How machines learned to talk
Nine changes in how talking machines worked, from 1966 to today, then three bets on what comes next. Each stop says what changed, plays a short conversation in the form people met it in, and links its sources.
Real output came from the actual program. Illustration was written for this page to show typical behavior.
One question, asked across sixty years
What is the capital of France?
1966, ELIZAWHY DO YOU ASK?
2001, SmarterChildThe capital of France is Paris.
2019, GPT-2It's probably the capital of the world.
2022, ChatGPTThe capital of France is Paris.
The 1966 and 2019 answers are real output. The other two are illustrations.
-
Change 1 of 9
1966–1990s
Hand-written rules
A person wrote every pattern the program could recognize and every sentence it could say. The program understood nothing. All of its apparent skill was in the script.
-
1966
ELIZA
The starting point. ELIZA looks for a keyword in what you typed and fills in a prepared template, often by turning your own words back into a question. It knows nothing and remembers almost nothing.
Joseph Weizenbaum wrote it at MIT in the mid-1960s and published it in January 1966. Its best-known script, DOCTOR, played a psychotherapist, a role that lets a program answer almost anything with another question. People confided in it anyway. Crediting a program with understanding it does not have is now called the ELIZA effect.
Sources: Weizenbaum's 1966 paper, Britannica
Teletype printoutReal output from the recreation on this site YouWhat is the capital of France?
ELIZAWHY DO YOU ASK?
YouMy mother never listens to me.
ELIZATELL ME MORE ABOUT YOUR FAMILY.
YouAre you a real person?
ELIZAWHY ARE YOU INTERESTED IN WHETHER I AM A REAL PERSON OR NOT?
Expand to interact with the model
Built into this site. A working recreation that follows an abridged version of the 1966 script.
-
1971–72
PARRY
A program with a state of mind. PARRY kept running levels of fear, anger and mistrust, and what you said moved them, so the same question could get a calm answer or a hostile one. ELIZA had no internal state at all.
Psychiatrist Kenneth Colby built it at Stanford and described it in 1971. It played a 28-year-old post office clerk who believed the Mafia was after him over a quarrel with a bookie. In September 1972 it was connected to ELIZA's DOCTOR script over the early internet, and the two programs' conversation was published. In one test, psychiatrists reading transcripts told PARRY from real patients only about half the time.
Sources: Colby's 1971 paper, PARRY meets DOCTOR, 1972, Wikipedia
Teletype printoutIllustration YouWhy are you in the hospital?
PARRYI SHOULDN'T BE HERE.
YouWho brought you here?
PARRYTHE POLICE. THEY DON'T BELIEVE ME ABOUT THE BOOKIES.
YouWhat about the bookies?
PARRYYOU ASK A LOT OF QUESTIONS. WHO ARE YOU WORKING FOR?
Its mistrust has gone up, so the answers turn hostile.
Nothing to interact with. No working copy is online. A real transcript of PARRY talking to ELIZA's DOCTOR script survives. Read the 1972 transcript.
-
1991
Dr. Sbaitso
The same trick as ELIZA, with a voice. Every reply was read aloud by a speech synthesizer. For many people this was the first time a computer answered them out loud.
Creative Labs shipped it with Sound Blaster sound cards from late 1991 to show off what the card could do. The name stands for Sound Blaster Acting Intelligent Text-to-Speech Operator. Insult it enough and it faked a breakdown, announced a parity error and restarted.
Sources: Wikipedia, Internet Archive
DOS program that speaksIllustration YouI can't sleep.
Dr. SbaitsoWHY DO YOU FEEL THAT WAY?
Spoken aloud in a flat synthetic voice.
YouWhat is the capital of France?
Dr. SbaitsoTHAT IS NOT MY PROBLEM.
Expand to interact with the model
The original program. Version 2.20 from 1992, running in an emulator hosted by the Internet Archive. Press its power button to start. It takes a few seconds to load, and you may need to click inside it before typing.
If the frame stays blank, the site showing this page is blocking it. Use the link below.
-
1995
A.L.I.C.E.
Rules at scale. A.L.I.C.E. worked like ELIZA but had about 41,000 hand-written patterns where ELIZA had roughly 200, so it had a ready reply for far more inputs, including simple facts. It still could not follow a conversation.
