Foundation Models · since 2020
LLMs
Large language models — transformer-based models trained at scale to generate and reason over text. The product layer that turned the architecture into a consumer phenomenon, and the substrate every agentic and multimodal development now builds on.
35
events traced
37
source records
30 Jun 2026
first signal
11 Jul 2026
last activity
Who drove it
Key movements
Nov 2022★
public inflection (ChatGPT)
175B → ~T↑
frontier scale
900M+↑
weekly active users (ChatGPT, Feb 2026)
The story, event by event
Every point below is traced to a real source — nothing on this page is invented.
11 Jun 2020impact 84
GPT-3 API opens the platform era
A 175-billion-parameter model made available behind a paid API turned a research architecture into a product surface anyone could build on.
GPT-3 scaled the GPT recipe to 175 billion parameters with a 2,048-token context window and showed strong few-shot performance — new tasks learned from a handful of examples in the prompt, with no gradient updates at all.
OpenAI shipped GPT-3 through a commercial API rather than a paper-only release, which meant the model's capability could be rented by any developer — the moment a research result became a platform other companies built products on top of.
This is the commercial origin point of the LLM-as-platform era — every API-first product built on a foundation model traces back to this decision.
4 Mar 2022impact 82
InstructGPT — alignment makes LLMs usable
Fine-tuning GPT-3 with human feedback produced a model that follows instructions and refuses harmful requests instead of just completing text.
OpenAI applied reinforcement learning from human feedback (RLHF) to GPT-3, training on human rankings of model outputs so the model learned to follow the intent of an instruction rather than merely predicting likely next tokens.
In human evaluations, outputs from the 1.3B-parameter InstructGPT model were preferred over outputs from the 175B-parameter base GPT-3 — a 100×-smaller aligned model beating a much larger unaligned one on the thing users actually cared about.
Alignment, not scale, was the unlock that made an LLM usable as a product — the direct precursor to ChatGPT.
30 Nov 2022impact 96
ChatGPT — the consumer inflection
A free chat interface over an instruction-tuned GPT model put a large language model in front of hundreds of millions of people almost overnight.
OpenAI launched ChatGPT on November 30, 2022 as a free research preview built on the InstructGPT lineage, wrapping the model in a conversational interface rather than a raw API console.
UBS estimated ChatGPT reached roughly 100 million monthly active users by January 2023, about two months after launch — commonly cited as the fastest-adopted consumer software product to date.
This is the consumer inflection point — the moment LLMs stopped being a developer tool and became a mass-market product everyone had an opinion about.
24 Feb 2023impact 68
Open-weight frontier (Llama)
Meta released open-weight foundation models that any developer could download and run, breaking the assumption that frontier-quality LLMs had to stay closed.
Meta's Llama 1, released February 2023, shipped model weights from 7B to 65B parameters under a research license, showing that a smaller open model trained on more tokens could match or beat larger closed models on many benchmarks.
Llama 2, released July 2023, extended this to an open-weight commercial license, letting companies build and ship products on top of the model rather than only researchers — splitting the ecosystem into a proprietary-frontier track and an open-deployment track that persists today.
This created the open-weight track that now runs parallel to the proprietary frontier — the reason self-hosted and fine-tuned LLMs are viable at all.
14 Mar 2023impact 80
GPT-4 — frontier reasoning + multimodal
GPT-4 delivered a step-change in reasoning ability and was the first widely-used frontier model that could also take images as input.
OpenAI released GPT-4 on March 14, 2023, reporting substantially improved performance over GPT-3.5 on a wide range of professional and academic benchmarks, including simulated bar-exam scoring in the top percentile of test-takers.
GPT-4 accepted image inputs alongside text — a multimodal frontier model at consumer scale, rather than a research-only demo — setting a capability bar the rest of the field spent the following two years chasing.
GPT-4 became the reference point every competing lab benchmarked against for the next two years.
12 Sept 2024impact 86
Reasoning models shift the frontier
OpenAI's o1 made a model spend extra computation thinking through a problem before answering, turning inference-time compute into a new lever for capability.
o1, released September 2024, was trained to produce an extended chain-of-thought before its final answer, trading additional inference-time compute for improved performance on reasoning-heavy tasks such as competition math and coding.
This established test-time compute as an axis of scaling distinct from pre-training scale — a model could get measurably better at a task simply by being allowed to think longer, not by being made bigger.
This opened a second scaling axis alongside pre-training — the frontier now moves on both bigger models and longer thinking.
