Close Menu
    Trending
    • Why I Still Don’t Believe in AI. Like many here, I’m a programmer. I… | by Ivan Roganov | Aug, 2025
    • The Exact Salaries Palantir Pays AI Researchers, Engineers
    • “I think of analysts as data wizards who help their product teams solve problems”
    • These 5 Programming Languages Are Quietly Taking Over in 2025 | by Aashish Kumar | The Pythonworld | Aug, 2025
    • Chess grandmaster Magnus Carlsen wins at Esports World Cup
    • How I Built a $20 Million Company While Still in College
    • How Computers “See” Molecules | Towards Data Science
    • Darwin Godel Machine | Nicholas Poon
    AIBS News
    • Home
    • Artificial Intelligence
    • Machine Learning
    • AI Technology
    • Data Science
    • More
      • Technology
      • Business
    AIBS News
    Home»Artificial Intelligence»Do You Really Need a Foundation Model?
    Artificial Intelligence

    Do You Really Need a Foundation Model?

    Team_AIBS NewsBy Team_AIBS NewsJuly 16, 2025No Comments10 Mins Read
    Share Facebook Twitter Pinterest LinkedIn Tumblr Reddit Telegram Email
    Share
    Facebook Twitter LinkedIn Pinterest Email


    are in every single place — however are they at all times the appropriate selection? In at present’s AI world, it looks as if everybody needs to make use of basis fashions and brokers.

    From GPT to CLIP to SAM, corporations are racing to construct purposes round giant, general-purpose fashions. And for good purpose: these fashions are highly effective, versatile, and sometimes straightforward to prototype with. However do you really want one?

    In lots of instances — particularly in manufacturing eventualities — an easier, custom-trained mannequin can carry out simply as properly, if not higher. With decrease price, decrease latency, and extra management.

    This text goals that will help you navigate this determination by masking:

    • What basis fashions are, and their execs and cons
    • What {custom} fashions are, and their execs and cons
    • How to decide on the appropriate strategy primarily based in your wants, with actual world examples
    • A visible determination framework to wrap all of it up

    Let’s get into it.

    Basis Fashions

    A basis mannequin is a big, pretrained mannequin skilled on huge datasets throughout a number of domains. These fashions are designed to be versatile sufficient to resolve a variety of downstream duties with little or no extra coaching. They are often seen as generalist fashions.

    They arrive in varied varieties:

    • LLMs (Massive Language Fashions) akin to GPT-4, Claude, Gemini, LLaMA, Mistral… We hear lots about them for the reason that launch of ChatGPT.
    • VLMs (Imaginative and prescient-Language Fashions) akin to CLIP, Flamingo, Gemini Imaginative and prescient… They now are usually used an increasing number of, even in options like ChatGPT.
    • Imaginative and prescient-specific fashions akin to SAM, DINO, Steady Diffusion, FLUX. They’re a bit extra specialised and principally utilized by practitioners, but extraordinarily highly effective.
    • Video-specific fashions akin to RunwayML, SORA, Veo… This discipline has made unbelievable progress within the final couple of years, and is now reaching spectacular outcomes.

    Most are accessible by APIs or open-source libraries, and plenty of assist zero-shot or few-shot studying.

    These fashions are normally skilled at a scale that’s simply not reachable by most corporations, each when it comes to knowledge and computing energy. That makes them actually engaging for a lot of causes:

    • Normal-purpose and versatile: One mannequin can deal with many alternative duties.
    • Quick to prototype with: No want on your personal dataset or coaching pipeline.
    • Pretrained on huge, various knowledge: They encode world data and normal reasoning.
    • Zero/few-shot capabilities: They work moderately properly out of the field.
    • Multimodal and versatile: They’ll generally deal with textual content, pictures, code, audio, and extra, which will be exhausting to breed for small groups.

    Whereas they’re highly effective, they arrive with some drawbacks and limitations:

    • Excessive operational price: Inference is pricey, particularly at scale.
    • Opaque conduct: Outcomes will be exhausting to debug or clarify.
    • Latency limitations: These fashions are usually very giant and have excessive latency, which will not be ideally suited for real-time purposes.
    • Privateness and compliance issues: Knowledge typically must be despatched to third-party APIs.
    • Lack of management: Troublesome to fine-tune or optimize for particular use instances, generally not even an possibility.
    Professionals and cons of basis fashions. Picture by creator.

    To recap, basis fashions are very highly effective: they’re skilled on huge datasets, can deal with textual content, picture, video and extra. They don’t must be skilled in your knowledge to work. However they’re normally not price efficient, might have excessive latency and will required sending your knowledge to 3rd events.

    The choice is to make use of {custom} fashions. Let’s now see what which means.

    Customized Fashions

    A {custom} mannequin is a mannequin constructed and skilled particularly for an outlined process utilizing your individual knowledge. This may very well be so simple as a logistic regression or as complicated as a deep studying structure tailor-made to your distinctive downside.

