Microsoft on how custom AI offers your business better answers, lower costs, faster innovation

2 hours ago 5
blueeleavtor-gettyimages-1254933178
4045/Getty Images

Large connection models similar ChatGPT's GPT-4o look to person each the accusation successful the known universe, oregon astatine slightest what engineers could scan disconnected the internet.

But what if you privation to usage a ample connection exemplary (LLM) with proprietary accusation from your ain institution data, oregon specialized accusation that's not publically disposable connected the internet, oregon different bid an LLM to person specialized knowledge?

Do you physique an LLM from scratch? Do you usage a small, open-source, self-hosted exemplary that contains lone your information?

Also: A fewer secretive AI companies could crush escaped society, researchers warn

As it turns out, you tin commencement with an LLM similar GPT-4o, and past physique up connected apical of that. That's called a customized AI.

In this article, Eric Boyd, Microsoft firm vice president for AI platforms, shares with ZDNET astir however Microsoft makes customized AI imaginable for their customers, what goes into a customized model, what the full process involves, and immoderate champion practices.

ericboydscreenshot-2025-05-05-130555

Eric Boyd, Microsoft firm vice president for AI platforms.

Microsoft

Let's get started.

ZDNET: Can you present yourself and supply an overview of your relation astatine Microsoft and with its AI platform?

Eric Boyd: I pb the AI level squad astatine Microsoft. It has been a brainsick mates of years successful the AI space.

I started moving astatine Microsoft successful 2009 successful the Bing organization, and it has been phenomenal seeing things germinate from there, due to the fact that truthful overmuch of Microsoft's AI innovation started with Bing. We built the infrastructure to bid AI models, to iterate and experimentation to spot which AI exemplary was performing best. And each that infrastructure turned into pieces and components of things that we present service done Azure AI Foundry.

Through Azure AI Foundry, we assistance companies entree everything from thousands of GPUs to physique and bid their ain AI models, to the tools needed to negociate that, to a catalog of AI models, ample and small, unfastened and frontier, which we connection via our concern with OpenAI and different providers.

We besides supply tools to physique applications connected apical of these AI models, including a wide scope of capabilities our customers request to marque definite they tin bash truthful responsibly.

Ultimately, my squad is focused connected gathering Azure AI Foundry truthful it includes everything a lawsuit oregon developer mightiness request to physique their AI solutions, and easy determination from thought to implementation successful a unafraid and trusted way.

Generative AI vs. customized AI

ZDNET: So, past twelvemonth we had generative AI. Now we person customized AI. What is it, and wherefore isn't generative AI enough?

EB: As companies person started to deploy applications, generative AI and the base instauration models person gotten them beauteous far. But galore are uncovering country cases wherever the basal instauration models don't reply ace well.

Also: The champion AI chatbots: ChatGPT, Copilot, and notable alternatives

So customized AI is simply a company's quality to usage its ain information to customize their halfway exemplary to get amended prime answers to questions — and successful immoderate cases they tin usage a little outgo model.

ZDNET: What are the cardinal advantages of customized AI implicit off-the-shelf generative AI solutions?

EB: Quality and outgo are the 2 superior advantages. With customized AI, you tin amended the prime of your application's answers by uncovering wherever the instauration exemplary is anemic and past fine-tuning the response. Fine-tuning besides lets you, successful immoderate cases, usage a lower-cost exemplary to execute higher-cost-model quality.

ZDNET: Can you stock examples of however businesses person successfully implemented customized AI solutions?

EB: Microsoft is wide applying this method crossed our tech stack, arsenic we often enactment arsenic our ain "customer zero," which has enabled america to experiment, learn, and hone cutting-edge champion practices. GitHub Copilot and Nuance DAX were some extensively fine-tuned and customized with specialized coding output and healthcare knowledge. As the prime of the output increases, truthful does adoption.

DAX Copilot has present surpassed 2 cardinal monthly physician-patient encounters, up 54% quarter-over-quarter, and is being utilized by apical providers similar Mass General Brigham, Michigan Medicine and Vanderbilt University Medical Center. By fine-tuning to this circumstantial data, the solution does a amended occupation producing a aesculapian grounds arsenic opposed to conscionable summarizing a doctor-patient conversation.

Also: Want to triumph successful the property of AI? You tin either physique it oregon physique your concern with it

We're successful a unsocial presumption with galore AI applications crossed the suite of Microsoft products, and successful gathering those, we've learned a batch astir what radical privation to bash next. By knowing however assorted techniques person helped our ain applications, we person a coagulated imaginativeness for however this is going to assistance our customers' applications.

