| | | | Your daily update on artificial intelligence frameworks and leadership strategies, enriched with knowledge from India's forefront AI innovators and leaders. |
|
|
|
|
|
|
| Hello there! Don't be surprised; we just gave AIM Daily XO an exciting new makeover. Everything's different—the look, the feel, the entire experience! But it's still your friendly AI Human, Amit Raja Naik, bringing it to you and sharing amazing AI updates with a twist. I hope you enjoy this new flavour! Well, Databricks just hit it out of the park with DBRX. Let’s dive in to see what’s going on there. Happy reading. 🤗 |
| | | Design: Amit Raja Naik Databricks released the world's most powerful open-source model The aftermath was quite fatal. It outperformed every known open-source LLM on the planet across reasoning, coding and maths. This includes Meta’s LLaMA, Mixtral, Google’s Gemini Pro, and even xAI’s Grok (released barely a few weeks ago). Forget poor open-source models; it even surpassed OpenAI’s closed-source models like GPT-3.5 and GPT-4. In other words, enterprises now have no choice but to switch to an open-source model that is as powerful as closed-source models and gives them the cost advantage of open-source. Comparing its price with closed-source models like GPT-4, the company’s VP of generative AI, Naveen Rao, said it is 1/10th of the dollars per token. |
| Citing GPT-4, he said it is $120 per 1 million output tokens, whereas DBRX with a 32K context window is only $6.2 per 1 million token output—i.e., 20 times lower than GPT-4 for 1 million token outputs. |
|
|
|
|
| Many have started embracing the model. It is also trending #1 on Hugging Face. Some of the early adopters are Perplexity AI, You.com, Accenture and NVIDIA. A bunch of avid developers are already experimenting with DBRX on M2 Ultra. |
| | | Design: Nikhil Kumar But, how was Databricks able to achieve this? The answer to this led us to its MoE architecture. “The economics are so much better for serving. They’re more than 2X better in terms of flops and floating point operations required to do the serving,” said Rao. He added that they’re the only ones who made it work at scale. “We’re not aware of anyone else who’s done it.” The MoE architecture directly addresses key enterprise barriers like cost, privacy/control, and complexity that have hindered AI adoption, said Rao in an exclusive interview with AIM before the release. Databricks is NOT alone Just a few weeks ago, Elon Musk’s xAI open-sourced Grok-1. Rao wasn’t impressed. “I don’t think Grok-1 is that great despite its large 314-billion parameter size,” he added, saying that it gets really difficult to evaluate as it takes a lot of compute. |
| He also pointed out that Grok’s capabilities are not commensurate with its massive scale and said: “For a model that size, I would expect a higher level of capability.” |
|
|
| Sadly, Rao sounded disappointed with xAI's Grok open-source model – its quality and capabilities relative to its massive scale – suggesting DBRX is a superior open-source alternative that outperforms Grok on most benchmarks. |
|
|
|
|
| xAI releases Grok 1.5 |
| | | Design: Diksha Mishra Recently, xAI releasedGrok 1.5. The model comes with improved reasoning capabilities and a 128K context window. It will soon be available on X. “Grok 2 should exceed current AI on all metrics. In training now,” said Elon Musk. Meanwhile, Meta is currently leading the open-source LLM race with Meta’s Llama 2, and the release of Llama 3 is also on the horizon. Rao is unperturbed. He said that most of them would not be able to achieve the same economics and quality. “I won’t be surprised if their model has the same economics as ours, but it is going to be of worse quality.” He said that the key is to look at not just performance quality but also cost and time – like how fast it is, what it costs, and the quality – all put together. “When you look at these together, I don’t think anyone’s going to beat us in a long time,” said Rao, confidently. Is DBRX really better than others? “We have to see to believe it,” said Manoj Shinde, the founder of and creator of BeautyGPT. He noted that some models performed better than others depending on the use cases; most benchmarks have been largely insignificant. |
| “If one were to choose the cheapest model, then Mistral AI (via Anyscale) with an average cost of $0.031 per 1 million tokens, (compared to $6.2 per million token) [would be the best bet],” shared Shinde. |
|
|
|
|
| Databricks’ latest DBRX 132B model uses a 36-layer MoE architecture with 16 experts, including 4 active ones, which is very unique. However, it still lags behind proprietary models such as Mistral and Anthropic. ARKInvest’s Brett Winton said that the resultant performance is SOTA for open source, particularly on coding, where it blows the competition out of the water. “Against commercial models, DBRX is a better coder than GPT-4 (at its initial release) and Gemini Pro (😬) (via the humanEval benchmark),” he added. The performance gain of DBRX definitely comes from better data and tokeniser from GPT-4, and obviously, MoE architecture, particularly when compared with MPT approaches. |
| | | Can we make India the world’s first AI-driven economy? |
| | | Anil Bhasin, Databricks India’s VP and country manager, is not only optimistic but firmly believes in this visionary goal. |
| | | Language is the OS of the Future |
| | | | | | | INDIA - LatentView Analytics acquires Decision Point for $39.1 million to enhance its data analytics capabilities, expand its presence in North America and Europe, and strengthen its focus on AI-led business solutions, particularly in the consumer packaged goods sector.
- Minfy Technologies collaborates with AWS India to enhance cloud services and AI utilisation, aiming for international expansion and business growth.
- Gyan AI Research introduces PARAMANU-AYN, a specialised NLP model for the Indian legal domain, showing promising results but with acknowledged limitations.
- Krutrim partners with Databricks to particularly for Indian languages, aiming to enhance AI solutions in India.
|
| | | Thank you. But before you leave. India’s biggest summit on diversity and inclusion in tech and AI, Rising 2024, is barely a week away. Catch us there on April 4&5 at the Hilton Convention Center, Manyata Tech Park, Bengaluru. Grab your passes today >>> Enjoying AIM Daily XO? Share it with colleagues or friends – they can sign up here. We love hearing from our readers! Have thoughts on our new format? Questions, comments, or ideas are always welcome. If there’s a specific topic in AI or analytics that you're curious about, tell us! Reach out to us at info@analyticsindiamag.com. Stay tuned for more insights in our next edition! Cheers! Amit Raja Naik |
|
|
|
|
|
|
| | | AIM Daily XO is published by AIM Media House, a global Media and Analyst Firm dedicated to Artificial Intelligence |
|
| | |
|
| |
|
|
|
|
|
|
Комментариев нет:
Отправить комментарий
Примечание. Отправлять комментарии могут только участники этого блога.