After GPT-3.5 made chat assistants mainstream, GPT-4 (OpenAI, 2023) raised the capability bar: stronger reasoning/exams performance, longer contexts in product variants, and public multimodal (vision) interfaces. Full internals remain closed; this lecture teaches what is documented enough to engineer against—and what to leave as unknown.
GPT-5 (August 2025) continues the family as a routed multi-tier generation with stronger reasoning and tool patterns—still closed-weight, still requiring grounding and evals in production.
Learning Objectives
By the end of this lesson, students should be able to:
- Place GPT-4 on the GPT timeline after GPT-3.5.
- Separate verified product capabilities from rumored architecture details.
- Describe multimodal (text+image) usage patterns at the API/product level.
- Relate system cards / evals to responsible deployment thinking.
- Contrast GPT-4-class APIs with open weights for build-vs-buy decisions.
- Connect still-present limits: hallucination, context cost, tool needs.
GPT-4 is OpenAI’s 2023 flagship GPT-family model generation, reported as a large multimodal system with substantially improved performance on academic and professional benchmarks versus GPT-3.5, accessed primarily via API and ChatGPT—with limited public disclosure of exact parameter counts and training details.
Timeline
Few-shot scale.
Chat alignment.
Stronger + multimodal.
Routed family (2025).
What We Can Say Confidently
| Topic | Documented / observed | Often unknown publicly |
|---|---|---|
| Interface | Chat/completions APIs; vision inputs in products | Exact vision encoder design |
| Quality | Large gains on many exams/benchmarks vs 3.5 | Full training corpus composition |
| Safety | System card; mitigations discussed | All internal red-team data |
| Scale | “Large” multimodal LM | Official single param count |
Text reasoning
- Harder exams / coding help
- Better instruction following
- Still can be wrong
Vision
- Images as inputs in products
- Diagrams, screenshots, photos
- OCR + reasoning patterns
Systems use
- Tools / browsing in products
- Longer context variants
- Cost/latency tiers
Engineering Against a Closed Model
“If it sounds confident, the architecture details on social media must be true.” Prefer primary sources (technical reports, system cards, API docs). Unverified rumors about layer counts or expert mixtures are not curriculum facts.
Strengths and Tradeoffs
Strengths
- State-of-the-art practical reasoning for many tasks.
- Multimodal product surfaces.
- Rich tooling ecosystem.
Tradeoffs
- Closed weights and limited reproducibility.
- Cost, rate limits, vendor lock-in.
- Hallucinations and policy constraints remain.
Knowledge Check
- Short Answer: In what year was GPT-4 announced? Answer: 2023.
- True/False: OpenAI published a single official public parameter count for GPT-4 like GPT-3’s 175B. Answer: False—exact count was not publicly confirmed like GPT-3.
- Multiple Choice: A documented GPT-4 theme is: (a) multimodal inputs in products, (b) being a CNN-only ImageNet model, (c) replacing electricity. Answer: (a).
- Short Answer: Name one artifact for safety/eval discussion. Answer: The GPT-4 system card (or similar reports).
- True/False: GPT-4 eliminated hallucinations. Answer: False.
- Multiple Choice: For curriculum honesty we should: (a) invent layer counts, (b) distinguish known product facts from rumors, (c) ignore APIs. Answer: (b).
- Short Answer: Why might a team still use DistilBERT? Answer: Cost, latency, privacy, or sufficient NLU accuracy.
- Short Answer: What precedes GPT-4 in this module? Answer: GPT-3.5.
- Multiple Choice: GPT-4 vs GPT-3.5 generally: (a) stronger on hard benchmarks, (b) identical always, (c) cannot chat. Answer: (a).
- True/False: Vendor APIs remove the need for evaluation harnesses. Answer: False.
Key Takeaways
- GPT-4 is the 2023 multimodal capability jump in the GPT line.
- Engineer from APIs, evals, and system cards—not rumors.
- Vision + stronger reasoning; limits remain.
- Closed models trade control for capability.
- Next: GPT-5 as continuing evolution themes.
Hands-on idea: Design a mini eval set (10 questions) and score any available strong model vs. a small open model.
Discussion prompt: When is “unknown architecture” acceptable in a regulated industry?
Recap: GPT-4 raised multimodal and reasoning capability under closed details. Continue with GPT-5.