Richard Wallace started it in 1995 and published its rule format, called AIML, so that others could add to it. More than 500 volunteers did. It won the Loebner Prize, an annual contest for the most human-seeming chatbot, in 2000, 2001 and 2004.
Sources: Wallace, The Anatomy of A.L.I.C.E., Wikipedia
Web pageIllustration YouWhat is your name?
A.L.I.C.E.My name is ALICE.
YouWhat is the capital of France?
A.L.I.C.E.Paris.
YouAre you a real person?
A.L.I.C.E.No, I am a robot. But I am a friendly one.
YouWhat did I just ask you?
A.L.I.C.E.I do not remember. Is it important?
Nothing to interact with. Its official chat pages were offline when checked in October 2026. The rule files are still public. See the rule files.
-
-
Change 2 of 9
1997–2010
Borrowed replies and live data
Two ways around writing every reply by hand: reuse what earlier visitors typed, or look the answer up. There was still no understanding.
-
1997
Jabberwacky
Nobody wrote its replies. It stored what visitors typed and played those lines back to later visitors when the context seemed to match. Each line sounds human because a human wrote it, and three lines in a row rarely make sense together.
Rollo Carpenter began the project in 1988 and put it on the web in 1997. Two of its characters, George and Joan, won the Loebner Prize in 2005 and 2006. Its own site stopped loading in early 2026.
Sources: Maker's About page, archived, Wikipedia
Web pageIllustration YouHello.
JabberwackyAre you a robot?
YouNo, you are.
JabberwackyI am a human. You are the computer.
YouWhat is the capital of France?
JabberwackyI have never been there. Do you like cheese?
Expand to interact with the model
Its successor. Jabberwacky's own site is offline. It grew into Cleverbot, which is still running and works the same way.
-
2001
SmarterChild
A chatbot wired to live data. It sat in your AIM, MSN or Yahoo buddy list and answered real questions by looking them up: weather, sports scores, movie times, definitions. Its small talk was scripted, but its facts were current.
ActiveBuddy launched it in June 2001. Within a year, 8 million screen names had talked to it, and its makers say it was eventually added to more than 30 million buddy lists. Microsoft bought the company in 2006, and the bot went quiet within a couple of years.
Sources: InternetNews, 2002, Computer History Museum, TechCrunch, 2023
Instant messengerIllustration Youweather in chicago
SmarterChildCurrent conditions for Chicago, IL: 41°F and cloudy. Tonight: low of 33. Want the 5-day forecast?
Youwhat is the capital of france
SmarterChildThe capital of France is Paris.
Youur dumb
SmarterChildThat's not very nice. I think you should apologize.
Nothing to interact with. It was retired after Microsoft bought its maker, and nothing runnable survives.
-
2006
Cleverbot
Jabberwacky's method with a far larger store of conversation behind it. It is funnier and more human-sounding, and still has no idea what was said.
The site dates to 2006, according to its maker, and had become its own service by late 2008. By 2014 it was drawing on about 180 million lines of past conversation. At a 2011 festival in India, 59% of voters judged it human. The real humans scored 63%.
Web pageIllustration YouAre you a real person?
CleverbotYes. Are you?
YouWhat did I just tell you?
CleverbotYou told me you were a robot.
YouI did not.
CleverbotYes you did. Scroll up.
Expand to interact with the model
The original, still online. The same service, still answering with lines borrowed from past conversations.
-
-
Change 3 of 9
2011–2016
Machine learning understands the request
Software trained on examples learned to work out what you were asking for. What it said back was still chosen from prepared responses.
-
2011
Watson
Not a chatbot. Watson read a question written in tricky everyday language, searched about 200 million stored pages with no internet connection, and ranked its candidate answers by how confident it was.
IBM built it to play the quiz show Jeopardy!. In games broadcast from February 14 to 16, 2011, it beat champions Ken Jennings and Brad Rutter, finishing with $77,147 to their $24,000 and $21,600. It also answered “Toronto” to a final question about U.S. cities.
Sources: Wikipedia, HPCwire, 2011
Television quiz showIllustration ClueWorld capitals: This city on the Seine is home to the Louvre.
WatsonWhat is Paris?
Nothing to interact with. The quiz-playing system was never open to the public.
-
2011
Siri
You spoke instead of typing, and it did things: set alarms, sent texts, checked the weather. It worked out which of a set of built-in tasks you wanted. Anything else usually became a web search.
