The Drum · Weekly AI news, freeSubscribe →
Timeline
The rise of large language models represents one of the most significant technological shifts of the 2020s. OpenAI's ChatGPT, launched in November 2022, democratized access to advanced AI capabilities and sparked a global race among tech companies to develop and deploy their own generative AI systems. What began as a research curiosity evolved into a transformative technology that reshaped how millions of people work, learn, and create. The trajectory from GPT-2 to GPT-4 and beyond reveals both the accelerating pace of AI development and the mounting questions about safety, accuracy, and societal impact. Today, large language models power everything from customer service to scientific research, yet fundamental questions about their limitations and risks remain contested.
The Story
The story begins with foundational research into neural networks and transformer architectures, accelerating through the 2010s as computing power and data availability increased exponentially. OpenAI's release of ChatGPT in late 2022 catalyzed mainstream awareness and triggered an industry-wide sprint to develop competing systems. Google, Meta, Anthropic, and others rushed to announce their own large language models, each claiming advantages in speed, accuracy, or safety. What remains unresolved is whether current scaling approaches will continue to yield improvements, how to align these systems with human values, and what regulatory frameworks should govern their deployment. The narrative arc suggests we are still in the early chapters of this transformation, with the most consequential decisions about AI governance and safety still ahead.
2017
Transformer Architecture Published
Researchers at Google introduce the transformer architecture in the paper 'Attention Is All You Need.' This foundational innovation enables more efficient processing of sequential data and becomes the basis for all modern large language models.
2018-06
OpenAI Releases GPT-1
OpenAI publishes the first Generative Pre-trained Transformer model, demonstrating that unsupervised learning on large text corpora could produce useful language capabilities. The model contains 117 million parameters.
2019-02
GPT-2 Released with Safety Concerns
OpenAI releases GPT-2, a 1.5 billion parameter model that shows significant improvements in text generation quality. The company initially withholds the full model due to concerns about misuse, sparking debate about responsible AI disclosure.
2020-06
GPT-3 Demonstrates Scaling Laws
OpenAI releases GPT-3 with 175 billion parameters, showing that larger models exhibit emergent capabilities and can perform tasks with minimal examples. The model's versatility generates significant media attention and venture capital interest in AI.
2020-09
Google Releases BERT and T5
Google publishes multiple large language models including BERT and T5, establishing itself as a major player in the LLM space. These models achieve state-of-the-art results on various benchmarks.
2021-01
OpenAI Launches GPT-3 API
OpenAI makes GPT-3 available through a commercial API, allowing developers and businesses to build applications on top of the model. This move begins the transition from research artifact to commercial product.
2021-06
Meta Releases OPT-175B
Meta releases Open Pretrained Transformer, a 175 billion parameter model comparable to GPT-3. The company emphasizes open access and reproducibility in AI research.
2021-12
DeepMind Releases Gato
DeepMind introduces Gato, a multi-modal model that can perform tasks across text, images, and control domains. The work suggests that scaling and multi-task training could lead to more general AI systems.
2022-01
OpenAI Releases Whisper
OpenAI releases Whisper, a speech recognition model trained on 680,000 hours of multilingual audio data. The model demonstrates strong performance across languages and accents.
2022-03
Google Releases PaLM
Google announces Pathways Language Model, a 540 billion parameter model showing improvements in reasoning and few-shot learning. The company positions PaLM as evidence of its continued leadership in AI.
2022-04
Anthropic Founded
Former OpenAI researchers, including Dario and Daniela Amodei, found Anthropic to focus on AI safety and alignment. The company raises $124 million in Series A funding to develop safer large language models.
2022-09
Meta Releases LLaMA
Meta releases LLaMA, a family of open-source language models ranging from 7 billion to 65 billion parameters. The models achieve competitive performance with closed-source alternatives, democratizing access to capable LLMs.
2022-11-30
ChatGPT Launches to Public
OpenAI releases ChatGPT, a fine-tuned version of GPT-3.5 optimized for conversational interaction. The product reaches 1 million users in five days and 100 million users within two months, becoming the fastest-growing application in history.
2022-12
Google Announces Bard
Google announces Bard, its conversational AI assistant powered by LaMDA. The announcement comes in response to ChatGPT's rapid adoption and signals Google's commitment to competing in the generative AI space.
2023-01
ChatGPT Reaches 100 Million Users
ChatGPT becomes the fastest-growing consumer application ever, reaching 100 million monthly active users. The milestone demonstrates mainstream adoption of large language models.