    They typically require extra upfront work however provide better management, decrease price, and higher efficiency on slender duties. Many highly effective and business-driving fashions are literally {custom} fashions, some well-known and broadly used, some addressing actually area of interest issues:

    • Netflix’s suggestion engine, utilized by billions, is a {custom} mannequin
    • Most churn prediction fashions, broadly utilized in many subscription-based corporations, are {custom} fashions (generally only a well-tuned logistic regression)
    • Credit score scoring fashions

    When utilizing {custom} fashions, you grasp each single step, making them actually highly effective for a number of causes:

    • Job-specific and optimized: You management the mannequin, the coaching knowledge, and the analysis.
    • Decrease latency and value: Customized fashions are normally smaller and cheaper. It’s crucial in edge or real-time environments.
    • Full management and explainability: They’re simpler to debug, retrain, and monitor.
    • Higher for tabular or structured knowledge: Basis fashions excel with unstructured knowledge. Customized fashions are likely to do higher on tabular knowledge.
    • Improved knowledge privacy: No must ship knowledge to exterior APIs.

    However, it’s a must to prepare and deploy your {custom} fashions your self to get enterprise worth out of them. It comes with some drawbacks:

    • Labeled knowledge could also be required: Which will be costly or time-consuming to get.
    • Slower to develop: Customized fashions require coaching a mannequin, implement pipelines, deploy and keep. That is time consuming.
    • Expert sources wanted: In-house ML experience is a should.

    Be happy to dig into deployment methods and the way to decide on the most effective strategy in that article:

    Professionals and cons for {custom} fashions. Picture by creator.

    In a single phrase, {custom} fashions give extra management and are normally cheaper to scale. However it comes at the price of a costlier and longer improvement section — to not point out the abilities. Then how to decide on properly whether or not to make use of a {custom} mannequin or a basis mannequin? Let’s attempt to reply that query.

    Basis Mannequin or Customized Mannequin: Learn how to Select?

    When to Select a Customized Mannequin

    I might say {that a} {custom} mannequin should be the default selection total. However to be extra truthful, let’s see in what particular instances it’s clearly a greater resolution than a basis mannequin. It comes down just a few necessities:

    • Groups & Assets: you’ve a machine studying engineer or knowledge crew, you possibly can label or generate coaching knowledge, and also you’re in a position to spend time coaching and optimizing your mannequin
    • Enterprise: both you’ve a extremely particular case to resolve, you’ve privateness necessities, you want low infra price, otherwise you want low latency and even edge deployment
    • Lengthy-term targets: you need management, and also you don’t need to depend on third-party APIs

    If you end up in a number of of those conditions, a {custom} mannequin could also be the best choice. Some typical examples I confronted in my profession have been in that scenario, for instance:

    • Constructing an in-house, {custom} forecasting mannequin for YouTube video income: you possibly can’t compromise on privateness, and no basis mannequin will do properly sufficient on such particular use instances
    • Deploying real-time video resolution on smartphone: when it is advisable to work at greater than 30 frames per second, no VLM can deal with the duty but
    • Credit score scoring for a financial institution: you possibly can’t compromise on privateness, and might’t use third-party options

    If you wish to dig into it, right here is an article about the right way to forecast YouTube video income:

    How Jellysmack Monetized YouTube Videos with Predictive Algorithms
    A Revolutionary Idea in the Creator Economy

    That being stated, whereas in some instances basis fashions usually are not the answer, let’s see after they really are a viable possibility.

    When to Select a Basis Mannequin

    Let’s make the equal train for basis fashions: let’s first test the necessities that make them an excellent possibility, and let’s take a look at some typical enterprise instances the place they’d thrive:

    • Crew & Assets: you don’t essentially have labeled knowledge, nor ML engineers or knowledge scientists, however you do have AI or Software program engineers
    • Enterprise: you need to check an concept rapidly or ship an MVP, you’re effective with utilizing exterior APIs, and latency or scaling price aren’t main issues
    • Job Traits: your process is open-ended, otherwise you’re exploring a novel or artistic downside house

    Listed below are some typical examples the place basis fashions have confirmed worthwhile

    • Prototyping a chatbot for inner assist or data administration: you’ve an open-ended process, with low necessities on latency and scale
    • Many early-stage MVPs with out long-term infra issues are good candidates

    As of now, basis fashions are actually well-liked for a lot of MVPs revolving round textual content and picture, whereas {custom} fashions have confirmed their worth in lots of enterprise instances. However why not combining each? In some instances, it’s potential to get the most effective options with hybrid approaches. Let’s see what which means.

