← Master Index
Vol. 22 Module 22.1 Lecture

Microsoft (Phi, Copilot)

Frontier LLM Providers

How This Lesson Fits the Module & Volume

Module 22.1 opened with labs that train flagships (OpenAI → Qwen). Microsoft is different: it is a distribution and SLM giant. Phi is Microsoft’s small-language-model family. Copilot is the product brand across GitHub, Microsoft 365, Windows, Security, and Copilot Studio—often running OpenAI models via Azure OpenAI / Microsoft Foundry-class platforms, not “a single Microsoft frontier LLM” you download.

This lecture closes the frontier-provider module and hands off to Module 22.2 Stability AI (image ecosystem). Deeper Copilot IDE detail lives in 22.6 GitHub Copilot. Azure as a cloud AI platform returns in Module 22.5. Vol. 21 coding assistants still require HITL diffs. Do not invent unpublished Phi training mixtures or fake Copilot seat prices.

Learning Objectives

By the end of this lesson, students should be able to:

  • Separate Phi (SLM weights / Azure SKUs) from Copilot (product suite) from Azure OpenAI (hosted OpenAI models).
  • Explain Microsoft’s role: partner + distributor + SLM lab, not a substitute for the OpenAI lecture.
  • Fill a qualitative Pricing / Strengths / Weaknesses card for Phi vs Copilot vs calling OpenAI directly.
  • Know when Phi fits edge/on-prem cheap inference vs when Copilot seats are the real buy.
  • Sketch a Phi chat call (Azure or local) without treating Copilot UI as an API.
  • Navigate next to Module 22.2 Stability AI for image vendors.
Definition

Phi is Microsoft’s family of comparatively small language (and later multimodal/SLM) models designed to punch above parameter count via curated/synthetic-heavy data narratives—many releases are open weights; some appear as Azure catalog SKUs. Copilot is Microsoft’s assistant product line (GitHub Copilot, Microsoft 365 Copilot, Copilot Studio, Windows Copilot, Security Copilot, …) that sits inside existing Microsoft workflows. Under the hood, Copilots often call Azure OpenAI or other Azure models—including OpenAI GPT-class SKUs Microsoft hosts. Microsoft is therefore both a Phi vendor and the primary enterprise on-ramp to OpenAI.

Three Microsoft AI Buys (Do Not Collapse Them)

BuyWhat it isWhat it is not
PhiSLM family: local, ONNX/DirectML, Azure, HFA GPT-4-class flagship you should expect to win unbounded reasoning
Copilot productsSeat-based assistants inside GitHub, M365, Windows, etc.A raw token API; not “Phi branded as Copilot” in general
Azure OpenAI / Azure AIHosted OpenAI + other models with Azure IAM, VNet, loggingOpenAI Consumer ChatGPT; not open GPT weights
Copilot StudioLow-code Copilot builder on Microsoft Graph + connectorsA replacement for Vol. 18 FastAPI when you need a real backend

Pricing / Strengths / Weaknesses

AxisPricing (qualitative)StrengthsWeaknesses
Phi (self-host / device) Open weights: you pay CPU/NPU/GPU, not Microsoft tokens. Fits laptops, edge, and cheap batch. Azure-hosted Phi is usage-metered like other Azure models—read live meters, no fake $. Low latency/cost at small size; on-prem/offline stories; good teaching contrast with frontier APIs; Windows/ONNX path. Capability ceiling below flagship GPT/Claude/Gemini; synthetic-data eval caution (Vol. 11.5 Phi); not a full multimodal estate.
GitHub Copilot Per-seat subscription (Individual / Business / Enterprise tiers)—qualitatively a developer license, not an API token invoice. Confirm current packaging. Fastest coding-assistant distribution in IDEs; Vol. 21 HITL product already exists; enterprise policy controls on Business/Enterprise plans. Seat cost can exceed API-only for tiny teams; model choice is Microsoft’s menu; not a general chatbot platform.
Microsoft 365 Copilot Per-user add-on on top of M365—among the more expensive seat products in the Copilot family (still: no invented dollars). Grounding uses tenant Graph data under Microsoft’s cloud terms. Unmatched if the customer already lives in Outlook/Teams/SharePoint; enterprise identity already solved. Wrong buy if you needed a custom Vol. 21 research assistant on your own RAG; lock-in to M365; privacy review still required (Vol. 20).
Azure OpenAI Token-metered, often close to OpenAI list with Azure commitment/reservation options. Enterprise support and networking are the extra you pay for. Procurement/compliance path for OpenAI models; SSO, private networking, content filters, regional deployment. You still do not own weights; dual-run (openai.com + Azure) complicates data maps; SKU names lag or differ from OpenAI.com.

Pick Phi when

  • Edge, offline, or cheap classification/summarize
  • You want an SLM you can actually run on-device
  • Frontier API cost/privacy is the problem, not max IQ

Pick Copilot seats when

  • Developers already in VS Code/JetBrains + GitHub
  • Knowledge work is already in M365
  • You want a product, not to build Vol. 21 from scratch

Pick Azure OpenAI when

  • The RFP requires Azure + OpenAI models
  • VNet/private endpoints beat openai.com keys
  • You still wrap it behind your API (Vol. 18.2)

Why Microsoft closes 22.1

  • Shows “provider” ≠ only frontier trainers
  • Phi teaches SLM vs LLM procurement
  • Copilot teaches seats vs tokens—the other half of pricing

Do not

  • Call Copilot “the Phi API”
  • Skip HITL because Copilot wrote the diff
  • Assume Azure OpenAI removes Vol. 20 obligations

Practical API Sketch

Phi via a local or Azure-compatible chat endpoint. Copilot itself is a product UI—not this client. Azure OpenAI uses a different base URL and deployment name than openai.com.