2023-02
Microsoft Invests $10 Billion in OpenAI
Microsoft announces a $10 billion investment in OpenAI as part of a multi-year partnership. The company integrates ChatGPT capabilities into Bing and Office products, escalating competition with Google.
2023-03-14
OpenAI Releases GPT-4
OpenAI releases GPT-4, a multimodal model that accepts both text and image inputs. The model shows significant improvements in reasoning, accuracy, and safety compared to GPT-3.5, though the exact parameter count remains undisclosed.
2023-03-31
Google Releases Bard to Public
Google opens Bard access to the public in the United States and United Kingdom. The move represents Google's formal entry into the conversational AI market following ChatGPT's success.
2023-04
Anthropic Releases Claude
Anthropic releases Claude, a large language model trained using constitutional AI methods to improve safety and alignment. The model emphasizes harmlessness and honesty in its design.
2023-05
Open Letter Calls for AI Pause
Prominent AI researchers and technology leaders, including Elon Musk and Steve Wozniak, sign an open letter calling for a six-month pause on training large language models. The letter raises concerns about existential risks from advanced AI systems.
2023-05-30
OpenAI Releases GPT-4 Turbo
OpenAI announces GPT-4 Turbo with an extended context window of 128,000 tokens, improved instruction following, and lower pricing. The update demonstrates rapid iteration on the GPT-4 architecture.
2023-06
EU Proposes AI Act Regulations
The European Union advances the AI Act, proposing comprehensive regulations for high-risk AI systems including large language models. The legislation aims to establish safety standards and transparency requirements.
2023-07
Meta Releases Llama 2
Meta releases Llama 2, an improved open-source language model available for commercial use. The model achieves performance comparable to closed-source alternatives while maintaining open access.
2023-09
Anthropic Raises $5 Billion
Anthropic announces a $5 billion funding commitment from Google, establishing the company as a major competitor in the LLM space. The investment reflects growing competition for AI talent and capabilities.
2023-10
OpenAI Releases GPT-4 Vision
OpenAI enables image input capabilities for GPT-4, allowing users to submit images for analysis and discussion. The multimodal capability expands the model's utility across new use cases.
2023-11-06
OpenAI Releases GPTs and Custom Models
OpenAI announces GPTs, allowing users to create customized versions of ChatGPT without coding. The platform enables developers to build specialized applications on top of OpenAI's models.
2023-11-13
Sam Altman Removed as OpenAI CEO
OpenAI's board removes Sam Altman as CEO, citing concerns about his communication with the board. The move triggers internal turmoil and employee resignations, with Altman later reinstated after negotiations.
2023-12
OpenAI Releases GPT-4 Turbo with Vision
OpenAI releases an updated version of GPT-4 Turbo that includes vision capabilities and improved performance. The model supports a 128,000 token context window and knowledge cutoff in April 2024.
2024-01
Mistral AI Releases Mixtral
Mistral AI releases Mixtral, an open-source mixture-of-experts model demonstrating competitive performance with larger closed-source models. The release accelerates the open-source LLM movement.
2024-02
OpenAI Releases Sora
OpenAI announces Sora, a text-to-video generation model that creates minute-long videos from text descriptions. The capability extends generative AI beyond text and images into video synthesis.
2024-03
Google Releases Gemini Ultra
Google releases Gemini Ultra, its most capable multimodal model to date. The model demonstrates improvements in reasoning, coding, and creative tasks across text, image, and video modalities.
2024-05
OpenAI Announces GPT-4o
OpenAI releases GPT-4o, a multimodal model optimized for speed and efficiency. The model processes text, images, and audio, with native support for real-time voice conversations.
2024-06
Anthropic Releases Claude 3.5 Sonnet
Anthropic releases Claude 3.5 Sonnet, demonstrating improvements in reasoning and coding capabilities. The model achieves competitive performance with GPT-4 while maintaining focus on safety and alignment.
2024-07
UK Establishes AI Regulation Framework
The United Kingdom publishes its AI regulation framework, taking a lighter-touch approach compared to the EU. The framework emphasizes principles-based regulation and sector-specific oversight.
2024-08
Meta Releases Llama 3.1
Meta releases Llama 3.1, an open-source model family with improved performance and a 128,000 token context window. The release reinforces Meta's commitment to open-source AI development.
2024-09
OpenAI Announces o1 Reasoning Model
OpenAI introduces o1, a model trained to spend more time reasoning before responding. The approach shows improvements on complex reasoning tasks and represents a shift toward process-based training methods.