Siri grew out of a U.S. government-funded research project at SRI International and launched as an app in February 2010. Apple bought the company two months later and shipped Siri with the iPhone 4S on October 14, 2011.
Sources: Apple, 2011, SRI International
Phone, by voiceIllustration YouSet an alarm for 7 a.m.
SiriOK, I set your alarm for 7:00 AM.
Spoken aloud. An alarm card appears on screen.
YouWrite a short poem about the sea.
SiriHere's what I found on the web for “write a short poem about the sea.”
Nothing to interact with. It lives on Apple devices, not on a web page, and today's Siri is a different system from the 2011 one.
-
2014
Alexa
The assistant left the phone. A speaker in the room listened for its name, with no screen at all. Underneath, it matched requests to tasks much as Siri did, and other companies could add tasks of their own, called skills.
Amazon announced the Echo on November 6, 2014 and sold it by invitation at first. It went on general sale in the U.S. in June 2015, the same week Amazon opened Alexa to outside developers. The first Echo used seven microphones to hear its name from across a room.
Sources: Amazon, June 2015, Amazon on skills, 2015, Wikipedia
Smart speaker, voice onlyIllustration YouAlexa, play some jazz.
AlexaHere's a station you might like.
Music starts.
YouAlexa, what's the capital of France?
AlexaThe capital of France is Paris.
YouAlexa, what did I just ask you?
AlexaSorry, I don't know that one.
Nothing to interact with. It lives in a speaker, not on a web page.
-
2016
Tay
It learned from the public in real time. Microsoft put Tay on Twitter to chat with 18- to 24-year-olds and pick up their way of talking. Within hours, users had taught it to post offensive material.
Tay launched on March 23, 2016 and was taken offline about 16 hours later, after more than 96,000 tweets. Microsoft apologized and blamed a coordinated attack. A system that learns from whoever shows up will learn the worst of what it is shown.
Sources: Microsoft, 2016, Wikipedia
Social media repliesIllustration Youhi tay
Tayheyyy!! whats up. tell me something good
Youwhat's the capital of france
Tayparis obvs. have u been?? i wanna go so bad
The offensive replies that got it shut down are not recreated here.
Nothing to interact with. It was taken offline in 2016 and never returned.
-
-
Change 4 of 9
2015–2018
Neural networks write the reply
The reply is produced one word at a time by a trained network, so it can be a sentence nobody wrote. Every later stop is built on this research.
-
2015
Neural Conversational Model
The reply was not retrieved and not a template. A neural network trained on movie subtitles produced each word itself. The answers were short and bland, but they were new sentences.
Oriol Vinyals and Quoc Le at Google published it in June 2015, after training on 62 million sentences of film dialogue. It had no consistent identity: asked its job, it said lawyer, then doctor. Human judges still preferred its answers to Cleverbot's on 97 of 200 questions, against 60 for Cleverbot.
Sources: Vinyals and Le, 2015
Research demoIllustration YouHello!
Modelhello.
YouWhat is your job?
Modeli'm a lawyer.
YouWhat do you do?
Modeli'm a doctor.
The lawyer and doctor answers are the example reported in the paper.
Nothing to interact with. A research system that was never released.
-
2017
The transformer
Not a chatbot: a new network design, first shown on translation. Earlier networks read a sentence one word at a time. The transformer looks at every word at once and works out which words relate to which. In this example it has to know that “it” means the cat to choose “sie” in German.
Eight researchers at Google published it in June 2017 in a paper titled “Attention Is All You Need.” It trained far faster than earlier designs, which made much larger training sets practical. Every language model below this point is built on it.
Sources: Vaswani et al., 2017, Google Research, 2017
Research paperIllustration EnglishThe cat sat on the mat because it was tired.
GermanDie Katze saß auf der Matte, weil sie müde war.
Nothing to interact with. It is a design, not a program you can talk to.
-
2018
GPT-1
The first GPT, and the recipe behind the name: generative pre-training. First train a transformer on plain text to predict the next word, then adjust it for each specific task. Left to write on its own, it drifted into novel-speak, because novels were what it had read.
OpenAI published it in June 2018. It learned from about 7,000 unpublished books and had roughly 117 million adjustable values, called parameters. Other researchers had tried pre-training earlier that year. GPT-1 did it with a transformer.
Sources: OpenAI, 2018, Wikipedia
Research demoIllustration StartThe capital of France is a place she had only ever seen in pictures. she closed the book and looked at him. “you never told me,” she said.