    When to Use Hybrid Options

    In lots of real-world workflows, the most effective reply is a mixture of each approaches. For instance, listed below are just a few frequent hybrid patterns that may leverage the most effective of each worlds

    • Basis mannequin as a labeling device: use SAM or GPT to create labeled knowledge, then prepare a smaller mannequin.
    • Data distillation: prepare a {custom} mannequin to imitate the outputs of a basis mannequin.
    • Bootstrapping: begin with basis mannequin to check, then change to {custom} later.
    • Function extraction: use CLIP or GPT embeddings as enter to an easier downstream mannequin.

    I used a few of these approaches in previous tasks throughout my profession, and so they generally enable to get state-of-the-art options, utilizing the generalistic energy of basis fashions and the flexibleness and scalability of {custom} fashions.

    • In laptop imaginative and prescient tasks, I used Steady Diffusion to create various and lifelike datasets, in addition to SAM to annotate knowledge rapidly and effectively
    • Small Language Fashions are getting traction, and generally get benefit of information distillation to get the most effective out of LLMs whereas remaining smaller, extra specialised and extra scalable
    • One may also use instruments like ChatGPT to simply annotate knowledge at scale earlier than coaching {custom} fashions

    Here’s a concrete instance of utilizing basis fashions in hybrid options for laptop imaginative and prescient:

    In a phrase, in lots of instances when coping with unstructured knowledge, a hybrid strategy will be highly effective and provides the most effective of each worlds.

    Conclusion: Determination Framework

    Let’s now summarize with a call chart when to go for a basis mannequin, when to go for a {custom} mannequin, and when to discover a hybrid strategy.

    Determination chart to decide on the appropriate strategy: {custom} mannequin, basis mannequin or hybrid. Picture by creator.

    In just a few phrases, all of it comes all the way down to the venture and the necessity. Certain, basis fashions are buzzing proper now, and they’re on the coronary heart of the present brokers revolution. Nonetheless, many very worthwhile enterprise issues will be addressed with {custom} fashions, whereas basis fashions are confirmed highly effective in lots of unstructured knowledge issues. To decide on properly, a correct evaluation of the wants and necessities with stakeholders and engineers, together with a call framework stays an excellent resolution.

    What about you: have you ever confronted any scenario the place the most effective resolution will not be what you may suppose?

    References

    • Talked about LLMs: GPT by OpenAI, Claude by Anthropic, Llama by Meta, Gemini by Google, and we may cite extra akin to Mistral, DeepSeek, and so on…
    • Imaginative and prescient-related fashions: SAM by Meta, CLIP by OpenAI, DINO by Meta, StableDiffusion by StabilityAI, FLUX by Black Forest Labs
    • Video-specific fashions: Veo by Google, RunwayML, SORA by OpenAI…



    Source link

    Share. Facebook Twitter Pinterest LinkedIn Tumblr Email
    Previous ArticlePerceptron Learning Algorithm Explained | by Prathik C | Jul, 2025
    Next Article Fyre Festival IP, Assets Sell on eBay for $250K: ‘So Low’
    Team_AIBS News
    • Website

    Related Posts

    Artificial Intelligence

    “I think of analysts as data wizards who help their product teams solve problems”

    August 2, 2025
    Artificial Intelligence

    How Computers “See” Molecules | Towards Data Science

    August 2, 2025
    Artificial Intelligence

    Mastering NLP with spaCy – Part 2

    August 1, 2025
    Add A Comment
    Leave A Reply Cancel Reply

    Top Posts

    Why I Still Don’t Believe in AI. Like many here, I’m a programmer. I… | by Ivan Roganov | Aug, 2025

    August 2, 2025

    I Tried Buying a Car Through Amazon: Here Are the Pros, Cons

    December 10, 2024

    Amazon and eBay to pay ‘fair share’ for e-waste recycling

    December 10, 2024

    Artificial Intelligence Concerns & Predictions For 2025

    December 10, 2024

    Barbara Corcoran: Entrepreneurs Must ‘Embrace Change’

    December 10, 2024
    Categories
    • AI Technology
    • Artificial Intelligence
    • Business
    • Data Science
    • Machine Learning
    • Technology
    Most Popular

    How Do the iPhone 16E and Google Pixel 9A Compare to More Expensive Models?

    April 10, 2025

    Former Google Engineer Risks Everything on Brain Tech

    April 11, 2025

    Secure Your Business with 320+ Hours of Cybersecurity Courses for $60

    January 25, 2025
    Our Picks

    Why I Still Don’t Believe in AI. Like many here, I’m a programmer. I… | by Ivan Roganov | Aug, 2025

    August 2, 2025

    The Exact Salaries Palantir Pays AI Researchers, Engineers

    August 2, 2025

    “I think of analysts as data wizards who help their product teams solve problems”

    August 2, 2025
    Categories
    • AI Technology
    • Artificial Intelligence
    • Business
    • Data Science
    • Machine Learning
    • Technology
    • Privacy Policy
    • Disclaimer
    • Terms and Conditions
    • About us
    • Contact us
    Copyright © 2024 Aibsnews.comAll Rights Reserved.

    Type above and press Enter to search. Press Esc to cancel.