# Phi — local OpenAI-compatible runtime (Ollama / ONNX / vLLM) OR Azure deployment. # pip install openai # export PHI_BASE_URL=http://127.0.0.1:11434/v1 # export PHI_API_KEY=ollama # many local servers expect a dummy key from openai import OpenAI import os client = OpenAI(base_url=os.environ["PHI_BASE_URL"], api_key=os.environ.get("PHI_API_KEY", "na")) MODEL = "phi4" # teaching stand-in; pin the exact local or Azure deployment name r = client.chat.completions.create( model=MODEL, messages=[ {"role": "system", "content": "Tiny on-device assistant. Refuse tasks that need frontier reasoning."}, {"role": "user", "content": "When do we deploy Phi locally vs buy M365 Copilot seats vs Azure OpenAI?"}, ], temperature=0.2, max_tokens=300, ) print(r.choices[0].message.content) # Azure OpenAI (GPT-class, not Phi): different azure_endpoint + azure_deployment. # from openai import AzureOpenAI # AzureOpenAI(api_version="...", azure_endpoint="https://YOUR.openai.azure.com") # GitHub Copilot / M365 Copilot are seat products — no "COPILOT_API_KEY" for tenant Graph magic here. # Still HITL on code diffs (Vol. 21). Next module: image vendors, starting with Stability AI.

Related Lectures

LectureRole
Alibaba Cloud (Qwen)Previous open multilingual lab
Phi (11.5)SLM lineage
OpenAIModels Azure OpenAI often hosts
Coding assistants / GitHub Copilot (22.6)Product vs this vendor card
OpenAI SDK / Azure client variantsHow you call Azure OpenAI or local Phi
Stability AI (22.2)Next module: image ecosystem
Common Misconception

Copilot is just Phi in a trench coat. Most Copilot SKUs ride Azure-hosted frontier models (often OpenAI). Second: Microsoft does not offer OpenAI models (Azure OpenAI does). Third: Phi replaces GPT-class reasoning. Fourth: M365 Copilot seats equal Azure token bills. Fifth: this lecture invents seat dollar prices. Sixth: buying Copilot removes HITL, eval, and privacy reviews.

Knowledge Check

  1. Short Answer: Contrast Phi and Copilot in one sentence. Answer: Phi is Microsoft’s SLM family; Copilot is a seat-based assistant product suite (often powered by Azure-hosted frontier models, not Phi alone).
  2. True/False: Azure OpenAI lets you download GPT flagship weights. Answer: False.
  3. Multiple Choice: GitHub Copilot is primarily: (a) a per-seat IDE coding assistant, (b) an open-weight 70B, (c) Veo. Answer: (a).
  4. Short Answer: When is Phi a better buy than a flagship API? Answer: Edge/offline/cheap small tasks where SLM quality is enough (privacy or cost gated).
  5. True/False: Microsoft 365 Copilot billing is the same meter as Azure OpenAI tokens. Answer: False.
  6. Multiple Choice: Microsoft’s role in 22.1 is best described as: (a) SLM lab + Copilot distributor + Azure on-ramp to OpenAI, (b) only a video lab, (c) only Mixtral. Answer: (a).
  7. Short Answer: Which Vol. 11.5 lecture covers Phi? Answer: Phi.
  8. True/False: This lecture asks you to memorize invented Copilot dollar prices. Answer: False.
  9. Multiple Choice: After this lecture, Module 22.2 starts with: (a) Stability AI (SD/SDXL/SVD/ControlNet), (b) Education AI, (c) Gemma only. Answer: (a).
  10. Short Answer: Name one Vol. 22.6 lecture that deepens GitHub Copilot as a coding tool. Answer: GitHub Copilot.

Key Takeaways

  • Microsoft 22.1 = Phi (SLMs) + Copilot (seats/products) + Azure OpenAI (hosted frontier, often OpenAI).
  • Do not collapse those three. Pricing is seats vs tokens vs your GPUs—qualitative only here.
  • Phi for edge/cheap; Copilot when the work already lives in GitHub/M365; Azure OpenAI when procurement demands Azure.
  • HITL, eval, privacy still apply. Copilot is not unsupervised production.
  • Module 22.1 complete. Next: 22.2 Stability AI.
Trainer’s Guide

Lab: Three-column buy decision for one company: (1) Phi on-device FAQ bot, (2) GitHub Copilot for engineering, (3) Azure OpenAI behind FastAPI for a Vol. 21 chatbot. Students must state qualitative cost posture, data path, and a non-goal (no fake prices, no “Copilot = Phi”). Optional local Phi spike.

Whiteboard: Recap all nine 22.1 providers as a catalog wall. Arrow out: image vendors (22.2) starting at Stability. Cap the module: you can now pick a vendor the way Vol. 21 taught you to name a product.

Recap: Microsoft closes frontier providers with Phi SLMs, Copilot seats, and Azure OpenAI distribution. Next module: Stability AI.