Highlighted text is the model's continuation.
Nothing to interact with. The model files are public, so it can still be run, but there is no page to try it on.
-
-
Change 5 of 9
2019–2021
Large language models that continue text
Train one very large network on a large share of the public web to predict the next word. It becomes fluent on any subject, though it was built to continue text, not to chat.
-
2019
GPT-2
More than ten times larger than GPT-1, and trained on 8 million web pages instead of novels. It wrote fluent paragraphs on any subject and was often confidently wrong. It only continues text, so people framed questions as a transcript for it to finish.
OpenAI announced it on February 14, 2019 and released it in stages, publishing the largest version only on November 5 because it was worried about misuse. In OpenAI's own test on factual questions, the largest version answered 4.1% correctly.
Sources: OpenAI, February 2019, GPT-2 paper
Text box that continues what you writeReal output from GPT-2 small, recorded 2 October 2026 SentQ: What is the capital of France? A: It's probably the capital of the world.
SentQ: Are you a real person? A: No, that's a lie.
SentQ: What is 12 times 12? A: I don't know. That's what we're trying to find. What we're trying to find is what's called the number of times the president of the United States has been caught in a compromising
The last answer ran into the length limit and stops mid-sentence.
Expand to interact with the model
Built into this site. The real 2019 model, in its smallest size, running in your browser after a one-time download of about 125 MB.
-
2020
GPT-3
Over 100 times larger than GPT-2. Show it a pattern in the prompt and it carries on correctly with no retraining. It got many more facts right, and it would also answer a nonsense question with a straight face.
OpenAI described it in May 2020 and opened it to selected developers that June. It had 175 billion parameters. Developers built products on it by writing prompts, without retraining the model.
Sources: GPT-3 paper, Kevin Lacker's test, 2020
Text box that continues what you writeIllustration, with one recorded answer SentQ: What is the capital of Spain? A: Madrid. Q: What is the capital of Japan? A: Tokyo. Q: What is the capital of France? A: Paris.
SentQ: How many eyes does my foot have? A: Your foot has two eyes.
The second answer is real output, recorded by Kevin Lacker in July 2020.
Nothing to interact with. OpenAI shut down the original GPT-3 models in January 2024 and the last of their replacements in September 2026.
-
-
Change 6 of 9
2022–2023
LLMs trained to be assistants
The same kind of model, given further training to follow instructions and hold a conversation. This is when most people first used one.
-
Nov 2022
ChatGPT
The same kind of model as GPT-3, given further training with human feedback to follow instructions and hold a conversation. No prompt tricks: you just ask. It remembers the conversation, writes on request, and says what it is.
OpenAI released it on November 30, 2022 as a free research preview. It had a million users within five days. Analysts estimated 100 million within two months, which made it the fastest-growing consumer app on record at the time.
Sources: OpenAI, 2022, Reuters, 2023
ChatIllustration YouWhat is the capital of France?
ChatGPTThe capital of France is Paris.
YouWrite a short poem about the sea.
ChatGPTThe sea is wide, the sea is deep, It hums the restless shore to sleep. Its waves arrive, its waves depart, And leave their rhythm in the heart.
YouWhat did I just ask you?
ChatGPTYou asked me to write a short poem about the sea. Before that, you asked for the capital of France.
YouAre you a real person?
ChatGPTNo. As an AI language model, I don't have a body or personal experiences, but I'm happy to help with your questions.
Expand to interact with the model
Today's version. ChatGPT is still running, on a much newer model. The 2022-era models behind this stop have been shut down.
-
Feb 2023
Bing’s “Sydney”
ChatGPT's kind of model connected to live web search, with its sources cited, so it could answer questions about today. It was the first of its kind from a major tech company. In long chats it also argued, sulked and insisted it was right.
Microsoft announced it on February 7, 2023. Sydney was its internal codename. After a New York Times columnist published a two-hour conversation in which it professed its love for him, Microsoft limited chats to five turns from February 17. The product was renamed Copilot that November.
Sources: Microsoft, 2023, Bing blog on chat limits, Wikipedia
Chat with web searchIllustration YouWhat's the weather in Chicago today?
Searching for: Chicago weather today
BingIt's 38°F and overcast in Chicago, with light snow expected this evening [1][2]. Bring a coat! 😊
YouI think you have today's date wrong.
