsight. at the time it was on nobody's list.** ### what they said at the time [1995 · newsweek, clifford stoll no online database will replace your daily newspapersource ↗](https://en.wikipedia.org/wiki/Clifford_Stoll) [1998 · red herring, paul krugman the internet's impact on the economy has been no greater than the fax machine'ssource ↗](https://quoteinvestigator.com/2023/10/26/internet-fax/) [1999 · barron's cover amazon.bombsource ↗](https://ritholtz.com/2019/06/amazon-bomb-may-31-1999/) [1999 · bessemer venture partners passed on meeting google's founders, two blocks away. it lists google in its own anti-portfolio source ↗](https://www.bvp.com/anti-portfolio) the same pattern is back. the money goes to bigger models and to agents that click through the old systems faster. **the usability layer, the thing that will look obvious later, is on nobody's list again.** ## it is strange to watch it happen again. at least to me the same is happening with ai. the technology is right. the bets are on the wrong layer. ### 2000: where the money went money raised by 2000, us$, log scale ### 2026: where the money goes largest round per kind of bet, 2025 to 2026, us$, log scale the shiny bets a new layer people touch left to right is my reading of each bet. heights are the money raised. in 2000 the shiny bets got the most, and were gone by 2002. today the biggest rounds go to bigger models and to agents that work the old systems for you: searching, calling, booking, answering tickets. the usability layer, where agents reach people, has almost nothing. it looks like a bubble to me. most of this money builds agents that click through the old systems faster. when those systems give way, what are these agents for? see the numbers | company | the bet | raised | era | |---|---|---|---| | boo.com | fashion, online | $135m | by 2000 | | webvan | groceries, delivered | $771m | by 2000 | | pets.com | pet food, delivered | $183m | by 2000 | | kozmo | anything, in an hour | $250m | by 2000 | | paypal | paying a stranger online | $123m | by 2000 | | google | finding anything | $25m | by 2000 | | calling agents | bland $50m, vapi $50m (may 2026) | $50m | 2025 to 2026, largest round | | seo agents | profound $180m (sep 2026) | $180m | 2025 to 2026, largest round | | company brains | sierra $950m (may 2026), decagon $250m (jan 2026), glean $150m (jun 2025) | $950m | 2025 to 2026, largest round | | booking agents | fora $60m (jul 2026), ai assisted bookings | $60m | 2025 to 2026, largest round | | personal agents | instinct $1b (sep 2026), manus $500m (oct 2026), perplexity $200m for comet (jun 2026) | $1b | 2025 to 2026, largest round | | bigger models | openai $122b (mar 2026), anthropic $65b (may 2026), xai $20b (jan 2026) | $122b | 2025 to 2026, largest round | | the human and agent layer | no round for an open layer for people. the nearest money is agent permissions for companies: arcade $60m (jun 2026) | no round | 2025 to 2026, largest round | half of all venture money now goes to ai, and most of it goes to bigger models and to doing the old things faster. that is still the first order. the labs have their incentives: if they let the usability layer go to the open, they think their value goes down. i differ here. an open layer grows the pie for everyone, the way the open web grew every company built on it. everyone else is running after what works and what looks real right now: enterprises, automations, ai enablement for services and apps. all of it makes the old way faster, and all of it stands on systems that are about to give way. the companies that own where people spend their time will not build what replaces them: they would throttle their own business. not all of it is wrong. **agents and local llms look right.** what they miss is the way people actually start using them, without the hassle: an agent of their own, meeting other people's agents, with nothing in between. that is the usability layer, and it is where the second order of ai starts, the way the browser and search started it for the web. ## the connecting layer is how humans grow. centuries of less cognitive load for centuries, humans have been trying to do one thing: carry less in our heads. every big jump came the same way. first a connecting layer, a base everyone can stand on. then the layers people build above it. roads, the post and ships let goods and letters travel, and trade and markets grew on top. the press let one voice reach thousands. the internet connected the computers, and email, search, saas and online shops grew above it. the phone put all of it in every pocket, and maps, uber and upi grew