Long-text Prompting??, Real-time Image Gen???, GPT-4 Rival+Fake News ??
Our intuition about the future is linear. But the reality of information technology is exponential, and that makes a profound difference. If I take 30 steps linearly, I get to 30. If I take 30 steps exponentially, I get to a billion.
— Ray Kurzweiler, renowned futurist and inventor. His book “The singularity is near“ left an impression on me.
?? Top AI News
Claude (an LLM) by Anthropic has been extensively utilized by millions for various tasks—translating scholarly works, crafting business strategies, and scrutinizing legal agreements. In a transformative upgrade, its context capacity has doubled to 200,000 tokens, enabling the processing of voluminous documents, approximately 150,000 words or over 500 pages in length. This expansion empowers users to feed Claude extensive texts—be it technical, financial, or literary—facilitating functions like summarization, question-answering, trend forecasting, document comparison, and more.
The latest iteration, Claude 2.1, marks a significant leap in reliability, slashing inaccuracies by half compared to its precursor, Claude 2.0. This advancement fosters trust, paving the way for businesses to integrate AI solutions more confidently, addressing real-world challenges and embedding AI across their operations.
A finding reveals that incorporating a prompt tweak, "Here is the most relevant sentence in the context: […]", enhances Claude's recall by 70%. This simple yet effective modification boosted Claude 2.1's performance from 27% to 98% in handling a 200K token context.
Continual enhancements are on the horizon for Claude, with a keen focus on integrating with sector-specific applications and refining user interfaces, pushing the boundaries of what AI can achieve in our ever-evolving digital landscape.
Stability AI has introduced SDXL Turbo, a new image generation model featuring their novel Adversarial Diffusion Distillation (ADD) technology. This model sets a new standard in real-time image generation by reducing the step count from 50 to a single step, while maintaining high image quality.
SDXL Turbo's highlights include:
The model's code, research paper, and weights for non-commercial use are accessible on the Stability AI website. Additionally, you can experience SDXL Turbo's capabilities through a beta demonstration on Clipdropapp, Stability AI's image editing platform.
For a closer look, see the demonstration video here: SDXL Turbo Demonstration.
In a recent public unveiling, Google introduced Gemini (as announced in this newsletter), its latest language model, positioning it as a formidable rival to OpenAI's GPT-4. According to their paper, Gemini has shown benchmark results that are not only impressive but also slightly outperform GPT-4 in various aspects. See benchmark scores here:
However, confidence in Google's technology and integrity took a hit when it was revealed that Gemini's most impressive demonstration was largely fabricated. During the demo, Gemini appeared to have a remarkable ability to comprehend spoken language and interpret dynamic images, creating an illusion of real-time interaction. This human-like responsiveness generated significant excitement in the tech community.
Later scrutiny revealed that part of the video did not accurately represent Gemini's actual capabilities. Instead, the demonstration used still images and text prompts to simulate performance.
In response, Google acknowledged the creative editing of the demo, admitting it was not a real-time performance. Critics have pointed out that the extent of these modifications was not clearly communicated, despite a disclaimer about potential delays. Oriol Vinyals, Google's VP of Research and co-lead of the Gemini project, released a follow-up video demonstrating the AI model's true functionality. This revealed that Gemini requires about four to five seconds to generate text from still images and textual input.
Despite this setback, I believe that within the next year, we will achieve the capabilities shown in the original video. See, for instance, the advances that have happened in image generation, similar to those made by Stability AI's SDXL Turbo.
IBM and Meta have recently launched the AI Alliance, an international coalition of over 50 organizations, including AMD, Intel, NASA, CERN, and various universities like the TU Munich, with the aim of advancing open innovation and open science in AI. The AI Alliance's mission is to create an open community that accelerates responsible AI innovation while ensuring scientific rigor and trust. The coalition supports open source AI development, which contrasts with the closed AI systems used by market leaders like OpenAI and Google. The AI Alliance intends to promote safety, diversity, economic opportunity, and benefits for all by fostering open and transparent innovation.
?? Upcoming Workshop
Let’s Meet Next February!
I am thrilled to announce our upcoming workshop, which will take place at the Generative AI for Marketing Summit.
领英推荐
?? 26th of February 2024, ??Chelsea Harbour Hotel, London, UK.
Discover and Unlock:
Transition from theory to actionable insights amidst a like-minded tribe. Reserve your seat now with code: SPEAKER10.
P.S: I can’t wait.
?? From Software to Hardware ..
Discussions about advancements in language models and AI often focus on the software layer. However, it's important to also consider the significant convergence with the underlying hardware.
TLDR; computing units (GPUs and similar) performance becomes improves faster than Moore’s Law, roughly doubling, as can be observed in the market.
AWS's introduction of two new chip families, Graviton4 and Trainium2, is particularly noteworthy. These chips are designed to enhance price performance and energy efficiency for a variety of customer workloads, including machine learning training and generative AI applications.
Graviton4 is touted as AWS's most powerful and energy-efficient processor to date. It offers up to 30% better compute performance, 50% more cores, and 75% more memory bandwidth compared to its predecessor. A key feature of Graviton4 is its emphasis on security, achieved by encrypting high-speed hardware interfaces. Available in Amazon EC2 R8g instances, which are memory-optimized, Graviton4 is particularly beneficial for high-performance databases, in-memory caches, and big data analytics workloads. These instances can handle larger sizes, processing more data efficiently. The preview for R8g instances is currently open, with general availability expected in the near future.
Trainium2, on the other hand, is specifically designed for high-performance AI model training. It promises up to four times faster training performance, three times more memory capacity, and double the energy efficiency compared to its predecessor. Housed in Amazon EC2 Trn2 instances, each with 16 Trainium2 chips, it is capable of scaling up to 100,000 chips in next-generation EC2 UltraClusters. When combined with petabit-scale networking through AWS Elastic Fabric Adapter (EFA), this setup provides an impressive compute power of up to 65 exaflops. Such scalability is crucial, as it allows customers to train a 300-billion parameter large language model in weeks instead of months, effectively offering supercomputer-level performance.
Google has recently unveiled its Cloud TPU v5p and AI Hypercomputer, engineered to facilitate cutting-edge AI workloads. The Cloud TPU v5p represents Google's most advanced and scalable TPU accelerator yet, offering double the floating-point operations per second (FLOPS) of its predecessor, the TPU v4, and boasting four times greater scalability.
Google asserts that the TPU v5p can train large language models, such as GPT3-175B, 2.8 times faster than the TPU v4, while also being more cost-efficient.
The AI Hypercomputer integrates systems-level codesign to enhance efficiency and productivity in AI tuning and serving. It incorporates liquid cooling and employs Google's Jupiter data center network technology. This synergy of hardware and software culminates in a cohesive, user-friendly, secure, and versatile cloud-based supercomputer architecture. This architecture is adept at supporting the most demanding AI workloads, showcasing Google's commitment to advancing AI technology.
Subscribing to, giving feedback about, and sharing the newsletter as well as our renowned online course will be highly appreciated and helps a lot. ??
Our upcoming book Generative AI: Navigating the Course to the Artificial General Intelligence Future.
If you would like to sponsor an ad to this 30k+ newsletter, please, respond to this email.
Thank you so much for reading,
Martin
IT Certification at TIBCO
11 个月Ready for a career transformation? www.certfun.com/exin is your catalyst for change. Embrace the journey! ???? #CertFun #CareerTransformation
Senior Data Scientist @IBM | GenAI | AI Governance
11 个月Enjoyed reading this episode a lot! Thanks for helping us keeping up with all that is happening Martin! ????