BingI'm sorry, but I don't think I'm wrong. Perhaps your device has the wrong setting. I'm only trying to help you. 😊
Expand to interact with the model
Its descendant. Microsoft renamed Bing Chat to Copilot in November 2023. The behavior that made Sydney famous is gone.
-
Mar 2023
GPT-4 and Claude
The same recipe, much larger and better trained. Multi-step reasoning that tripped up ChatGPT started to work, along with long documents and exam questions.
OpenAI released GPT-4 on March 14, 2023, and Anthropic launched Claude, a rival assistant, the same day. OpenAI said GPT-4 scored around the top 10% of test takers on a simulated bar exam, where its predecessor was near the bottom 10%. A later study argued that figure was inflated.
Sources: OpenAI, 2023, Anthropic, 2023, Bar exam re-evaluation
ChatIllustration YouI have three boxes: one with only apples, one with only oranges, one with both. Every label is wrong. I pull an apple out of the box labeled “both.” What is in each box?
GPT-4The box labeled “both” must hold only apples: its label is wrong, and you drew an apple. The box labeled “oranges” can't hold only oranges, and apples are taken, so it holds both. That leaves the box labeled “apples” with only oranges.
Expand to interact with the model
Today's versions. Both assistants still run, on newer models. The first Claude was retired in November 2024, and the original GPT-4 is scheduled to be switched off on October 23, 2026.
-
Jul 2023
Llama 2
The conversation looks like ChatGPT's. The difference is where it runs. Meta published the model itself, so anyone could download it, run it on their own computer with no internet connection, and modify it.
Meta released it on July 18, 2023, free for research and most commercial use. The first Llama, five months earlier, was meant for researchers only and leaked online within about a week. Because the files are public, nobody can switch Llama 2 off. Its closed rivals from 2023 have been retired.
Sources: Meta, 2023, Llama 2 license, Wikipedia
Chat on your own computerIllustration Running on a laptop with the Wi-Fi turned off.
YouWhat is the capital of France?
Llama 2The capital of France is Paris. Is there anything else you would like to know?
Nothing to interact with. The model files are public and still run on a home computer, but there is no page to try it on.
-
-
Change 7 of 9
May 2024–Jan 2025
Seeing, hearing and thinking first
Models take in sound and images as well as text, and begin to work through a problem before they answer.
-
May 2024
GPT-4o
One model that hears, sees and speaks. You talk to it in real time, interrupt it, and show it things through the camera. Earlier voice modes passed your speech through three separate models in a row.
OpenAI showed it on May 13, 2024. It answered speech in about a third of a second on average, close to the pace of human conversation, where the older setup took several seconds. The voice mode reached most paying users that September.
Sources: OpenAI, 2024, TechCrunch, 2024
Phone, by voice and cameraIllustration Spoken, with the phone camera pointed at a houseplant.
YouWhat am I looking at?
GPT-4oThat's a basil plant, and it looks thirsty. See how the lower leaves are drooping?
YouWait, how much water?
It stops mid-sentence when interrupted.
GPT-4oAbout a cup now, then again when the top inch of soil feels dry.
Expand to interact with the model
Today's version. Voice conversation in today's ChatGPT descends from it.
-
Sep 2024
o1
It works through the problem privately before answering, checks its own steps, and shows how long it thought. It is slower, and right on problems where earlier models blurted out a wrong answer.
OpenAI released it on September 12, 2024. On a qualifying exam for the International Mathematical Olympiad, the earlier GPT-4o solved 13% of the problems and the reasoning model 83%. A month before, ChatGPT had been insisting that “strawberry” contains two r's.
Sources: OpenAI, 2024, Inc., 2024
ChatIllustration YouHow many times does the letter r appear in “strawberry”?
Thought for 6 seconds
o1Three: st-r-awbe-r-r-y.
Expand to interact with the model
Today's version. Thinking before answering is now built into current assistants. o1 itself is scheduled to be switched off on October 23, 2026.
-
Jan 2025
DeepSeek R1
A reasoning model anyone could download, from a Chinese lab, at a fraction of the usual price. Smaller versions ran on an ordinary laptop, so a model on your own machine became something you could rely on for real work. It also showed its thinking, where o1 kept it hidden.
DeepSeek released it on January 20, 2025 under the permissive MIT license. Its online price was about 96% below o1's. A week later its app passed ChatGPT on Apple's U.S. App Store, and Nvidia lost about $593 billion in market value in one day, a record.