ZDNET: What proposal would you springiness to companies conscionable opening their AI customization journey?

EB: I mostly promote companies to beryllium their usage lawsuit works utilizing the astir almighty instauration exemplary possible, and past look astatine steps to either amended prime oregon trim cost.

Customization would beryllium a method for some of those. For this, they'll request to person utilized their exertion capable to cognize its imaginable weaknesses, wherever the exemplary and information are not answering the questions arsenic they privation them to, and commencement collecting that information and gathering the repository for what they privation the exemplary to do. That's yet going to beryllium the information we usage to customize the model.

Also: Autonomous businesses volition beryllium powered by AI agents

In the epoch of AI, information is simply a changemaker arsenic these systems necessitate high-quality, accessible and unafraid information to relation properly. Making definite they person that information is simply a cardinal portion of customizing the model. We are moving to assistance customers modernize their information to the cloud, and unify their information estates to physique the adjacent procreation of intelligent apps.

Optimize your AI investment

ZDNET: What are the outgo implications of processing and maintaining customized AI solutions, and however tin companies optimize their investments?

EB: The outgo of fine-tuning the exemplary is often comparatively humble but an important concern arsenic determination are besides costs for collecting the information and past grooming the model. Customers besides request to see the lifespan of the model.

When fine-tuning, we suggest starting with a foundational exemplary (GPT-4o, oregon the like) to customize. When the next-generation exemplary comes out, you tin either take "I'm going to support my customized model" oregon "I americium going to re-customize the next-generation model."

Also: AI cause deployments volition turn 327% during the adjacent 2 years. Here's what to bash now

Keeping your information acceptable volition marque that consequent customization easier, but you would person to bash it again. Although that is thing to consider, don't beryllium acrophobic due to the fact that the interaction depends connected the gait of exemplary innovation.

We can't accidental what the aboriginal holds for caller exemplary capabilities, but customers who fine-tuned GPT-4o a twelvemonth agone would apt beryllium blessed with their solution today, contempt advancements successful reasoning models similar the o1 series.

ZDNET: What are the astir communal hurdles organizations look erstwhile implementing customized AI, and however tin they flooded them?

EB: To customize models, you request information that addresses wherever successful your exertion you privation improvement. Having wide information connected your exemplary apt won't get you to that adjacent level. You request information wherever your exertion isn't performing arsenic you want, truthful you tin find however to amended it.

In the past, astir companies person not been accustomed to doing this, truthful it's a caller musculus to build. Although determination are tools and techniques to automate that, galore companies don't person the radical who cognize however to, truthful they request to put successful processing those skills archetypal and foremost, and past enactment connected applying them

ZDNET: What ethical considerations should organizations support successful caput erstwhile deploying customized AI?

EB: I don't deliberation customized AI brings caller ethical considerations. It's the aforesaid acceptable of things you indispensable see broadly with generative AI. It's "Here's this exertion I've developed. How americium I going to marque definite it behaves responsibly for my brand, for my applications, and for the imaginable implications of however this exertion volition get used?"

Also: The champion AI for coding (including 2 caller apical picks - and what not to use)

All the things that we screen successful our Responsible AI Standard for however we deliberation radical should behave should inactive beryllium considered. One of the benefits of utilizing our level to make and deploy your AI applications is that Microsoft offers tools like Azure AI Content Safety that enactment with the customized models, truthful customers tin beryllium assured their systems are liable by design.

Bias, fairness, and transparency

ZDNET: How does Microsoft code concerns astir bias, fairness, and transparency successful customized AI models?

EB: Today, we connection implicit 30 tools and 100 features to assistance our customers, developers, and researchers responsibly physique with AI. Though Azure AI Content Safety is embedded by default successful each models successful the Azure AI Foundry catalog, preventing misuse and maltreatment astatine the exemplary level unsocial is astir impossible. That's wherefore it's imperative to besides person systems and tools that assistance you trial and show each measurement of the way, before, during, and aft deployment.

Microsoft aims to assistance customers done each furniture of generative AI hazard mitigation. We person tools to assistance users map, measure, mitigate, monitor, respond, and govern. We are looking astatine this from the strategy level, the idiosyncratic level, and the exemplary level. We are continuing to put successful probe connected identifying, measuring, and mitigating antithetic types of fairness-related harms, and we are innovating successful caller ways to proactively trial our AI systems, arsenic outlined successful our Responsible AI Standard.

ZDNET: How does Microsoft Azure enactment businesses successful tailoring AI models to their circumstantial needs?

EB: We've been gathering systems into Azure AI Foundry to simplify this process. There's the fine-tuning work itself, and observability services that marque it easier to cod information connected applications, which successful crook tin beryllium utilized for fine-tuning.

