Fine-Tuning, Prompt Engineering, and Embeddings: How to Combine These Techniques for Optimal AI Performance
Przemek Majewski
Living with Diabetes | AI Strategist | DLabs.AI CEO | Ex-CERN
Hi, welcome to the first edition of Bits and Bytes in 2024!
Can you believe it's been two years since this newsletter launched? Time flies, especially in the world of AI. It's been a thrilling ride to explore and share the rapid developments in technologies like ChatGPT, Midjourney, and other AI breakthroughs.?
It makes me wonder if, one day, these updates will be part of history lessons, showcasing how AI reshaped our world. We certainly live in interesting times, and they promise to get even more intriguing.
In this edition, we're covering a range of exciting topics:
So, let's dive into these fascinating developments!
What's up with DLabs.AI: A Brief Recap of 2023 (and Plans for 2024)
An intense year is behind us. As far as AI development is concerned, I feel it was the most groundbreaking year of our generation. The development of LLM and Generative AI has transformed the technology market, which simultaneously struggled with a crisis resulting in many layoffs, investment freezes, and other challenges.
It was also a challenging year at DLabs.AI. On the one hand, we had to accelerate our pace to adapt to the changing AI landscape and the influx of companies that, following the implementation of GPT and the "hype" around AI, suddenly tried to position themselves as AI experts (remember the meme about what's the fastest thing in the world? If you missed it, I've included it below):
Did we succeed? There have certainly been ups and downs, but we've strived to develop our competencies and deliver as much value as possible through AI implementation.
First and foremost, we created three GPT-powered Agents of our own experimental prototypes: SugarAssist (Diabetes Assistant), DLeads (B2B Sales Outreach Agent), and Student Assistant (to help choose the best education and career path).
It allowed us to gain a deeper understanding of how LLMs work, which we quickly put to use with clients such as Boldly (for whom we built an AI-based virtual assistant capable of extracting crucial information from emails and streamlining meeting organization), among others.
We've also decided to dive deeper into understanding our clients: both from a marketing perspective (through in-depth analysis of sales meetings, sales funnels, buyer personas, etc.) and from a business perspective by offering additional tailored services.
I'll tell you more about that in the next newsletter (we're currently finessing the offer). But being the type of people who like to put ideas into practice quickly, we've conducted numerous experiments to better understand the market and adjust our business accordingly.?
That's why I'm looking at 2024 with optimism, eagerly anticipating the exciting projects we have on the horizon.
Fine-Tuning, Prompt Engineering, and Embeddings: Understanding the Key Differences
As alluded to above, DLabs.AI has been working with LLMs a lot in recent months. And as the person liaising with customers on the daily, I'm often asked to explain the differences between fine-tuning, prompt engineering, and embeddings (alongside when to use each one).?
My answer? Well, the truth is that each serves a unique purpose in customizing and enhancing the capabilities of AI models, and when used collectively, they can address complex tasks with remarkable finesse.
Since the question crops up a lot, I decided to dive into these concepts, elucidating their differences, synergies, and practical applications.
Fine-Tuning: Tailoring AI to Specific Needs
Fine-tuning transforms a generalist pre-trained model into a specialist. It's like sharpening a knife: the model, initially broad in knowledge, is honed for a specific domain through additional training on a task-specific dataset.
Example: Fine-tuning GPT on medical texts transforms it from a general language model to a specialist in medical diagnoses.
Prompt Engineering: The Art of Asking Right
Prompt engineering is designing input prompts to guide a pre-trained model to the desired output. It's about crafting language to steer the AI in a specific direction.
Example: To get a poem about the ocean, a precise prompt like "compose a haiku about the tranquil ocean at dawn" yields better results than a vague "write a poem."
Embeddings: The Linguistic DNA of AI
Embeddings are numerical representations of text, turning language into a vector space.
Example: The word "apple" becomes a vector, associated with related concepts like "fruit" and "red," distinct from unrelated ones like "astronomy."
How To Combine These Methods for Enhanced AI Customization
Fine-Tuning with Embeddings: During fine-tuning, the embeddings within the model are refined. So, our medical GPT, initially familiar with "apple" in a dietary context, begins to understand "apple" in terms of its health benefits or as part of patient dietary advice.
Prompt Engineering with Embeddings: Skilled prompt engineering makes use of the model's embeddings. By understanding which prompts lead to vectors that produce better results, we can craft inputs that align with the desired output (for instance, using "apple" in a prompt about healthy foods guides the AI to discuss nutritional aspects).
As you can see, fine-tuning, prompt engineering, and embeddings are integral to customizing AI. They enable specialized applications, expanding AI’s capabilities in various domains. Interested in learning more about the subject?