9 Jul 2025impact 60
Introduction of T5Gemma Models
release details: {'availability': 'pretrained and instruction-tuned', 'community_impact': 'new opportunities for research and development'}; performance comparison: {'model': 'T5Gemma 9B-9B', 'compared_to': 'Gemma 2 9B', 'DROP_score_difference': 4, 'GSM8K_score_difference': 9}
14 Aug 2025impact 60
Introduction of Gemma 3 270M Model
downloads: 200000000; model parameters: 270000000; performance comparison: exceeded larger proprietary models
12 Sept 2025impact 60
Introduction of VaultGemma Model
training method: differential privacy; model parameters: 1000000000
9 Dec 2025impact 60
Introduction of FACTS Benchmark Suite for LLM Factuality Evaluation
collaboration: Kaggle; benchmark suite: FACTS Benchmark Suite; examples available: 3513
17 Dec 2025impact 60
Introduction of Gemini 3 Flash Model
swe bench score: 78%; token processing: over 1 trillion tokens daily
20 Feb 2026impact 50
Collaboration to Enhance Local AI Support
focus: Local AI community support and enhancement; projects: ggml, llama.cpp; collaborators: Georgi Gerganov, Hugging Face
17 Mar 2026impact 60
Growth in Open Source AI Ecosystem
model count: 2000000; user growth: 13000000; dataset count: 500000
20 Mar 2026impact 60
Fine-tuning embedding models with synthetic data
commands: 6; improvements: {'Recall@10': 'over 10%', 'Recall@60': '26%'}; execution time: 2-3 hours
31 Mar 2026impact 60
ChatGPT Mainstream Adoption & Diversification
ChatGPT saw significant mainstream adoption and demographic diversification in Q1 2026, with increased usage across various age groups, genders, and emerging markets.
Users with typically feminine names now account for over half of gender-inferred ChatGPT users, indicating a shift in user demographics.
Message volume increased across all age groups, with users under 35 continuing to lead, while older age groups also showed notable growth.
The fastest adoption gains were observed in emerging markets across Latin America, Asia-Pacific, and Africa, signaling global expansion.
Workplace tasks like content creation, health-related documentation, and information retrieval were among the fastest-growing specialized uses.
The continued expansion of ChatGPT's user base and its application in diverse, specialized tasks underscores the increasing integration of LLMs into daily life and professional workflows, solidifying their mainstream status.
23 Apr 2026impact 70
Release of GPT-5.5
safeguards: advanced cybersecurity safeguards; gdpval score: 84.9; fast mode performance: 1.5x faster token generation
5 May 2026impact 65
GPT-5.5 Instant release with improved accuracy and personalization
GPT-5.5 Instant, an updated version of ChatGPT, has been released with significant improvements in accuracy, personalization, and user experience.
GPT-5.5 Instant produced 52.5% fewer hallucinated claims than GPT-5.3 Instant on high-stakes prompts covering areas like medicine, law, and finance.
It also reduced inaccurate claims by 37.3% on especially challenging conversations users had flagged for factual errors.
The update introduces memory sources for improved personalization control and better contextual understanding from past interactions.
These advancements enhance the reliability and user experience of large language models, pushing the frontier of practical AI applications.
5 May 2026impact 45
ChatGPT Ads Platform Expansion with Self-Serve Manager
ChatGPT expands its advertising platform with a new self-serve Ads Manager and cost-per-click bidding, aiming to increase advertiser participation while maintaining user privacy.
Advertisers can now create ChatGPT ads through partners or a new beta self-serve Ads Manager.
Advertisers receive aggregated performance insights that help them understand campaign impact, without access to individual conversations, ensuring user privacy.
This progression signifies a further step in the commercialization and monetization of large language models, integrating advertising as a core product feature and expanding access for businesses.
5 May 2026impact 60
Introduction of GPT-5.5 Instant
classification: High capability; safety approach: similar to previous models; preparedness categories: Cybersecurity, Biological & Chemical Preparedness
6 May 2026impact 60
vLLM V1 Migration Success
fixes: logprobs, runtime defaults, weight-update path, fp32 lm_head; final result: V1 matched V0
6 May 2026impact 40
Enhancement of Open ASR Leaderboard
visits: 710000; datasets: Appen Inc., DataoceanAI
7 May 2026impact 55
ChatGPT Advertising Pilot Expansion
ChatGPT expands its advertising pilot to several international markets, aiming to monetize the platform while maintaining user trust and privacy.
The advertising pilot is expanding to include the United Kingdom, Mexico, Brazil, Japan, and South Korea, building on initial tests in the U.S. (article_id: 5e1a5813-9395-4ae6-b70a-b755745c1102).
Early results from the pilot indicate no negative impact on consumer trust metrics, low dismissal rates of ads, and ongoing improvements in ad relevance based on user feedback (article_id: 5e1a5813-9395-4ae6-b70a-b755745c1102).
OpenAI emphasizes that ads do not influence the quality of answers provided by ChatGPT and that future expansion will be conducted responsibly as safeguards mature (article_id: 5e1a5813-9395-4ae6-b70a-b755745c1102).
This expansion signifies a critical step in the commercialization and diversification of LLM platforms, demonstrating a viable monetization strategy that balances user experience with revenue generation and broadens access to advanced AI features.
7 May 2026impact 55
Trusted Contact safety feature
ChatGPT introduces a new 'Trusted Contact' safety feature to provide support during discussions of self-harm.
Users can designate one adult as their Trusted Contact, allowing a nominated individual to be notified if the system detects discussions of self-harm.
This feature was developed with guidance from mental health professionals and organizations specializing in suicide prevention.
ChatGPT will continue to encourage users to contact crisis hotlines or emergency services when appropriate, complementing the Trusted Contact feature.
