Peluncuran GPT-4

OpenAI merilis GPT-4 — LLM multimodal (input teks & gambar) yang kinerjanya mendekati atau melebihi manusia pada ujian standar (BAR, SAT, GRE, AP, USMLE).

GPT-4: multimodal, ~1.8T parameters (rumor MoE 8x220B). State of the art 2023 H1. API dengan vision, JSON mode. Microsoft Bing 365 Copilot.

Print

Peluncuran GPT-4

Ringkasan

OpenAI merilis GPT-4 pada 14 Maret 2023 — evolusi besar dari GPT-3.5 dengan kemampuan multimodal (input teks DAN gambar).

Pencapaian

  • Multimodal: menerima input teks + gambar
  • Context window: 8.192 token (standar) atau 32.768 token (extended)
  • Ujian standar:
    • BAR (ujian律师): 90% percentile (vs GPT-3.5: 10%)
    • SAT Evidence-Based Reading: 93% (vs GPT-3.5: 85%)
    • GRE Quantitative: 80% (vs GPT-3.5: 25%)
    • AP Biology: 85-100% percentile
    • USMLE: mendekati level praktisi medis
  • Coding: HumanEval pass@1 67% (vs GPT-3.5: 48%)

Arsitektur

  • MoE (rumor): 8 expert x 220B parameters = 1,8T total
  • Training data: sampai September 2021
  • Training compute: $100M+ (rumor)
  • Inference cost: 3x GPT-3.5

Aplikasi Pasca GPT-4

  • Microsoft Bing Chat (Februari 2023) — basis GPT-4
  • Microsoft 365 Copilot (Maret 2023) — Office + AI
  • GitHub Copilot X (Maret 2023) — coding AI
  • Duolingo Max (Maret 2023) — bahasa AI
  • Khan Academy Khanmigo — tutor AI
  • Be My Eyes (Visual AI) — aksesibilitas
  • Stripe — customer support
  • Morgan Stanley — wealth management research

Dampak

GPT-4 menetapkan ulang ekspektasi publik. Beberapa perusahaan memecat pekerja customer service karena “AI sudah bisa”. Industri AI menjadi infrastruktur.

Connected to

Not yet written

The following pages are referenced but don't exist yet — they'd make good future additions.

  • /concepts/large-language-model
  • /sources/openai

References

  1. Wikipedia

Type at least 2 characters to search.

Press to navigate, to open, esc to close.