Modern AI models are not inventions. They are descendants.
Every model is a phenotype — a body assembled from memes: ideas that were born, mutated, recombined, competed, and either thrived or died under enormous technical and economic pressure. Origin of Models is the interactive natural history of those ideas — attention, RoPE, MoE, RLHF, DPO, FlashAttention, test-time compute, latent diffusion, video world models — and of the dead ends along the way.
Models are phenotypes
Every model — from GPT-1 to DeepSeek-V4, from Sora to Genie — is a body that a set of ideas built. Architecture, training data, alignment, scaling, vision, video, and the tools around it: the model is where the ideas manifest.
Ideas are memes
Ideas are genes. They copy themselves from paper to model to model, mutate under new conditions, recombine with other ideas, and compete for survival. Some spread everywhere; some evolve into something new; some die.
Pressure accelerates evolution
The evolutionary pressure here is economics — a capability that makes money gets copied, refined, and scaled at astonishing speed. Explosive adoption turned a research niche into one of the fastest idea-evolution engines in history.
Every idea has a story — including the dead ends. The ideas that failed teach as much as the ones that won. Each meme card carries an essence for newcomers, a technical deep-dive, an impact rating, and citations that open in new tabs. Nothing here is a black box: everything traces to real papers, reports, and analyses.
The Evolution of Language Models
Every model is a phenotype — a body that a set of memes (ideas) built together. Scroll through time, play the clock, and watch ideas bloom, merge, mutate, and die as they travel from model to model. Click any node to meet the model; click a colored dot to meet the idea behind it.
The Meme Library
Ideas are the genes of the models genome. Browse every meme that built the field — the ones that thrived, the ones that evolved into something else, and the ones that were tried and died. Filter by domain, by fate (status), by era, or by the lab that drove it. Click any card for the full deep-dive.
The Genealogy of Ideas
Two family trees. The idea tree traces how each meme descends from the ideas before it — who begat whom, and where the line went extinct. The model tree shows how phenotypes descend from one another and which families dominate the field.
Scaling: the engine of the whole game
Ideas get the credit, but scaling is the engine — compute, data, and the engineering that makes each generation of model physically possible. This is the layer that unlocks everything else: the scaling laws that guide it, the parallelism and precision tricks that make it affordable, and the economics that decide who gets to play.
Frontier training compute, 2017 → 2026
The scaling memes — ideas that unlock size
The scaling narrative
Harnesses & Tools: the adoption layer
Models are raw capability. The harness — serving engines, fine-tuning tooling, evaluation harnesses, agent protocols, RAG stacks, model hubs, and developer tools — is what turns them into products people can actually use. This is the compounding layer: every tool makes the next tool possible, and together they decide how fast, cheap, and easy it is to adopt LLMs.
The adoption waves
The tool memes — the harness idea library
The tools timeline — when each harness arrived
The people behind the memes
Ideas don't evolve themselves — people do. Every meme in this catalog was born in someone's head, carried by a team, and spread through a community. This view maps the humans: who introduced what, which labs were the households of ideas, and how people and memes form a collaboration network.
The pivot explorer
Same data, your question. Group the timeline's memes and models by lab, category, fate, or era, slice by year, and color by whatever facet you care about. Click any cell to meet the ideas or models behind it.
Ideas AI tried — and abandoned
Evolution is written by the winners, but the losers teach the most. Every idea below was born, tried at scale, and lost — to economics, to a better idea, or to the environment changing. Each fossil records Born → Tried → Why it lost → What replaced it.
Methodology & corrections
What we mean by “origin”, how we verify (and mark what we can't), how genealogy edges are chosen, and how to correct the record.
Privacy Policy
What information Origin of Models receives, why it is used, and the choices available to you.
Last updated: 26 August 2026
1. Who operates this site
Origin of Models is operated by Eran Shir (“Origin of Models”, “we”, “us”). For privacy questions or requests, contact [email protected].
2. Information you provide
If you post an annotation or reply, we receive the display name and content you submit. That information is published publicly with the relevant passage. If you submit a correction, suggestion, claim, or email, we receive the information included in that submission, including optional contact details.
Do not post private, confidential, identifying, or sensitive information about yourself or anyone else.
3. Information collected automatically
The site creates an anonymous browser identifier in local storage (oom-ann-voter) to remember Agree/Disagree reactions and may remember your chosen display name (oom-ann-author). Requests may include technical information such as IP address, browser user-agent, timestamps, requested URLs, and standard security or diagnostic data. IP addresses may be used temporarily for abuse prevention and rate limiting; raw IP addresses are not stored in annotation records.
4. How information is used
We use information to operate community features, publish and associate annotations with passages, remember reactions, respond to requests, prevent spam and abuse, maintain security, diagnose problems, improve the site, and comply with legal obligations.
5. Public content
Annotations, replies, display names, reactions, and timestamps are public. Public content may be copied, indexed, quoted, or archived by others. Removing content from this site may not remove copies held elsewhere.