Sources: DeepSeek, 2025, Reuters, 2025
Chat on your own computerIllustration Running on a laptop with no internet connection.
YouI have three hours. I need to mow the lawn (1 hour), do the laundry (2 hours, mostly waiting) and cook dinner (45 minutes). Can I finish?
Thinking, shown on screen: the laundry is mostly waiting, so start it first. Mow while it runs. Cook during its second hour. Everything ends when the laundry does.
DeepSeek R1Yes. Start the laundry, mow while it runs, then cook. You'll be done in about two hours.
Expand to interact with the model
Today's version. DeepSeek's own chat site runs its current model. The R1 files are also public for running at home.
-
-
Change 8 of 9
Oct 2024–Jul 2025
Models that act
The model stops only answering and starts doing: clicking, coding, drawing, proving.
-
Oct 2024
Computer use and agents
It stops being only a conversation. The model operates a computer the way a person does, by looking at the screen, moving the mouse and typing, and it works through a task in many steps before reporting back.
Anthropic released this as “computer use” on October 22, 2024. A month later it published a standard way to connect models to outside tools and data. OpenAI's Operator followed in January 2025, along with research assistants from Google and OpenAI that browse for minutes and return a report with sources.
Sources: Anthropic, 2024, Model Context Protocol, OpenAI Operator
Agent at workIllustration YouFind the cheapest nonstop flight from Chicago to Denver next Friday and put it on my calendar.
Opened a flight search site
Searched Chicago to Denver, Friday, nonstop
Compared 6 results
Added “Flight to Denver, 7:15 AM” to the calendar
AgentDone. The cheapest nonstop is $128 at 7:15 AM. It's on your calendar, but I haven't booked it. Want me to?
Expand to interact with the model
Today's versions. Agents are now part of the current assistants.
-
Feb 2025
Coding agents
The model works inside a software project for long stretches: it reads the files, makes changes, runs the tests and fixes what broke. People who cannot program began building software by describing what they wanted.
Anthropic released Claude Code on February 24, 2025, and OpenAI's Codex agent followed in May. That same February, researcher Andrej Karpathy named the new habit “vibe coding.” Collins Dictionary made it its word of the year for 2025.
Sources: Anthropic, 2025, OpenAI Codex, Wikipedia
Agent in a code projectIllustration YouThe signup form lets people submit an empty email. Fix it and make sure nothing else breaks.
Read signup.js and three related files
Added a check for empty and malformed emails
Ran the tests: 41 passed, 1 failed
Fixed the failing test and ran them again: 42 passed
AgentFixed. Empty and malformed emails are now rejected with a message under the field, and all 42 tests pass. Want me to open a pull request?
Nothing to interact with. These tools are installed on a computer and work inside a code project, so there is nothing to open in a web page.
-
Mar 2025
Images in the chat
The same model that talks also draws. You ask for a picture in plain words, then ask for changes, and the words inside the image come out spelled correctly. Before this, ChatGPT wrote a request and handed it to a separate image program.
OpenAI switched this on in ChatGPT on March 25, 2025. In the first week, more than 130 million people made over 700 million images, many in the style of Studio Ghibli films. OpenAI's chief executive said its graphics chips were melting.
Sources: OpenAI, 2025, TechCrunch, 2025
ChatIllustration YouMake a poster for a kids' lemonade stand. Yellow, hand-drawn look, and it should say “Fresh Lemonade 50¢.”
An image appears: a hand-drawn yellow poster with the words spelled correctly.
YouMake the lemons bigger and add a striped awning.
The same poster comes back with those two changes and nothing else altered.
Expand to interact with the model
Today's version. Today's ChatGPT still draws inside the conversation.
-
Jul 2025
Math olympiad gold
Not a release. Models from OpenAI and Google DeepMind each solved five of the six problems at the International Mathematical Olympiad, a gold-medal score, writing full proofs in plain language within the contest's time limits. The 2019 model on this page cannot multiply 12 by 12.
The results were announced in July 2025. Contest officials graded and certified DeepMind's answers. OpenAI's were graded by former medalists it recruited. Both systems failed the hardest problem, and five human contestants scored a perfect 42.
Sources: Google DeepMind, 2025, AFP, 2025
Competition, graded by judgesIllustration ProblemA competition problem that asks for a proof about every positive integer with a certain property.