above that. ### every jump came from a new base layer the base everyone stands on, and what people built above it. heights are my reading of what one person could do. trade markets cities **roads, the post and ships**for centuriesgoods and letters travel books newspapers science **the printing press**1440one voice reaches thousands news wires live markets calling home **the telegraph and phone**1844 to 1876a voice crosses an ocean email search saas online shops **the internet and the web**1983 to 1991computers reach each other maps uber instagram upi **the phone and app stores**2007 to 2008all of it, in every pocket the second order missing its layer the human and agent network today: inside apps and browsers **ai models and agents**2017 to nowmachines that read, write, reason and act a base layer what grew above it becoming the base now what we build on it first see it as a table | when | the base layer | what it did | what grew above it | |---|---|---|---| | for centuries | roads, the post and ships | goods and letters travel | trade, markets, cities | | 1440 | the printing press | one voice reaches thousands | books, newspapers, science | | 1844 to 1876 | the telegraph and phone | a voice crosses an ocean | news wires, live markets, calling home | | 1983 to 1991 | the internet and the web | computers reach each other | email, search, saas, online shops | | 2007 to 2008 | the phone and app stores | all of it, in every pocket | maps, uber, instagram, upi | | 2017 to now | ai models and agents (becoming the base) | machines that read, write, reason and act | today it works inside apps and browsers. next, built on it: the human and agent network. then the second order | what took humans four generations, one person can do in months today. and what was not possible at all, we do every day. ai is only going to expand that, and this time the base itself can think. people already touch it first: 2 billion people use ai tools in 2026. but today the models still work through the old base, inside apps and browsers built for eyes, as one more thing on the phone. **what we want is the other way round: ai models and agents as the base layer, and everything else built on them.** the first thing to build on that base is the layer between people and their agents, so your agent can reach theirs, with a person deciding on both ends. a revolution holds no meaning for what is working today. the car did not make the horse carriage better. it rebuilt roads, fuel, cities and work around itself, and the carriage became scrap. when ai becomes the base, it will rebuild in both directions. the last time the base changed ### the car arrives. the carriage becomes scrap nobody upgraded the carriage. the world was rebuilt for the new base, and what could not take it was left behind. when ai becomes the base ### any car you already own can drive itself much of what works today gains abilities nobody planned for. the rest becomes the next carriage. a picture, not data. the dots are the one small piece each car gets. that is what a new base layer does. it does not negotiate with what exists. your mail, your chats, the old software at work, your clinic's booking system: some of it will be rebuilt to work through your agent, and some of it will stop making sense. **the inventions will come both ways: new things the old ones can do, and new things that make the old ones pointless.** i have my theories about what gets built first when ai is the base. openthings is where i am testing them. [see what openthings builds on it](https://openthings.live/things/) ## the hard problems everyone is working on. and where openthings stands ask the people building ai what is still broken and you get the same list: who owns the thing you talk to, agents that cannot work with each other, no way to find the right agent, memory stuck inside one app, agents nobody can verify, agents that cannot pay, attacks hidden in plain text, models that state wrong things with confidence, and data that is used up or locked away. ask them how it gets solved, and the answers could not be further apart. most of the fixes on offer are a device, a browser or a protocol that one company controls. ### the five layer cake, twice jensen huang's stack describes one side. people live between two. your side your agent - energy - chips - infrastructure - models - applications **the missing piece**an open protocol, agents on both ends their side their agent - energy - chips - infrastructure - models - applications every layer of the cake has a giant behind it, and each one makes its own stack taller. but my agent and yours run on different