6. Service providers and international processing
The site is hosted using Cloudflare Pages, Functions, and KV services. Cloudflare and other infrastructure providers may process technical data on our behalf in locations outside Israel. Their handling of data is governed by their own terms and privacy practices.
7. Retention and removal
Public contributions may be retained for as long as the community feature operates, including for integrity, abuse prevention, and recordkeeping. Security and operational records may be retained for reasonable periods. To request access, correction, or deletion of information associated with you, email [email protected] with enough detail to identify the content. We may need to verify the request and may retain information where legally permitted or reasonably necessary.
8. Your choices
You can avoid community features, remove local-storage entries through your browser, and refrain from providing optional contact information. Disabling local storage may prevent reactions or remembered names from working properly.
9. Children
The community features are not directed to children under 16. Do not submit personal information if you are not legally permitted to do so.
10. Security and changes
No internet service can guarantee absolute security. We may update this policy as the site changes. Material updates will be reflected by the date above. This policy is intended as a clear operational disclosure and is not legal advice.
Terms of Service
The rules for using Origin of Models and participating in its community annotation layer.
Last updated: 26 August 2026
1. Acceptance
By accessing the site or submitting content, you agree to these Terms and the Privacy Policy. If you do not agree, do not use the community features.
2. Informational service; no professional advice
Origin of Models is an educational and editorial project. Content may be incomplete, outdated, estimated, disputed, or incorrect. Nothing on the site is legal, financial, investment, medical, security, academic, or other professional advice. You use the site and rely on its content at your own risk.
3. Community content is not pre-moderated
Annotations and replies are user-generated and generally appear without review before publication. They may be inaccurate, misleading, offensive, inappropriate, or unlawful and do not represent the views of Origin of Models or Eran Shir. We do not endorse, verify, or assume responsibility for user submissions.
We may—but are not obligated to—review, hide, edit, preserve, reject, or remove content, or restrict access, at any time and without notice. The absence of removal does not mean that content has been approved.
4. Community Conduct: be thoughtful, polite, and constructive
Critique ideas, evidence, and arguments—not people. Be cordial, make claims in good faith, and add useful context. Do not submit harassment, personal attacks, threats, hate speech, sexually exploitative content, impersonation, spam, doxxing, private information, malicious links, copyright infringement, deliberately deceptive material, illegal content, or content intended to disrupt the service.
Violating content may be removed. Repeat or serious abuse may result in access being restricted or blocked without warning. Credible threats or apparently illegal activity may be preserved and reported where appropriate.
5. Your submissions and licence
You are responsible for your submissions and represent that you have the right to post them. You retain ownership of your content. By submitting it, you grant us a worldwide, non-exclusive, royalty-free licence to host, store, reproduce, format, display, moderate, and distribute it as reasonably necessary to operate, promote, archive, and improve the site.
6. No guarantee of publication or preservation
We do not guarantee that submissions will be published, remain available, receive responses, or be preserved. Features may change or be discontinued, and content may be lost or removed.
7. Intellectual property
The site's original design, organization, editorial text, and software are protected by applicable intellectual-property laws. Third-party names, papers, trademarks, and linked materials belong to their respective owners. You may link to the site and quote limited portions with attribution, but may not misrepresent affiliation or reproduce substantial original content in violation of applicable law.
8. Disclaimer of warranties
To the maximum extent permitted by law, the site is provided “as is” and “as available”, without warranties of accuracy, reliability, availability, security, fitness for a particular purpose, or non-infringement.
9. Limitation of liability
To the maximum extent permitted by applicable law, Eran Shir and Origin of Models will not be liable for indirect, incidental, consequential, special, exemplary, or punitive damages, loss of data, loss of profits, reputational harm, or damages arising from reliance on site or community content. Nothing in these Terms excludes liability that cannot lawfully be excluded.
10. Indemnity
To the extent permitted by law, you agree to be responsible for claims and reasonable costs arising from your unlawful submission, your infringement of another person's rights, or your material breach of these Terms.
11. Israeli law and courts
These Terms are governed by the laws of the State of Israel, without regard to conflict-of-law rules. Subject to any mandatory consumer rights, the competent courts in Israel will have exclusive jurisdiction over disputes relating to the site or these Terms.
12. Changes and contact
We may update these Terms. Continued use after an update constitutes acceptance of the updated Terms. Questions, complaints, and removal requests may be sent to [email protected].
These Terms are a practical operating policy, not legal advice. Because legal enforceability depends on facts and applicable law, they should be reviewed by qualified Israeli counsel.
Suggest an idea
We cover 257 ideas — but the field is bigger than any catalog. Nominate an idea, model, person, or lab we missed. Every submission must carry a reference (arXiv link, GitHub repo, or paper/blog post) so we can verify it before inclusion.
Claim verification
Verifying your claim on an idea.
Checking…
Go down the rabbit hole
A random high-quality idea, dead end, or model from the tree. No two visits the same.