Feel free to join my webinar, How to Build a GPT-Based System: From Idea to Deployment.
AI's Legal Drama: OpenAI and Microsoft Caught in Copyright Crosshairs
Let's start the year with a drama-like turn of events reminiscent of a Hollywood plot. OpenAI has found itself in the legal crosshairs over ChatGPT.?
A group of authors (including Nicholas Basbanes and Nicholas Gage) have accused the tech giants of using their copyrighted works without consent to train ChatGPT and other AI models. Their lawyer, Michael Richter, has criticized the companies for leveraging these works to fuel a billion-dollar industry without proper remuneration.
Adding to the intrigue, this case is part of a broader wave of similar lawsuits. Notable fiction and nonfiction writers, from comedian Sarah Silverman to "Game of Thrones" creator George R.R. Martin, have also challenged tech companies over the use of their creative output in AI program training.
This legal entanglement seems to be just the tip of the iceberg as more authors and media institutions step forward to question the tech giants' use of their work. A few days ago, The New York Times launched its own lawsuit against OpenAI and Microsoft, alleging that their use of its journalistic content to train AI chatbots could result in substantial financial losses.?
While OpenAI argues that such usage falls under the umbrella of 'fair use,' The Times contends that it infringes on copyright, potentially misleads the public, and undermines the integrity of their original content.
Well, these lawsuits certainly bring into focus crucial issues regarding 'fair use' in the context of AI development and the extent to which copyrighted materials can be utilized in AI training. The ramifications of these legal battles could be profound, shaping the future landscape of AI development and safeguarding intellectual property rights.
And by the way, have you ever had a déjà vu moment with ChatGPT's responses? Sometimes, I can't shake off the feeling that I might have had a sneak peek at the DLabs.AI blog. How about you??
Notice any uncanny echoes of your own content in its replies??
Art or AI? Midjourney v6 Blurs the Line
Just before the holidays, Midjourney introduced the alpha version of its latest image synthesis model, v6, capturing the attention of AI enthusiasts and artists alike. During the winter break, the AI community dived into exploring this new version, showcasing their experiments on various social platforms.
领英推荐
What’s intriguing about v6? It's the leap in detail and a new take on user prompts. Users are impressed by the remarkable detail and scenery but pointed out issues with high contrast and saturation. It seems with v6, users are learning to adapt to a 'less is more' prompting style.
Now, let's talk about visuals. Below you can find some examples:
In my opinion, the sample works generated by this tool are nothing short of stunning, pushing the limits of hyper-realism. Personally, I find it both astonishing and a bit unnerving.?
The level of detail is so intense it verges on the surreal. The sheer intricacy in these images could easily make you do a double take, questioning what's real and what's AI-generated. This advancement in Midjourney v6 not only showcases the potential of AI in art but also raises questions about the future of digital imagery and its impact on our perception of reality.
However, this tool isn't without controversy. Much like the recent legal challenges faced by OpenAI and Microsoft, Midjourney is also under scrutiny for its training methods. The use of web-scraped artwork without artists' consent has sparked debates in the AI community, with some critics questioning the ethics behind these practices.
Well, we live in interesting times, where each new AI breakthrough brings both wonder and new questions to ponder. What are your thoughts on the images produced by this tool??
Are you as captivated by their intricacy as I am?
OpenAI Unveils GPT Store for Custom Chatbot Sharing
Let's return to OpenAI as it finally launches its GPT Store: a dedicated platform where users can share their custom chatbots. The store (which faced several delays and was eagerly awaited since its announcement last November) is now live.?
And the innovative platform is set to broaden the potential use cases of ChatGPT, extending OpenAI's ecosystem far beyond the company's direct offerings. The buzz around the GPT Store began with the GPT Builder program, which created over 3 million user-generated bots.?
This platform allows users to share their unique versions of ChatGPT, opening up a wealth of personalized applications. As of now, the sharing and creation of custom GPTs are limited to subscribers of OpenAI's paid tiers. You might wonder, "Why would I share my custom chatbots?"?
Well. OpenAI is set to launch a revenue-sharing program for GPT creators. Tied to user engagement, the initiative (kicking off in Q1 this year) offers financial incentives for your creative contributions. While the specifics are still under wraps, this program promises to turn your innovative chatbot designs into potential earnings.
What about security? Ahead of the store's launch, OpenAI has implemented a robust review system to ensure all custom GPTs adhere to its brand guidelines and usage policies. Users can also report any GPTs they find objectionable or unsafe. The GPT Store will initially be accessible to ChatGPT Plus and Enterprise users, including those subscribed to the newly introduced Team tier.?