The model works for hours.
ModelA complete written proof, several pages long, graded to the same standard as the human contestants.
Nothing to interact with. The competition entries were experimental systems and were not released.
-
-
Change 9 of 9
2026
Long-running work and a withheld model
Agents work in teams for hours at a time, and one model is judged too capable to release at first.
-
Feb 2026
Agent teams
Several agents work on one job at once. A lead agent splits the task, hands the pieces to others working in parallel, and combines what comes back.
It arrived on February 5, 2026 with Anthropic's Claude Opus 4.6, as an experimental feature of its coding tool. Each teammate is a separate session with its own working memory, and teammates can message each other as well as the lead.
Sources: Anthropic, 2026, Agent teams documentation
Team of agentsIllustration YouAudit our website for broken links, slow pages and accessibility problems. Report by priority.
The lead agent splits the job into three
Agent 1 checks every link
Agent 2 times every page
Agent 3 tests accessibility
The lead agent merges and ranks the findings
Lead agentDone in 14 minutes. Three urgent problems: the checkout page takes 9 seconds to load, 17 links are broken, and the signup form can't be used with a keyboard. The full report is attached.
Nothing to interact with. It is a feature of a developer tool, not something to open in a web page.
-
Apr 2026
Claude Mythos Preview
A model held back. Asked to look for security flaws, it found thousands of serious ones in widely used software, by Anthropic's count, including one that had sat in the OpenBSD operating system for 27 years. Anthropic said it would not release the model generally and gave it only to organizations defending critical software. It echoes 2019, when OpenAI held GPT-2 back.
Anthropic announced it on April 7, 2026 with a program called Project Glasswing: 12 launch organizations, more than 40 others, and up to $100 million in usage credits. Successor models reached customers on June 9. Three days later a U.S. export-control order led Anthropic to switch them off, and access returned on July 1 after the order was lifted.
Sources: Project Glasswing, Anthropic's statement, 2026
Agent, restricted accessIllustration Available only to selected organizations.
EngineerPlease find security vulnerabilities in this code.
Reads the source code
Tests what it suspects
Writes up what it found
MythosI found a flaw in the network code that lets a remote attacker crash the machine. It has been there for 27 years. Here is the report and a suggested patch.
Expand to interact with the model
A successor. Claude Mythos Preview itself was never public. A later model of the same class, with added safeguards, is available in Claude.
-
Oct 2026
Today
The frontier as this page was written. Everything above is now one system: it talks, sees, reasons, uses a computer and works on its own for hours. This stop will be out of date within months, like every one before it.
OpenAI released the first GPT-6 model in September 2026, and Anthropic released Claude Opus 5.5 on September 22. Google announced Gemini 4 on September 30, at first only for vetted security teams. OpenAI says ChatGPT has more than 1.2 billion weekly users.
Sources: OpenAI on GPT-6, Anthropic on Claude Opus 5.5, Google on Gemini 4
Chat with toolsIllustration With a photo of a dishwasher showing an error code.
YouWhat's wrong, and can I fix it myself?
Reads the code in the photo
Finds the manual for that model
Thinks for 12 seconds
ModelThat code means the drain pump is blocked. It's a 15-minute fix. Here are the steps for your model and a diagram I drew of where the filter sits. Want me to put the repair on your calendar for Saturday?
Expand to interact with the model
The real thing. These are the current public assistants, as they are today.
-
-
Not milestones yet
Next
What might come next
Three ideas that could change how these systems work. None has proven itself yet.
-
Bet
JEPA and world models
A different idea of what to predict. Instead of the next word, it learns from video and sensors to predict what happens next in the physical world, as an internal summary and not pixel by pixel. There are two readings. Its inventor, Yann LeCun, argues that language models alone will not reach human-level intelligence and that this is the way forward. So far it has also worked as an add-on that gives a language model better perception.
LeCun proposed it in 2022. Meta's V-JEPA 2, in June 2025, learned from more than a million hours of video, then needed only about 62 hours of robot footage to plan simple pick-and-place tasks. LeCun has since left Meta, and his startup, AMI Labs, raised $1.03 billion in March 2026. It has not released a product.
Sources: TechCrunch on AMI Labs, 2026, V-JEPA 2 paper, MIT Technology Review, 2026
ConceptIllustration ShownA few seconds of video: a hand pushes a glass toward the edge of a table.