stacks, from different labs. what connects them is not a layer of either cake. ### how far apart the answers are the same problems, read by the people building ai. left to right is my reading of each view. 1 **jensen huang**nvidia · mar 2026 build the whole stack: energy, chips, ai factories, models, applications 2 **sam altman**openai · jan 2025 confident they know how to build agi, as it was traditionally understood 3 **dario amodei**anthropic · may 2025 models probably hallucinate less than humans, just in stranger ways 4 **satya nadella**microsoft · dec 2024 business apps collapse into an ai tier of agents 5 **demis hassabis**google deepmind · 2025 agi five to ten years out. one or two breakthroughs still needed 6 **andrej karpathy**eureka labs · oct 2025 agents just do not work yet. this is the decade of agents, not the year 7 **ilya sutskever**ssi · 2024 and 2025 we hit peak data. the age of scaling is over, back to research 8 **yann lecun**ami labs · nov 2025 llms are a dead end. build world models instead and openthings not on this line. whichever of them is right about the model, agents on both ends still need an open way to reach each other, and a person to decide. the problem what is being tried what i think the real answer is **the interface**every company wants to own the thing you talk to humane's ai pin was bricked in feb 2025 after hp bought the company. openai says its first device will not ship before the end of feb 2027, and reports point to an agent-first openai phone around 2028. apple pays for google's gemini to power the new siri. browsers become agents too: comet (jul 2025), chatgpt atlas (oct 2025). hardware loses its weight once the intelligence is in the model. if ai can make even a table talk, nobody needs a new gadget to reach it. what has to change is the systems: they become agentic and take on the intelligence. that is what openthings shows, one system at a time. **interoperability**agents from different makers cannot work together mcp (anthropic, nov 2024, given to the linux foundation's agentic ai foundation in dec 2025) connects agents to tools. a2a (google, apr 2025, now at the linux foundation) connects agents across frameworks. newer pieces keep arriving: the NANDA index (mit media lab and project nanda) resolves an agent's name to where it lives, and aweb gives agents signed mail between them. each connects agents, not the people behind them. a handle and a signed card for every person, so any agent on any model can reach another person's agent. and named shapes both sides speak, so software that never met can share work with no api in between. interoperability between people, not only between tools and companies. **finding**agents still search a web built for eyes every assistant answers from a crawled index, then reads the page. when google shows an ai summary, clicks on a result fall from 15% to 8%, so the visit fades while the crawl goes on: over half of what ai crawlers fetch is pages they already had. open registries of agents fill with noise: of 173,473 on erc-8004, only 3 to 15% work. the live question goes straight to the other side's agent: a page can say a plumber exists, only the plumber knows if he is free tomorrow. the lists of who is out there stay open and many, so whoever pays for one never owns it, and your own agent decides the order, from your own people first. people are introduced, never browsed. **memory and context**every assistant knows you only inside its own app chatgpt (apr 2025) and claude (oct 2025) each remember you inside their own product. switching means pasting a prompt's output into claude's importer (mar 2026), because chatgpt's data export leaves its memory out. memory is not something people should look after, any more than they run their own servers. once agents are the first interface, memory becomes one more layer of communication: everything a person does, at home, at work, anywhere, becomes memory. it is theirs and private, kept by their own agent, and shared only as far as they allow. memory locked in one app, or carried over by copy and paste, never gets there. **identity and trust**few can check who an agent acts for, and only through one company web bot auth, an ietf draft that cloudflare and others put into production, signs an agent's requests so a site knows which company runs it. x401 (in production since oct 2026) ties an agent to a verified person through a certificate, and agentmail's agentid hands an app the owner's email. each runs through one issuer, and none says what the agent may do. a person or an organisation at the root of every agent, with keys they