GenAI Weekly — Edition 27

GenAI Weekly — Edition 27

Your Weekly Dose of Gen AI: News, Trends, and Breakthroughs

Stay at the forefront of the Gen AI revolution with Gen AI Weekly! Each week, we curate the most noteworthy news, insights, and breakthroughs in the field, equipping you with the knowledge you need to stay ahead of the curve.

? Click subscribe to be notified of future editions



A Primer on AI Invoice Processing and Extraction

From the Unstract blog:

Tarun Singh

Invoice management, especially at scale, can be complex and time-consuming due to the variety of formats—like paper, email, PDFs, or EDI—each containing a mix of structured and unstructured data. Accurately extracting this information is crucial for smooth financial operations.

AI and Large Language Models (LLMs) have revolutionized invoice processing, enhancing efficiency and precision beyond traditional OCR technology. This article explores the challenges of invoice processing and how integrating LLMs with Unstract can transform the way businesses handle invoices.


Microsoft releases powerful new Phi-3.5 models, beating Google, OpenAI and more

Carl Franzen writing for Venture Beat :

VentureBeat Carl Franzen

The three new Phi 3.5 models include the 3.82 billion parameter Phi-3.5-mini-instruct , the 41.9 billion parameter Phi-3.5-MoE-instruct , and the 4.15 billion parameter Phi-3.5-vision-instruct , each designed for basic/fast reasoning, more powerful reasoning, and vision (image and video analysis) tasks, respectively.
All three models are available for developers to download, use, and fine-tune customize on Hugging Face under a Microsoft-branded MIT License that allows for commercial usage and modification without restrictions.
Amazingly, all three models also boast near state-of-the-art performance across a number of third-party benchmark tests, even beating other AI providers including Google’s Gemini 1.5 Flash, Meta’s Llama 3.1, and even OpenAI’s GPT-4o in some cases.

My take on this: Very, very impressive. Challenges the notion that just increasing the number of parameters in the model won’t amount to better capabilities.


Artificial intelligence is losing hype

From The Economist :

Silicon Valley’s tech bros are having a difficult few weeks. A growing number of investors worry that artificial intelligence (AI) will not deliver the vast profits they seek. Since peaking last month the share prices of Western firms driving the ai revolution have dropped by 15%. A growing number of observers now question the limitations of large language models, which power services such as ChatGPT. Big tech firms have spent tens of billions of dollars on ai models, with even more extravagant promises of future outlays. Yet according to the latest data from the Census Bureau, only 4.8% of American companies use ai to produce goods and services, down from a high of 5.4% early this year. Roughly the same share intend to do so within the next year.
Gently raise these issues with a technologist and they will look at you with a mixture of disappointment and pity. Haven’t you heard of the “hype cycle”? This is a term popularised by Gartner, a research firm—and one that is common knowledge in the Valley. After an initial period of irrational euphoria and overinvestment, hot new technologies enter the “trough of disillusionment”, the argument goes, where sentiment sours. Everyone starts to worry that adoption of the technology is proceeding too slowly, while profits are hard to come by. However, as night follows day, the tech makes a comeback. Investment that had accompanied the wave of euphoria enables a huge build-out of infrastructure, in turn pushing the technology towards mainstream adoption. Is the hype cycle a useful guide to the world’s ai future?
It is certainly helpful in explaining the evolution of some older technologies. Trains are a classic example. Railway fever gripped 19th-century Britain. Hoping for healthy returns, everyone from Charles Darwin to John Stuart Mill ploughed money into railway stocks, creating a stockmarket bubble. A crash followed. Then the railway companies, using the capital they had raised during the mania, built the track out, connecting Britain from top to bottom and transforming the economy. The hype cycle was complete. More recently, the internet followed a similar evolution. There was euphoria over the technology in the 1990s, with futurologists predicting that within a couple of years everyone would do all their shopping online. In 2000 the market crashed, prompting the failure of 135 big dotcom companies, from garden.com to pets.com . The more important outcome, though, was that by then telecoms firms had invested billions in fibre-optic cables, which would go on to became the infrastructure for?today’s internet.

My take on this: We’ll only know in the future.


OpenAI makes fine-tuning available for GPT-4o

From the OpenAI blog :

OpenAI

Developers can now fine-tune GPT-4o with custom datasets to get higher performance at a lower cost for their specific use cases. Fine-tuning enables the model to customize structure and tone of responses, or to follow complex domain-specific instructions. Developers can already produce strong results for their applications with as little as a few dozen examples in their training data set.
From coding to creative writing, fine-tuning can have a large impact on model performance across a variety of domains. This is just the start—we’ll continue to invest in expanding our model customization options for developers.

My take on this: This one is a very capable model and finetuning it should allow you to build powerful applications.


Condé Nast Signs Deal With OpenAI

Kate Knibbs writing for Wired :

WIRED Kate Knibbs

Condé Nast and OpenAI have struck a multi-year deal that will allow the AI giant to use content from the media giant’s roster of properties—which includes the New Yorker, Vogue, Vanity Fair, Bon Appetit, and, yes, WIRED. The deal will allow OpenAI to surface stories from these outlets in both ChatGPT and the new SearchGPT prototype .
“It’s crucial that we meet audiences where they are and embrace new technologies while also ensuring proper attribution and compensation for use of our intellectual property,” Condé Nast CEO Roger Lynch wrote in a company-wide email. Lynch pointed to ongoing turmoil within the publishing industry while discussing the deal, noting that technology companies have made it harder for publishers to make money, most recently with changes to traditional search.
“Our partnership with OpenAI begins to make up for some of that revenue, allowing us to continue to protect and invest in our journalism and creative endeavors,” he wrote.
Lynch testified before Congress earlier this year on how AI companies like OpenAI trained their models, speaking in favor of licensing. He has previously been a vocal opponent of AI companies using content without first seeking permission, describing said data as “stolen goods.” After WIRED reported earlier this year on the web-scraping practices of the AI search engine startup Perplexity, Condé Nast sent a cease-and-desist letter demanding that the company cease using its content.
Specific terms of the partnership have not been disclosed. OpenAI declined to comment on the deal’s terms.

My take on this: Content creators and AI companies seem to be settling on something.


Midjourney Web is now open to all

From the Midjourney blog :

Midjourney

We've opened up our web platform to everyone! Now you can explore and create in a custom-built experience. Check out an overview of how to use the site.

My take on this: Midjourney was previously available publicly via Discord and to very few users on the web. Now, the web version is open to all.


Outport: Just-in-Time Model Hot-Swap

From their demo site :

Hot swapping enables efficient handling of fluctuating traffic loads across services and models.

It allows:

  • Spinning up/down different models on a single GPU node to manage variable demand
  • Loading models onto GPUs only when needed
  • Utilizing GPU resources for other tasks during low-traffic periods

In this demo, all models are cached into memory and loaded onto GPU just-in-time before generation - enabling a ~2 second load time. See in realtime how we can swap between four ~15GB models using a small L4 GPU with 22 GB of VRAM

My take on this: Very impressive demo. This can help with maximizing any GPU instances you might be using. At the minimum, you can load your models into GPU real quick.


If you've made it this far and follow my newsletter, please consider exploring the platform we're currently building: Unstract —a no-code LLM platform that automates unstructured data workflows.


For the extra curious

要查看或添加评论,请登录

社区洞察

其他会员也浏览了