This represents a significant step in integrating proactive user safety mechanisms into large language models, addressing critical ethical and societal concerns around AI interaction with sensitive topics.
7 May 2026impact 65
GPT-5.5 and GPT-5.5-Cyber for Cybersecurity with Trusted Access
OpenAI rolls out GPT-5.5 and a specialized GPT-5.5-Cyber model, integrated with a new Trusted Access for Cyber framework, to enhance cybersecurity capabilities for defenders.
GPT-5.5 is described as OpenAI's "smartest and most intuitive model to date," indicating a significant advancement in their core LLM technology.
GPT-5.5-Cyber is specifically designed to facilitate "specialized dual-use workflows" in cybersecurity, highlighting a targeted application of advanced LLMs.
The "Trusted Access for Cyber" framework, an identity and trust-based system, ensures responsible use of these enhanced capabilities, requiring "phishing-resistant account security protections" for users.
OpenAI is partnering with security vendors to improve the overall cybersecurity ecosystem through this deployment.
This marks a significant step in deploying frontier LLMs with specific safeguards for critical applications like cybersecurity, demonstrating the evolving strategies for responsible AI integration.
19 May 2026impact 60
OlmoEarth v1.1 Launch
model: OlmoEarth v1.1; compute cost reduction: up to 3x
25 May 2026impact 65
ChatGPT integrates trusted journalism with attribution via media partnership
OpenAI has partnered with major Brazilian news organizations, Grupo Folha and Grupo UOL, to integrate high-quality journalism into ChatGPT and provide AI tools to publishers.
More than 900 million weekly active ChatGPT users globally will gain access to summaries based on reporting from Folha de S.Paulo and UOL.
This marks OpenAI’s first media partnership in Brazil, a significant market with over 50 million monthly active ChatGPT users.
The collaboration emphasizes attribution, transparency, and linking back to original sources, aiming to integrate trusted reporting into AI experiences.
Grupo Folha and Grupo UOL will also receive access to OpenAI's Codex, ChatGPT Enterprise, and API to explore AI applications in journalism and enhance operations.
This partnership establishes a model for LLM providers to engage with traditional media, addressing content sourcing, attribution, and the integration of AI tools into journalistic workflows, which is crucial for the responsible evolution of LLMs as information platforms.
9 Jun 2026impact 50
Integration of GitHub CI with Hugging Face Jobs
gpu access: True; hardware support: various configurations; ci time improvement: 30%
18 Jun 2026impact 50
Evaluation of PEFT Techniques
LoRA usage: {'checkpoints': {'total': 10000, 'percentage': 95.0, 'loRA_checkpoints': 7111}, 'model_cards': {'total': 20834, 'percentage': 98.4, 'loRA_mentions': 20509}}
26 Jun 2026impact 40
Introduction of vLLM Server on Hugging Face Jobs
summary: The article details how to run a vLLM server using Hugging Face Jobs, highlighting its flexibility and ease of use.; command example: curl https:// --8000.hf.jobs/v1/models -H "Authorization: Bearer $(hf auth token)"
6 Jul 2026impact 50
Reddit's LLMs reduce spam
spam posts caught: 25000; spam views blocked: 23000000; reduction percentage: 20
9 Jul 2026impact 60
OpenAI announces GPT 5.6 as preferred model for Microsoft 365 Copilot
model: GPT 5.6; application: Microsoft 365 Copilot; partnership: OpenAI and Microsoft
9 Jul 2026impact 60
Launch of GPT-5.6 Family of Models
models: {'name': 'Sol', 'pricing': {'input': 5, 'output': 30}, 'efficiency': '54% more token efficient for coding tasks'}, {'name': 'Terra', 'pricing': {'input': 2.5, 'output': 15}}, {'name': 'Luna', 'pricing': {'input': 1, 'output': 6}}; superiority metric: Artificial Analysis Coding Agent Index
9 Jul 2026impact 60
Introduction of GPT-5.6 for Microsoft 365 Copilot
integration: Microsoft will access OpenAI models directly through the API; enhancements: Improved document drafting and editing in Word, Enhanced data analysis efficiency in Excel, Aids in creating polished presentations in PowerPoint, Facilitates complex work in Cowork
9 Jul 2026impact 70
Launch of GPT-5.6 Family of Models
safeguards: advanced safeguards against misuse; applications: coding, knowledge work, cybersecurity; performance benchmark: {'model': 'GPT-5.6 Sol', 'score': 53.6, 'comparison': {'against': 'Claude Fable 5', 'difference': 13.1}}
9 Jul 2026impact 60
Launch of GPT-5.6 and ChatGPT Work
OpenAI has launched GPT-5.6 and ChatGPT Work, enhancing AI accessibility for users.
OpenAI launched ChatGPT Work alongside GPT-5.6.
Mac and Windows users worldwide, including free ChatGPT users, should have immediate access to ChatGPT Work and GPT-5.6 via the ChatGPT desktop app.
This launch marks a significant step in making advanced AI tools more accessible to everyday users.
Lineage
Descended from
This is one node. The map holds the whole field.
Watch LLMs — and everything it connects to — grow as real news threads onto the map every day.