This new subscription model, designed for smaller teams, offers access to advanced features like GPT-4, DALL-E 3, and specialized data analysis tools, ensuring data privacy and customization for specific team needs.
So, the launch of the GPT Store is undoubtedly exciting, but have you considered its impact on the market? In my opinion, this feature could have a significant side effect and may lead to the downfall of several smaller SaaS companies that specialized in selling prompts.? A typical scenario where a single move by a technology giant effectively "wipes out" smaller players in the industry.?
Source: OpenAI
Apple's Anticipated Leap into Generative AI
As an avid Apple fan, I've watched their role in the AI arms race with keen interest, eagerly awaiting a significant stride.?
According to Mark Gurman, a renowned tech reporter and authority on all things Apple, the tech giant is finally gearing up to make notable AI advancements with its upcoming iOS release. Set for unveiling at its 2024 Worldwide Developers Conference in June, here's a sneak peek at what’s coming:
The most anticipated update? A major overhaul for Siri. Long critiqued for its limitations, Siri's revamp could be a game-changer for iPhone users (like myself) who have long desired more reliability and functionality from the virtual assistant.
Watching Apple's journey in the AI space, it's clear they've been trailing behind giants like OpenAI, Microsoft, Google, and Samsung, all of whom have already made notable advances in AI. While it looks like we might have to wait until 2025 for Apple's full AI transformation, this announcement is a promising sign.?
It offers a hopeful glimpse into a more AI-integrated future from a company we've always admired for its innovation.
Source: ZDNET
Deloitte Uses AI to ... Reduce Workforce Layoffs
As the AI wave sweeps through the tech world, leading to significant layoffs at giants like Duolingo and Google, Deloitte LLP is taking a notably different and proactive stance.?
Faced with the consulting industry's challenge of aligning staff levels with fluctuating demands, Deloitte is smartly leveraging AI as a balancing tool. This year, while Deloitte expanded its workforce by a massive 130,000, it also faced the tough decision of signaling potential job cuts to many in the U.S. and U.K. due to a demand slowdown.?
But now Stevan Rolls, Deloitte's global chief talent officer, is focusing on deploying AI to analyze and redistribute existing employees from quieter business areas to roles facing surging demand. It's not just about staff reallocation, though.?
Deloitte is also tapping into generative AI to automate tasks typically reserved for junior staff, including prepping documents for meetings or gathering data for client pitches (think of it as having a ChatGPT-style tool for streamlining internal processes).
With a workforce nearing 460,000, Deloitte's approach isn't just about managing numbers; it's about reimagining the role of AI in workforce dynamics and underscores a vital principle: AI and human resources can complement each other to create a more dynamic, efficient, and resilient workforce.?
It's a potent reminder that the key to successfully navigating the AI landscape isn't just about who has the most advanced technology.?
It’s also about who can use it most effectively to empower their human talent.
Source: Bloomberg
AI-Assisted Breakthrough in the Discovery of New Antibiotics
In a remarkable leap forward in healthcare, scientists have harnessed AI to unlock a new class of antibiotics targeting the formidable MRSA bacteria. This discovery, the first in 60 years, truly underscores the revolution AI is bringing to drug discovery.?
It's not every day that we see such a significant breakthrough, especially against drug-resistant superbugs posing a global health threat. The team, guided by Professor James Collins at MIT, turned to deep learning models to sift through data from thousands of compounds.?
Their focused effort on MRSA, bacteria notorious for causing severe infections, is particularly noteworthy because, by narrowing down to two promising compounds from a pool of millions, they've showcased AI's extraordinary precision and efficiency.
What's really striking is how this approach could radically change the future of medicine. It's like we've just seen the tip of the iceberg in terms of what AI can achieve in healthcare. For me, it's incredibly inspiring to think about the countless lives that could be saved with AI's help.
The story isn't just about a scientific achievement; it's about opening a new chapter in how AI is revolutionizing antibiotic discovery, offering new hope against drug-resistant superbugs.
Source: Euro News
_____
That wraps up our first newsletter of the year. I hope you found the insights helpful and the tech news fascinating.
Now, as we march into 2024, I wish you all success, inspiration, and innovation. I'd also love to see you at my webinar on January 17, which will offer an incredible opportunity for us to connect and dive deeper into the world of AI.?
Until then, take care, and have a fantastic January!
Content Writer in IT (now focusing on GenAI)
2 个月So simple and clear explanation of such complex concepts! Bravo!
AI practitioner | Building Language Learning in VR | Exploring the practical side of "AI That Works"
8 个月Thanks for an amazing article, Shemmy! It's truly fascinating that you've gathered this information into an article and made it so engaging to read. Thanks for a great resource!