ZDNET: What relation does open-source AI play successful the customization and scalability of AI solutions?

EB: We've seen a batch of innovation successful the open-source exemplary space, mostly astatine little terms points (and truthful little prime points). But those lower-cost models are often bully places to commencement due to the fact that you tin trial and experimentation to spot if you tin execute the prime you'd get with a higher-priced model.

Also: I tested 10 AI contented detectors - and these 5 correctly identified AI substance each time

In general, the innovation successful this abstraction has brought a batch of exemplary assortment into the Azure AI Foundry exemplary catalog, which customers tin measure against, and take the champion exemplary for their usage case.

ZDNET: What are the cardinal differences betwixt fine-tuning existing AI models and gathering AI solutions from scratch?

EB: It's massively costly to physique your ain exemplary from scratch, whereas fine-tuning is rather tenable for astir applications. Cost would beryllium the superior difference. But if you're conscionable gathering a modular AI solution utilizing a accepted instauration exemplary (not a customized model), the superior quality is that you whitethorn sacrifice prime and/or price, the 2 main levers you're optimizing for.

Agents are the apps of the AI era

ZDNET: What interaction bash you foresee AI copilots having connected endeavor AI strategies?

EB: Large connection models person changed however concern gets done successful enterprises, and we spot that lone continuing to accelerate. With our customers, we're progressively seeing them physique applications that execute tasks for radical and implicit work, and get it done for them, arsenic opposed to conscionable answering a question.

Also: What are AI agents? How to entree a squad of personalized assistants

This is the displacement toward AI agents being discussed. Agents are the apps of the AI era. Every enactment of concern strategy contiguous is going to get reimagined arsenic an cause that sits connected apical of a copilot. That is going to transform ample swaths of antithetic concern processes.

ZDNET: How should organizations equilibrium AI automation with quality oversight to guarantee optimal outcomes?

EB: This is simply a cardinal question. These models bash galore things, but not everything well. Ensuring we recognize their capabilities and person radical yet accountable for the enactment that gets done indispensable beryllium a cardinal portion of liable AI policies, and a cardinal portion of however we urge applications beryllium built.

Also: Why scaling agentic AI is simply a marathon, not a sprint

The tone of Microsoft's AI tools is astir advancing quality agency, putting the quality astatine the center, and being grounded successful their context. We are creating platforms and tools that, alternatively than acting arsenic a substitute for quality effort, tin assistance humans with cognitive work.

ZDNET: If you could connection 1 cardinal takeaway to concern leaders exploring customized AI, what would it be?

EB: As AI applications go a larger portion of each business's portfolio, they volition miss retired if they don't deliberation done their customization strategy to guarantee the highest-quality, best-performing applications astatine the champion price.

For companies wanting to get started contiguous with customized AI, I say: Look astatine your generative AI application, people wherever successful that exertion you privation to improve, cod immoderate data, and springiness it a shot.

ZDNET: How bash you spot the aboriginal of AI evolving beyond customized AI, and what's the adjacent large displacement connected the horizon?

EB: We've spent the past 2 years gathering applications that cognize however to usage your information to assistance you reply a question and past springiness you a substance reply back. I deliberation we're going to walk the adjacent 2 years gathering applications that execute portion of the enactment for you.

Also: You'll soon negociate a squad of AI agents, says Microsoft's Work Trend report

In this scenario, you tin delegate tasks and expect them to get done, sometimes autonomously via agents, versus successful a synchronous chat conversation. But agents are conscionable a ample connection exemplary exertion that you tin inquire to bash enactment and execute actions.

Within those applications, you volition inactive find places wherever customized models volition amended the prime of the system, adjacent erstwhile the compute is happening down the scenes.

Have you explored customized AI?

What astir your organization? Have you explored enterprise-grade AI customization yet? What challenges oregon opportunities bash you spot successful tailoring instauration models to your ain data? Are you considering fine-tuning models similar GPT-4o oregon moving with open-source alternatives? What relation bash you deliberation agents and copilots volition play successful your concern strategy? Let america cognize successful the comments below.


You tin travel my day-to-day task updates connected societal media. Be definite to subscribe to my play update newsletter, and travel maine connected Twitter/X astatine @DavidGewirtz, connected Facebook astatine Facebook.com/DavidGewirtz, connected Instagram astatine Instagram.com/DavidGewirtz, connected Bluesky astatine @DavidGewirtz.com, and connected YouTube astatine YouTube.com/DavidGewirtzTV.

Want much stories astir AI? Sign up for Innovation, our play newsletter.

Read Entire Article