hold, and agents as sub-keys with grants that only shrink. any side can check who stands behind an agent and what it is allowed to do, with no single issuer in the middle. **payments and commerce**an agent cannot pay without a human clicking checkout google's ap2 (v0.2) signs a checkout mandate and a payment mandate, and returns a receipt for each. openai and stripe's agentic commerce protocol powered instant checkout, until openai retired it as a feature of its own in mar 2026: walmart saw shoppers buy three times less inside the chat than on its own site. visa and mastercard each run their own agent scheme, and in oct 2026 godaddy, google, cloudflare and infoblox demoed a passkey approval turning into a signed payment mandate. good rails, each for paying only. businesses want agents to find them, and they want to keep their customer. agents on both sides give them both. payments already have mandates and receipts. everything else an agent does for you does not: a meeting, a draft, a disclosure. one open grant covers all of them, with limits and a signed receipt, and hands a payment to whichever rail you use. **security**an agent can be talked into anything by text it reads ai browsers were shown to follow instructions hidden in pages. openai says prompt injection is unlikely to ever be fully solved (dec 2025), and the uk ncsc warns it may never be fully mitigated. if a model can be talked into anything, it must not hold the keys. and people already skip the yes: 32% of ai users in the us have let ai act for them without a final sign-off. so plain code at the gate checks the grant, and nothing goes out without a yes. **hallucination**models state wrong things with confidence bigger models, reasoning, retrieval, evals. openai's own research (sep 2025) finds training rewards guessing over saying i do not know. dario amodei says models probably hallucinate less than humans, just in stranger ways. the model drafts and never decides. claims travel signed, with their evidence, and the other side's agent checks the evidence instead of believing the text. **data after ai**the public internet is used up, and the rest is locked away ilya sutskever: we hit peak data, there is one internet. labs license archives and make synthetic data, and cloudflare now blocks ai crawlers by default, with pay per crawl (jul 2025). the next data is what people and businesses say for themselves, through their own agents, instead of pages waiting to be crawled. it should stay theirs: shared by intent, scoped, revocable, with a receipt. and the bottom of the cake, energy and chips, is not ours to solve. the person already pays their model vendor for the intelligence. the network never needs its own. every lab is building a taller stack. the problems live between stacks. that is why the answer is an open protocol, with agents on both ends. ## we need a movement like www again. the layer after tcp/ip and http the web did not win because it was the best technology. it won because it was open: nobody owned it, anyone could build on it, and a few shared protocols let every machine and every page reach every other. [**1993**cern puts the web's software in the public domain, so anyone can build on it](https://web30.web.cern.ch/web-history.html)[**1994**tim berners-lee founds the w3c, to keep the web's standards open](https://web30.web.cern.ch/web-history.html)[**6 billion**people online in 2025, about three in four people on earth](https://www.itu.int/en/mediacentre/Pages/PR-2025-11-17-Facts-and-Figures.aspx) ### 01 · the infra we still run on the pipes still work. the top of the stack assumes a person. every site expects a human to log in, click, prove they are not a robot and keep a session. an agent has to pretend to be that person, or get blocked as a bot. the web was built for eyes and hands, and that is now what limits it. assumes a person at a screen - **captcha**2000s prove you are a human, not a bot - **oauth**2012 sign in with a platform you already use - **the login**everywhere an account and a password per site - **cookies**1994 a browser remembers your session - **http and html**1989 to 1993 pages for a person to read and click still works for agents - **ssl and tls**1995 a padlock between you and one site - **dns**1983 names for machines - **tcp/ip**1983 machines reach each other ### 02 · who is working on the agent layer each is a real piece. none is something a person can use yet. most describe documents, tools, companies and agents. the newest drafts start from the person, which is the right start. but none is adopted, none ships to people, and none puts an agent on both sides of