4 novembre 2023
Grok est une IA calquée sur le Guide de l’auto-stoppeur galactique, donc destinée à répondre à presque n’importe quoi et, beaucoup plus difficile, même à suggérer les questions à poser !
Grok est conçu pour répondre aux questions avec un peu d’esprit et a un côté rebelle, alors s’il vous plaît, ne l’utilisez pas si vous détestez l’humour !
Un avantage unique et fondamental de Grok est qu’il dispose d’une connaissance en temps réel du monde via la plate-forme X. Il répondra également à des questions piquantes qui sont rejetées par la plupart des autres systèmes d’IA.
Grok est encore un produit bêta très précoce – le mieux que nous puissions faire avec 2 mois d’entraînement – alors attendez-vous à ce qu’il s’améliore rapidement chaque semaine qui passe avec votre aide.
Merci
à l’équipe xAI
Pourquoi nous construisons Grok
Chez xAI, nous voulons créer des outils d’IA qui aident l’humanité dans sa quête de compréhension et de connaissance.
En créant et en améliorant Grok, nous visons à :
- Recueillez des commentaires et assurez-vous que nous construisons des outils d’IA qui profitent au maximum à l’ensemble de l’humanité. Nous pensons qu’il est important de concevoir des outils d’IA utiles aux personnes de tous horizons et de toutes opinions politiques. Nous voulons également donner à nos utilisateurs les moyens d’utiliser nos outils d’IA, sous réserve de la loi. Notre objectif avec Grok est d’explorer et de démontrer cette approche en public.
- Renforcer la recherche et l’innovation : nous voulons que Grok serve d’assistant de recherche puissant pour tout le monde, en l’aidant à accéder rapidement aux informations pertinentes, à traiter les données et à trouver de nouvelles idées.
Notre objectif ultime est que nos outils d’IA aident à la compréhension.
Le voyage vers Grok-1
Le moteur qui propulse Grok est le Grok-1, notre LLM de pointe, que nous avons développé au cours des quatre derniers mois. Grok-1 a connu de nombreuses itérations au cours de cette période.
Après l’annonce de xAI, nous avons entraîné un prototype de LLM (Grok-0) avec 33 milliards de paramètres. Ce modèle précoce se rapproche des capacités de LLaMA 2 (70B) sur des benchmarks LM standard, mais n’utilise que la moitié de ses ressources de formation. Au cours des deux derniers mois, nous avons apporté des améliorations significatives dans les capacités de raisonnement et de codage menant à Grok-1, un modèle de langage de pointe qui est nettement plus puissant, atteignant 63,2 % sur la tâche de codage HumanEval et 73 % sur MMLU.
Pour comprendre les améliorations de capacités que nous avons apportées avec Grok-1, nous avons mené une série d’évaluations à l’aide de quelques benchmarks d’apprentissage automatique standard conçus pour mesurer les capacités de mathématiques et de raisonnement.
GSM8k : Problèmes de mots mathématiques au collège (Cobbe et al. 2021), à l’aide de l’invite de chaîne de pensée.
MMLU : Les questions multidisciplinaires à choix multiples (Hendrycks et al. 2021) ont fourni des exemples contextuels à 5 coups.
HumanEval: Python code completion task, (Chen et al. 2021), zero-shot evaluated for pass@1.
MATH: Middle school and high school mathematics problems written in LaTeX, (Hendrycks et al. 2021), prompted with a fixed 4-shot prompt.
Benchmark | Grok-0 (33B) | LLaMa 2 70B | Inflection-1 | GPT-3.5 | Grok-1 | Palm 2 | Claude 2 | GPT-4 |
---|---|---|---|---|---|---|---|---|
GSM8k | 56.8% 8-shot |
56.8% 8-shot |
62.9% 8-shot |
57.1% 8-shot |
62.9% 8-shot |
80.7% 8-shot |
88.0% 8-shot |
92.0% 8-shot |
MMLU | 65.7% 5-shot |
68.9% 5-shot |
72.7% 5-shot |
70.0% 5-shot |
73.0% 5-shot |
78.0% 5-shot |
75.0% 5-shot + CoT |
86.4% 5-shot |
HumanEval | 39.7% 0-shot |
29.9% 0-shot |
35.4% 0-shot |
48.1% 0-shot |
63.2% 0-shot |
- | 70% 0-shot |
67% 0-shot |
MATH | 15.7% 4-shot |
13.5% 4-shot |
16.0% 4-shot |
23.5% 4-shot |
23.9% 4-shot |
34.6% 4-shot |
- | 42.5% 4-shot |
On these benchmarks, Grok-1 displayed strong results, surpassing all other models in its compute class, including ChatGPT-3.5 and Inflection-1. It is only surpassed by models that were trained with a significantly larger amount of training data and compute resources like GPT-4. This showcases the rapid progress we are making at xAI in training LLMs with exceptional efficiency.
