# Stability AI Launches StableLM, an Open\-Source ChatGPT Rival – A Potential Game Changer for AI Accessibility - Canonical: https://33rdsquare.com/stability-ai-launches-stablelm-an-open-source-chatgpt-rival-a-potential-game-changer-for-ai-accessibility/ - Published: 2024-09-03 - Author: Jordan Brown - Categories: [Artificial Intelligence & Machine Learning & ChatGPT](https://33rdsquare.com/category/tech/ai/) --- Earlier this week, artificial intelligence research company Stability AI unveiled its latest creation to the machine learning community – StableLM, an open source conversational AI system intended as an alternative to ChatGPT. By releasing StableLM as an open source platform, Stability AI hopes to catalyze innovation in natural language AI – bringing us one step closer to democratized access to foundational AI technology. ## How Does StableLM Work? Like ChatGPT and other natural language AI systems, StableLM is an auto-regressive language model. This means it predicts the next word in a sequence given all the previous words as context. More specifically, StableLM is a transformer-based model – it uses an attention mechanism to learn contextual relationships between words. The model ingests massive textual datasets, then is trained to predict the next word after a given sequence of words. To generate text, users provide an initial prompt and StableLM will continue predicting subsequent words that are likely to follow that prompt. This enables it to produce remarkably human-like text, hold conversations, summarize passages, generate code, and more in response to user input. StableLM was trained on The Pile, an open source dataset curated by Stability AI. Comprising over 1.5 trillion tokens scraped from the internet, The Pile is approximately 3x larger than the Common Crawl dataset used for GPT-3. This extensive training enables StableLM to tackle a variety of natural language tasks. | Model | Parameters | Training Dataset | Availability | | --- | --- | --- | --- | | StableLM | 3B – 7B (up to 65B planned) | The Pile (1.5 trillion tokens) | Open source | | ChatGPT | 175B | Internal OpenAI dataset | Proprietary | | GPT-3 | 175B | Common Crawl (570GB) | Restricted access | In its initial iterations, StableLM ranges from 3 billion to 7 billion parameters – significantly smaller than GPT-3 and ChatGPT. However, Stability AI asserts that with proper training, even more compact models can reach impressive performance. And in the open source spirit, they plan to keep rapidly iterating to produce larger versions up to 65 billion parameters. ## Testing StableLM‘s Capabilities StableLM is currently available in an early alpha version on GitHub and Hugging Face for anyone to experiment with. I chatted with the 7 billion parameter "StableLM-Tuned-Alpha-7b" model to get a feel for its conversational abilities. Though slower than ChatGPT, StableLM managed intelligent, eloquent responses to my open-ended prompts. When asked to write a poem, it replied: _Let me weave a verse for you, my friend / Of hopes and dreams that have no end / Where truth and light together blend / And joy reigns free o‘er hill and glen / Where every soul finds room to mend / In a world where all are kin again_ Impressive rhyme and thematic coherence for 7 billion parameters! With more training data, I could see the potential for StableLM generating even more sophisticated text. According to Gary Marcus, cognitive scientist and CEO of AI startup Robust.AI, "StableLM seems like a promising start, but people should temper their expectations about what it can currently do.” He predicts capabilities will improve rapidly as Stability AI releases larger models. ## Open Source vs. Proprietary AI Unlike ChatGPT which is proprietary, StableLM takes an open source approach – the model architecture, weights, training code, and datasets have all been published. This transparency brings many advantages. Researchers can closely analyze models for biases or limitations. Developers can build custom solutions on top of the core model. And the community can collaborate to continuously enhance open source AI. However, some experts argue closed development allows companies like OpenAI to pour more resources into optimizing proprietary models. Anthropic, an AI safety startup, wrote that "open-source models often represent understandably suboptimal implementations”. So while the open source approach promotes accessibility, Stability AI has their work cut out replicating the performance of industry giants. But by tapping into community resources, their rapid iterations on StableLM aim to compete while remaining open. ## Democratizing Access to Beneficial AI The release of StableLM mirrors Stability AI‘s previous open-sourcing of text-to-image generator Stable Diffusion. This demonstrated the immense value unleashed when foundations of AI become freely accessible. With StableLM, conversational AI could see the same democratization. As Gary Marcus put it, "Making AI accessible to all will likely require openness, transparency and collaboration.” If capabilities improve as intended, StableLM could expand access to AI that solves real-world problems, boosts business productivity, and enhances knowledge – not just brute compute resources. Democratization also reduces potential downsides of AI concentration in large corporations. Open systems allow broader input into developing AI safely and ethically. And the more who can build upon shared models, the greater diversity of beneficial applications we’ll see. No model is perfect, and StableLM has much maturing ahead. But its launch represents a stride towards openness in a technology with immense societal influence. As an AI enthusiast, I’m excited to see where the journey takes us! --- Source: [Stability AI Launches StableLM, an Open\-Source ChatGPT Rival – A Potential Game Changer for AI Accessibility](https://33rdsquare.com/stability-ai-launches-stablelm-an-open-source-chatgpt-rival-a-potential-game-changer-for-ai-accessibility/)