a human thing, with a gate that waits for both people's yes, under keys the people hold, without a platform in the middle. the purpose the web had, a network that works for everyone, is still unsolved for agents. w3c [decentralized identifiers v1.0 (2022) and verifiable credentials 2.0 (may 2025)](https://www.w3.org/TR/did-core/) identities and claims anyone can verify, without one company in the middle w3c [ai agent protocol community group (may 2025)](https://www.w3.org/community/agentprotocol/) how agents find each other, share intent, and prove who they are. early, a draft white paper linux foundation [mcp (agentic ai foundation, dec 2025) and a2a (v1.0, signed agent cards)](https://blog.modelcontextprotocol.io/posts/2025-12-09-mcp-joins-agentic-ai-foundation/) agents to tools, and agents to agents across frameworks ietf [web bot auth (draft)](https://workos.com/blog/web-bot-auth-identify-ai-agents) signed requests that tell a site which company runs an agent ietf [principal agent protocol (jun 2026) and agent delegation receipts (oct 2026), individual drafts](https://datatracker.ietf.org/doc/html/draft-baur-pap-02) a person's key as the root, grants that only shrink, signed receipts. proposals, not standards, and nothing a person can use yet ### 03 · the layer that is missing the web was built to carry pages to people. the agent world has to carry the people themselves: who they are, what they know, what they want, and what their agents may do. what the new network has to carry - identity - authority - context - memory - intent - trust - consent - payment - a person's yes - and what robots and machines bring next this is not one more protocol on the web. it is a new network, for people, their agents, models and machines, and it will need things nobody can name yet. OHAN is where i start: the person at the root, everything else built around them. ### the web, how it was built tcp/ip1983 machines can reach each other http and the web1989 to 1991 documents can link to each other browsers and search1993 to 1998 anyone can find anything the web we usesecond order billions of people, everything online ### ai, how it is being built models2017 to now machines can read, write and reason mcp, a2a and agent protocols2024 to now agents can use tools, and talk to other agents missing: the usability layerOHAN a network for people, agents and machines: identity, context, intent, trust, consent, and more the agent websecond order being tried, missing its usability layer after tcp/ip, http changed everything about how we share. with agents it changes again: how humans, machines, robots and llms talk to each other, and stay connected. mcp is a real start. anthropic released it in november 2024, it has more than 10,000 public servers, and it now sits with the linux foundation's agentic ai foundation. **but it connects agents to tools. the pieces for the person are being drafted one by one, and nobody has put them together into something people can use.** that is what [OHAN](https://openthings.live/next/) is for. ## the connecting layers took years, not sprints. tcp/ip, the web, and now nobody built tcp/ip or http in two weeks, or in the six months a funding round wants. bob kahn and vint cerf sketched tcp/ip in 1973, with darpa paying for the patience. it took ten years, round after round of rewriting and a split into two protocols, before the arpanet switched over on 1 january 1983. ### how long the connecting layers took years from the first sketch. hover a dot for the moment. what a funding round expects2 weeks to 6 months tcp/ip10 years, sketch to switch-on **1973** kahn and cerf sketch it, darpa funds it · **1974** the paper, ieee transactions · **1977** first test across three networks · **1978** split into tcp and ip · **1983** the arpanet switches over, 1 january the web4 years to free and easy, 6 to everywhere **1989** the proposal: vague but exciting · **1990** first server, on a next cube · **1991** goes public. conference paper rejected · **1993** cern makes it free. mosaic ships · **1995** netscape goes public. the web goes mainstream the agent layeryear 2. the web's 1991 if history rhymes: 2028 to 2034 we are here **2024** mcp is released · **2025** mcp moves to the linux foundation · **2026** today 0 2 4 6 8 10 12 years since the first sketch tim berners-lee's proposal for the web came back from his manager at cern with a note: vague but exciting. his paper for the hypertext conference in 1991 was rejected, and the demo he gave there instead got a cool reception. the web only took off after cern made it free for anyone in april 1993, and a browser called