Since these benchmarks can be found on the web and we can’t rule out that our models were inadvertently trained on them, we hand-graded our model (and also Claude-2 and GPT-4) on the 2023 Hungarian national high school finals in mathematics, which was published at the end of May, after we collected our dataset. Grok passed the exam with a C (59%), while Claude-2 achieved the same grade (55%), and GPT-4 got a B with 68%. All models were evaluated at temperature 0.1 and the same prompt. It must be noted that we made no effort to tune for this evaluation. This experiment served as a “real-life” test on a dataset our model was never explicitly tuned for.
Human-graded evaluation | Grok-0 | GPT-3.5 | Claude 2 | Grok-1 | GPT-4 |
---|---|---|---|---|---|
Hungarian National High School Math Exam (May 2023) | 37% 1-shot |
41% 1-shot |
55% 1-shot |
59% 1-shot |
68% 1-shot |
We provide a summary of the important technical details of Grok-1 in the model card.
Engineering at xAI
At the frontier of deep learning research, reliable infrastructure must be built with the same care as datasets and learning algorithms. To create Grok, we built a custom training and inference stack based on Kubernetes, Rust, and JAX.
LLM training runs like a freight train thundering ahead; if one car derails, the entire train is dragged off the tracks, making it difficult to set upright again. There are a myriad of ways GPUs fail: manufacturing defects, loose connections, incorrect configuration, degraded memory chips, the occasional random bit flip, and more. When training, we synchronize computations across tens of thousands of GPUs for months on end, and all these failure modes become frequent due to scale. To overcome these challenges, we employ a set of custom distributed systems that ensure that every type of failure is immediately identified and automatically handled. At xAI, we have made maximizing useful compute per watt the key focus of our efforts. Over the past few months, our infrastructure has enabled us to minimize downtime and maintain a high Model Flop Utilization (MFU) even in the presence of unreliable hardware.
Rust has proven to be an ideal choice for building scalable, reliable, and maintainable infrastructure. It offers high performance, a rich ecosystem, and prevents the majority of bugs one would typically find in a distributed system. Given our small team size, infrastructure reliability is crucial, otherwise, maintenance starves innovation. Rust provides us with confidence that any code modification or refactor is likely to produce working programs that will run for months with minimal supervision.
We are now preparing for our next jump in model capabilities, which will require reliably coordinating training runs on tens of thousands of accelerators, running internet-scale data pipelines, and building new kinds of capabilities and tools into Grok. If that sounds exciting to you, apply to join the team here.
Research at xAI
We give Grok access to search tools and real-time information, but as with all the LLMs trained on next-token prediction, our model can still generate false or contradictory information. We believe that achieving reliable reasoning is the most important research direction to address the limitations of current systems. Here, we would like to highlight a few promising research directions we are most excited about at xAI:
- Scalable oversight with tool assistance. Human feedback is essential. However, providing consistent and accurate feedback can be challenging, especially when dealing with lengthy code or complex reasoning steps. AI can assist with scalable oversight by looking up references from different sources, verifying intermediate steps with external tools, and seeking human feedback when necessary. We aim to make the most effective use of our AI tutors' time with the help of our models.
- Integrating with formal verification for safety, reliability, and grounding. To create AI systems that can reason deeply about the real world, we plan to develop reasoning skills in less ambiguous and more verifiable situations. This allows us to evaluate our systems without human feedback or interaction with the real world. One major immediate goal of this approach is to give formal guarantees for code correctness, especially regarding formally verifiable aspects of AI safety.
- Long-context understanding and retrieval. Training models for efficiently discovering useful knowledge in a particular context are at the heart of producing truly intelligent systems. We are working on methods that can discover and retrieve information whenever it is needed.
- Adversarial robustness. Adversarial examples demonstrate that optimizers can easily exploit vulnerabilities in AI systems, both during training and serving time, causing them to make egregious mistakes. These vulnerabilities are long-standing weaknesses of deep learning models. We are particularly interested in improving the robustness of LLMs, reward models, and monitoring systems.
- Multimodal capabilities. Currently, Grok doesn’t have other senses, such as vision and audio. To better assist users, we will equip Grok with these different senses that can enable broader applications, including real-time interactions and assistance.
We believe that AI holds immense potential for contributing significant scientific and economic value to society, so we will work towards developing reliable safeguards against catastrophic forms of malicious use. We believe in doing our utmost to ensure that AI remains a force for good.
If you share our optimism and want to contribute to our mission, apply to join the team here.
Early Access to Grok
We are offering a limited number of users in the United States to try out our Grok prototype and provide valuable feedback that will help us improve its capabilities before a wider release. You can join the Grok waitlist here. This release just represents the first step for xAI. Looking ahead, we have an exciting roadmap and will be rolling out new capabilities and features in the coming months.