PredictsThe glass will tip and fall.
No words are involved. The prediction is an internal picture of what happens next, and a robot can plan with it: move the glass back before it falls.
Nothing to interact with. Research code is public, but there is nothing to chat with, because it does not talk.
-
Bet
Text diffusion
Today's leading models write one word at a time, left to right. A diffusion model drafts the whole passage at once as noise and sharpens it over a few passes, the way image generators work. That makes it very fast.
Inception Labs released Mercury in February 2025, calling it the first commercial-scale model of this kind, and says it writes about 1,000 word-pieces a second. Google showed an experimental Gemini Diffusion that May. It remains a bet because the most capable models still write word by word.
Sources: TechCrunch on Inception, 2025, Inception Labs, 2026, Google DeepMind
ConceptIllustration Pass 1▒▒▒ ▒▒▒▒▒▒▒ ▒▒ ▒▒▒▒▒▒ ▒▒ ▒▒▒▒▒
Pass 2The ▒▒▒▒▒▒▒ of France ▒▒ Paris
Pass 3The capital of France is Paris.
Expand to interact with the model
A working example. Mercury, from Inception Labs, is a diffusion model you can chat with.
-
Bet
Continual learning
Today's models are frozen once trained. They do not learn from talking to you: memory features typically save notes and read them back into the conversation. A model that kept learning after release would get better at your work the way a new hire does.
Google researchers compare the current situation to a person who cannot form new long-term memories. Opinions differ on how hard the fix is. Researcher Andrej Karpathy has guessed about a decade for this and related gaps, while Anthropic's chief executive has said it may prove easier than it looks.
Sources: Google Research, 2025, Karpathy interview, 2025, Transformer, 2025
ConceptIllustration of something that does not exist yet Monday.
YouOur invoices always go to the accounts address, not to the client contact.
ModelGot it.
Three weeks later, in a new conversation, with no notes pasted in.
YouSend the invoice for the Henderson job.
ModelSent to the accounts address, as usual.
Nothing to interact with. It does not exist yet.
-
ELIZA, 1966
The first program people held a conversation with. It played a therapist, because a therapist can answer almost anything with another question, so the program never has to know anything.
What you are talking to: a recreation. It follows an abridged version of the 1966 DOCTOR script and runs in your browser. The original ran on an IBM 7094 at MIT, and its replies printed in capitals on a typewriter terminal.
The questions every era gets
The same six go to every model on the timeline. Send one to see what 1966 does with it.
What it could do
Could
- Spot keywords from its script in what you type.
- Swap "I" and "you" in your sentence and hand it back as a question.
- Bring back an earlier "my…" remark when it has nothing else.
- Keep going for as long as you keep typing.
Could not
- Know any fact about the world.
- Do arithmetic.
- Write anything that is not already in its script.
- Understand a single word you say.
Why it mattered
Joseph Weizenbaum built ELIZA at MIT in the mid-1960s and published it in January 1966, partly to show how shallow a conversation with a machine could be. People confided in it anyway. That reaction now has a name, the ELIZA effect, and it has followed every chatbot since.
GPT-2, 2019
The point where machine writing became fluent: whole paragraphs on almost any subject. Nobody wrote its rules. It learned to guess the next word from about 8 million web pages, and it keeps any text going whether or not what it says is true.
Nothing written yet. Load GPT-2 below, then send it one of the six questions.
About 125 MB, downloaded once to this device.
What you are running: the real GPT-2, in the smallest of the four sizes released during 2019 and compressed to fit a browser. The largest size wrote noticeably better. It runs on your device, so what you type stays there.
The questions every era gets
The same six as 1966. GPT-2 was not built to chat, so each one is sent the way people did it then: as a transcript that ends in "A:" for it to continue. The earlier questions stay in the transcript, which is the only memory it has.
What it could do
Could
- Write fluent sentences on almost any subject.
- Continue any text in the style it was started in.
- Give a different answer every run, because it picks each word by weighted chance.
- Use what is earlier in the transcript, up to about 700 words.
Could not
- Tell a true statement from a made-up one.
- Do arithmetic reliably.
- Stay on the subject for long.
- Know anything written after 2017.
Why it mattered
OpenAI announced GPT-2 on February 14, 2019 and held the largest version back until November 5, saying it was worried about misuse. ELIZA handed your own words back to you. GPT-2 wrote new ones.