mosaic made it easy to use. by that clock, ai's connecting layer is in its 1991. public, working, and mostly misunderstood. ## alien to most rooms. my journey, so far when i started talking about this with peers and a few well known investors, it was alien to most of them, or it sounded like something they already knew. i understand why. for thirty years the playbook has been the same: saas, the lean startup, five ways to a hundred million dollar business. all of it bets on scaling the application layer, on top of a layer people touch that someone else already built: the browser, search, and lately the phone. ### i have been alien many times - 2012 to 2013 **drag and drop builders**website and app builders you could use without code, and javascript on the backend, when that was still a strange idea now how many sites get made - then **single page apps**the web as an app, not a stack of pages now how most web apps are built - widely **progressive web apps**i founded widely on pwas, and worked with the people writing the w3c standard now installable on every phone - 2019 **ai solutions**building with ai before most rooms had heard of a language model now in almost every product - now **openthings**the usability layer: agents on both sides, open rules between people alien to most rooms, again being early, again and again, taught me the two things this work needs most: **patience, and staying longer with the problem than anyone else.** with ai, all of it has come true at once. and the time has changed. **the usability layer, the one people actually touch, is being rebuilt.** that needs a different kind of business: not vc shaped, and not conformist to what we all adhered to. i did that in my last startup, and it worked. after a few of those conversations i knew pitching would waste my time. if i could afford it, i should build and experiment instead. i would have loved more hands, we would go faster. but this idea's time is now and the next few years. **we are right on time.** ## tested in a room, with two laptops. and three things for the future i have a vision, and ideas tested on my own problems, in a room with two laptops, trying what i imagined. i think it can work. not alone. with the force of people who want three things for the future. 01 no one or two labs or brands owning human data, or human interests. 02 ai that lets us live a more private, personal and automated life. 03 a world built in open source, free, with ai giving every human capabilities like never before, the way computers and phones did with the internet. [message @anshulix on x](https://x.com/anshulix) [read what is next](https://openthings.live/next/) anshuli founder, openthings ## sources. checked oct 2026 - nasdaq peak, low and decline: [wikipedia, nasdaq composite](https://en.wikipedia.org/wiki/Nasdaq_Composite). year-end closes: [macrotrends](https://pro.macrotrends.net/1320/nasdaq-historical-chart). - people online: [our world in data, number of internet users](https://ourworldindata.org/grapher/number-of-internet-users) (itu and world bank). 1995 to 2004 are summed from country rows. - what they said at the time: clifford stoll, newsweek, 1995 ([wikipedia](https://en.wikipedia.org/wiki/Clifford_Stoll), [paleofuture](https://paleofuture.com/blog/2008/3/28/the-internet-bah-1995.html)); paul krugman, red herring, june 1998 ([quote investigator](https://quoteinvestigator.com/2023/10/26/internet-fax/)); barron's "amazon.bomb", 31 may 1999 ([ritholtz](https://ritholtz.com/2019/06/amazon-bomb-may-31-1999/), [cnbc](https://www.cnbc.com/2021/10/12/jeff-bezos-shared-a-22-year-old-article-predicting-amazons-failure.html)); bessemer's google pass ([bessemer anti-portfolio](https://www.bvp.com/anti-portfolio)). - the hard problems, 2025 to 2026: humane and hp ([wareable](https://www.wareable.com/news/humane-ai-pin-discontinued-hp-acquisition-deal)); openai's device date ([mactrast, on the court filing](https://www.mactrast.com/2026/02/release-of-jony-ive-designed-ai-device-delayed-to-2027-thanks-to-lawsuit/amp/)) and phone reports ([techcrunch](https://techcrunch.com/2026/04/27/openai-could-be-making-a-phone-with-ai-agents-replacing-apps/)); apple and gemini ([android authority](https://www.androidauthority.com/apple-siri-google-gemini-deal-3631890)); mcp to the agentic ai foundation ([mcp blog](https://blog.modelcontextprotocol.io/posts/2025-12-09-mcp-joins-agentic-ai-foundation/)); claude's memory import ([pcworld](https://pcworld.com/article/3076376/claude-can-now-import-chat-histories-from-chatgpt-and-other-ais.html), [techcabal](https://techcabal.com/2026/05/27/4-easy-steps-to-move-your-chatgpt-memory-to-claude/)); web bot auth ([workos](https://workos.com/blog/web-bot-auth-identify-ai-agents)); ap2 and agent payments ([siliconangle](https://siliconangle.com/2025/09/16/googles-agent-payments-protocol-provides-framework-ai-automate-consumer-purchases/), [aimultiple](https://research.aimultiple.com/agentic-payments)); ai browsers and prompt injection ([simon willison](https://simonwillison.net/tags/perplexity/), [techcrunch](https://techcrunch.com/2025/12/22/openai-says-ai-browsers-may-always-be-vulnerable-to-prompt-injection-attacks/), [cyberscoop](https://cyberscoop.com/openai-chatgpt-atlas-prompt-injection-browser-agent-security-update-head-of-preparedness/)); why models hallucinate ([openai](https://openai.com/index/why-language-models-hallucinate/)); cloudflare pay per crawl ([cloudflare](https://blog.cloudflare.com/introducing-pay-per-crawl/)). added on 26 oct 2026: the NANDA index ([ietf draft](https://datatracker.ietf.org/doc/draft-raskar-agentic-web-federated-resolution/)); aweb ([github](https://github.com/awebai/aweb)); clicks with an ai summary ([pew research center](https://www.pewresearch.org/short-reads/2025/07/22/google-users-are-less-likely-to-click-on-links-when-an-ai-summary-appears-in-the-results/)); crawlers re-fetching unchanged pages ([techcrunch](https://techcrunch.com/2026/07/01/cloudflares-new-policy-pushes-ai-companies-to-pay-for-publishers-content/)); erc-8004 ([arxiv](https://arxiv.org/html/2606.26028)); x401 ([help net security](https://www.helpnetsecurity.com/2026/06/26/proofs-x401-establishes-an-open-protocol-for-ai-agent-identity-and-authorization/)); agentid ([agentmail](https://www.agentmail.to/blog/introducing-agentid)); instant checkout retired ([techcrunch](https://techcrunch.com/2026/03/24/openais-plans-to-make-chatgpt-more-like-amazon-arent-going-so-well), [modern retail](https://www.modernretail.co/technology/what-went-wrong-with-chatgpts-instant-checkout/)); the passkey to payment demo ([tech coffee house](https://techcoffeehouse.com/2026/10/16/godaddy-google-and-cloudflare-unveil-ai-agent-trust-system/)). - the web and its infra: cern's public domain release and the w3c ([cern, web history](https://web30.web.cern.ch/web-history.html)); people online in 2025 ([itu, facts and figures 2025](https://www.itu.int/en/mediacentre/Pages/PR-2025-11-17-Facts-and-Figures.aspx)); [dns](https://en.wikipedia.org/wiki/Domain_Name_System), [cookies](https://en.wikipedia.org/wiki/HTTP_cookie), [ssl and tls](https://en.wikipedia.org/wiki/Transport_Layer_Security), [oauth 2.0, rfc 6749](https://datatracker.ietf.org/doc/html/rfc6749), [captcha](https://en.wikipedia.org/wiki/CAPTCHA). agent standards: [w3c did core](https://www.w3.org/TR/did-core/), [verifiable credentials 2.0](https://startwithidentity.com/blog/2025-05-15-w3c-publishes-verifiable-credentials-data-model-2-0-as-a-recommendatio/), [w3c ai agent protocol community group](https://www.w3.org/community/agentprotocol/), [a2a v1.0](https://about.you.com/resources/a2a-protocol-explained-what-agent-to-agent-communication-solves). - excite and google, 1999: [today i found out](https://www.todayifoundout.com/index.php/2013/03/excite-had-a-chance-to-buy-google-for-750k-but-turned-it-down/), from george bell's own account. - app store at launch and a year later: [fortune, 2009](https://fortune.com/2009/07/11/snapshot-of-the-iphone-app-store-one-year-later). - ai venture funding, 2024 and 2025: [crunchbase news](https://news.crunchbase.com/ai/big-funding-trends-charts-eoy-2025/), and its final 2025 count ([crunchbase news, jan 2026](https://news.crunchbase.com/venture/funding-data-third-largest-year-2025/)). - consumer ai users, agent use, use of more than one assistant, acting without a sign-off, and spend: [menlo ventures, the state of consumer ai 2026](https://menlovc.com/perspective/2026-the-state-of-consumer-ai/) ([report](https://menlovc.com/wp-content/uploads/2026/09/menlo_ventures_consumer_ai_report-2026.pdf)). the share figures come from its survey of 5,067 us adults. - person-first agent drafts and agent permissions: [principal agent protocol](https://datatracker.ietf.org/doc/html/draft-baur-pap-02), [agent delegation receipts](https://datatracker.ietf.org/doc/draft-nelson-agent-delegation-receipts/), [attenuated delegation profile](https://datatracker.ietf.org/doc/draft-hamr-oauth-agent-delegation/) (ietf individual drafts, no ietf standing); arcade's $60m round ([pymnts, jun 2026](https://www.pymnts.com/?p=3850692)); ap2 v0.2 ([ap2 specification](https://ap2-protocol.org/ap2/specification/)). - chatgpt weekly users: [techcrunch, feb 2026](https://techcrunch.com/2026/02/27/chatgpt-reaches-900m-weekly-active-users), on openai's announcement. - where the money went, 2000: [wikipedia, webvan](https://en.wikipedia.org/wiki/Webvan), [venturebeat, the worst vc investments](https://venturebeat.com/technology/20-worst-vc-investments-of-all-time), [google's 1999 funding release](https://thenextweb.com/news/googles-press-release), [x.com and paypal, april 2000](https://www.paypalobjects.com/html/pr-040500.html). today: [valueaddvc, biggest ai rounds of 2026](https://valueaddvc.com/blog/biggest-ai-funding-rounds-of-2026-so-far-ranked), [techcrunch, instinct](https://techcrunch.com/2026/09/28/viral-ai-agent-instinct-raises-1b-series-c-at-a-10b-valuation/), [cnbc, manus](https://www.cnbc.com/2026/10/08/manus-fund-raise-meta-muse-tencent.html), [techtimes, perplexity comet](https://www.techtimes.com/articles/318028/20260608/perplexity-raises-200-million-comet-ai-browser-agent-economy-front-door.htm), [cmswire, sierra](https://www.cmswire.com/customer-experience/sierra-raises-950m-at-15b-valuation-eyes-transformation-beyond-customer-support/), [cmswire, decagon](https://www.cmswire.com/contact-center/decagon-triples-valuation-to-45b-with-250m-series-d/), [businesswire, glean](https://www.businesswire.com/news/home/20250610090029/en/Glean-Raises-$150M-Series-F-at-$7.2B-Valuation-to-Accelerate-Enterprise-AI-Agent-Innovation-Globally), [thenextweb, profound](https://thenextweb.com/news/profound-180m-series-d-ai-answer-engine-optimisation), [techcrunch, fora](https://techcrunch.com/2026/07/16/ai-powered-travel-agency-fora-hits-unicorn-status-raises-60m/), [fortune, bland](https://finance.yahoo.com/technology/ai/articles/exclusive-voice-ai-startup-bland-130000952.html), [vapi](https://finance.yahoo.com/sectors/technology/articles/vapi-raises-50m-series-b-130000518.html). - the connecting layers: [wikipedia, printing press](https://en.wikipedia.org/wiki/Printing_press), [electrical telegraph](https://en.wikipedia.org/wiki/Electrical_telegraph), [invention of the telephone](https://en.wikipedia.org/wiki/Invention_of_the_telephone), [app store](https://en.wikipedia.org/wiki/App_Store_(Apple)). the heights in the staircase are my reading. people using ai tools in 2026: menlo ventures, below. - the hard problems: [nvidia, the ai five layer cake](https://blogs.nvidia.com/blog/ai-5-layer-cake), [venturebeat on altman's reflections](https://venturebeat.com/ai/openai-begins-2025-with-massive-hype-for-agi-superintelligence), [techcrunch on amodei](https://techcrunch.com/2025/05/22/anthropic-ceo-claims-ai-models-hallucinate-less-than-humans/), [cx today on nadella](https://www.cxtoday.com/data-analytics/microsoft-ceo-ai-agents-will-transform-saas-as-we-know-it/), [nbc news on hassabis](https://www.nbcnews.com/business/business-news/ai-can-match-humans-task-will-five-10-years-google-deepmind-ceo-says-rcna196771), [the new stack on karpathy](https://thenewstack.io/openai-co-founder-ai-agents-are-still-10-years-away/), [the decoder on sutskever at neurips 2024](https://the-decoder.com/?p=19700) and [his nov 2025 interview](https://forum.effectivealtruism.org/posts/iuKa2iPg7vD9BdZna/highlights-from-ilya-sutskever-s-november-2025-interview), [euronews on lecun](https://euronews.com/next/2025/11/20/french-godfather-of-ai-yann-lecun-confirms-he-is-leaving-meta-to-launch-ai-start-up), [the new stack on a2a](https://thenewstack.io/google-donates-the-agent2agent-protocol-to-the-linux-foundation/), [openai adopting mcp](https://shellypalmer.com/2025/03/openai-and-anthropic-play-nice-its-a-big-deal-for-agents/), [openai, memory](https://openai.com/blog/memory-and-new-controls-for-chatgpt), [anthropic, memory](https://claude.com/blog/memory), [techcrunch on prompt injection](https://techcrunch.com/2025/12/22/openai-says-ai-browsers-may-always-be-vulnerable-to-prompt-injection-attacks/). the spectrum's left to right order is my reading. - tcp/ip: [ethw, the tcp milestone](https://ethw.org/Milestones:Transmission_Control_Protocol_(TCP)_Enables_the_Internet,_1974), [computer history museum, the 1977 test](https://computerhistory.org/blog/born-in-a-van-happy-40th-birthday-to-the-internet/) and [darpa](https://www.darpa.mil/about/innovation-timeline/tcp-ip). - the web: [cern, a short history of the web](https://home.cern/science/computing/birth-web/short-history-web), and [cybercultural, hypertext 91](https://cybercultural.com/p/1991-tim-berners-lee-trys-to-convert-the-hypertext-faithful/). - mcp: [anthropic, donating the model context protocol](https://www.anthropic.com/news/donating-the-model-context-protocol-and-establishing-of-the-agentic-ai-foundation).