www.fgks.org   »   [go: up one dir, main page]
Jump to content

GPT-4: Difference between revisions

From Wikipedia, the free encyclopedia
Browse history interactively
← Previous editNext edit →
Content deleted Content added
VisualWikitext
just keeping the tone.
No edit summary
Line 9: Line 9:
}}
}}


'''GPT-4''' ('''Generative Pre-trained Transformer 4''') is a [[multimodal learning|multimodal]] [[large language model]] created by [[OpenAI]],<ref name="ars-technica">{{Cite web |last=Edwards |first=Benj |date=March 14, 2023 |title=OpenAI's GPT-4 exhibits "human-level performance" on professional benchmarks |url=https://arstechnica.com/information-technology/2023/03/openai-announces-gpt-4-its-next-generation-ai-language-model/ |access-date=March 15, 2023 |website=[[Ars Technica]] |archive-date=March 14, 2023 |archive-url=https://web.archive.org/web/20230314225236/https://arstechnica.com/information-technology/2023/03/openai-announces-gpt-4-its-next-generation-ai-language-model/ |url-status=live }}</ref> the fourth in the [[Generative pre-trained transformer|GPT]] series. It was released on March 14, 2023, and will be available via [[API]] and for [[ChatGPT]] Plus users.<ref name="ars-technica" /> [[Microsoft]] confirmed that versions of [[Microsoft Bing|Bing]] using GPT had in fact been using GPT-4 before its official release. As a [[Transformer (machine learning model)|transformer]], GPT-4 was pretrained to predict the next token (using both public data and "data licensed from third-party providers"), and was then fine-tuned with [[reinforcement learning from human feedback]].<ref name="GPT4Tech" />{{Rp|page=2}}
'''GPT-4''' ('''Generative Pre-trained Transformer 4''') is a [[multimodal learning|multimodal]] [[large language model]] created by [[OpenAI]],<ref name="ars-technica">{{Cite web |last=Edwards |first=Benj |date=March 14, 2023 |title=OpenAI's GPT-4 exhibits "human-level performance" on professional benchmarks |url=https://arstechnica.com/information-technology/2023/03/openai-announces-gpt-4-its-next-generation-ai-language-model/ |access-date=March 15, 2023 |website=[[Ars Technica]] |archive-date=March 14, 2023 |archive-url=https://web.archive.org/web/20230314225236/https://arstechnica.com/information-technology/2023/03/openai-announces-gpt-4-its-next-generation-ai-language-model/ |url-status=live }}</ref> the fourth in the [[Generative pre-trained transformer|GPT]] series. It was released on March 14, 2023, and will be available via [[API]] and for [[ChatGPT]] Plus users.<ref name="ars-technica" /> [[Microsoft]] confirmed that versions of [[Microsoft Bing|Bing]] using GPT had in fact been using GPT-4 before its official release. As a [[Transformer (machine learning model)|transformer]], GPT-4 was pretrained to predict the next token (using both public data and "data licensed from third-party providers"), and was then fine-tuned with [[reinforcement learning from human feedback|reinforcement learning from human and AI software]] for human alignment and policy compliance.<ref name="GPT4Tech" />{{Rp|page=2}}


== Training and capabilities ==
== Training and capabilities ==
OpenAI wrote in their blog post announcing GPT-4 that "GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5."<ref>{{Cite web |last=Wiggers |first=Kyle |date=March 14, 2023 |title=OpenAI releases GPT-4, a multimodal AI that it claims is state-of-the-art |url=https://techcrunch.com/2023/03/14/openai-releases-gpt-4-ai-that-it-claims-is-state-of-the-art/ |access-date=March 15, 2023 |website=TechCrunch |archive-date=March 15, 2023 |archive-url=https://web.archive.org/web/20230315003723/https://techcrunch.com/2023/03/14/openai-releases-gpt-4-ai-that-it-claims-is-state-of-the-art/ |url-status=live }}</ref> The organization produced two versions of GPT-4 with context windows of 8192 and 32768 tokens, a significant improvement over GPT-3.5 and GPT-3, which were limited to 4096 and 2048 tokens respectively.<ref>{{Cite web |title=OpenAI API |url=https://platform.openai.com/ |access-date=2023-03-16 |website=platform.openai.com |language=en}}</ref> Unlike its predecessor, GPT-4 can take images as well as text as inputs.<ref>{{cite web |author1=Alex Hern |author2=Johana Bhuiyan |title=OpenAI says new model GPT-4 is more creative and less likely to invent facts |url=https://www.theguardian.com/technology/2023/mar/14/chat-gpt-4-new-model |website=The Guardian |access-date=15 March 2023 |date=14 March 2023}}</ref>
OpenAI wrote in their blog post announcing GPT-4 that "GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5."<ref>{{Cite web |last=Wiggers |first=Kyle |date=March 14, 2023 |title=OpenAI releases GPT-4, a multimodal AI that it claims is state-of-the-art |url=https://techcrunch.com/2023/03/14/openai-releases-gpt-4-ai-that-it-claims-is-state-of-the-art/ |access-date=March 15, 2023 |website=TechCrunch |archive-date=March 15, 2023 |archive-url=https://web.archive.org/web/20230315003723/https://techcrunch.com/2023/03/14/openai-releases-gpt-4-ai-that-it-claims-is-state-of-the-art/ |url-status=live }}</ref> The organization produced two versions of GPT-4 with context windows of 8192 and 32768 tokens, a significant improvement over GPT-3.5 and GPT-3, which were limited to 4096 and 2048 tokens respectively.<ref>{{Cite web |title=OpenAI API |url=https://platform.openai.com/ |access-date=2023-03-16 |website=platform.openai.com |language=en}}</ref> Unlike its predecessor, GPT-4 can take images as well as text as inputs.<ref>{{cite web |author1=Alex Hern |author2=Johana Bhuiyan |title=OpenAI says new model GPT-4 is more creative and less likely to invent facts |url=https://www.theguardian.com/technology/2023/mar/14/chat-gpt-4-new-model |website=The Guardian |access-date=15 March 2023 |date=14 March 2023}}</ref>


OpenAI adopted a closed approach with regards to the technical details of GPT-4; the technical report explicitly refrained from specifying the model size, architecture, hardware, or training method. The report claimed that "the competitive landscape and the safety implications of large-scale models" were factors that influenced this decision.<ref name="GPT4Tech" /> The precise parameter count of GPT-4 remains unknown, but ''[[The Verge]]'' cited rumors that GPT-4 would substantially increase the parameter count from [[GPT-3]]'s 175 billion to 100 trillion. OpenAI CEO [[Sam Altman]] described these rumors as "complete bullshit".<ref name="verge" />
OpenAI adopted a closed approach with regards to the technical details of GPT-4; the technical report explicitly refrained from specifying the model size, architecture, hardware, or training method, apart from the fact it uses reinforcement learning from human and AI feedback. The report claimed that "the competitive landscape and the safety implications of large-scale models" were factors that influenced this decision.<ref name="GPT4Tech" /> The precise parameter count of GPT-4 remains unknown, but ''[[The Verge]]'' cited rumors that GPT-4 would substantially increase the parameter count from [[GPT-3]]'s 175 billion to 100 trillion. OpenAI CEO [[Sam Altman]] described these rumors as "complete bullshit".<ref name="verge" />


U.S. Representatives [[Don Beyer]] and [[Ted Lieu]] confirmed to the [[New York Times]] that Altman visited [[United States Congress|Congress]] in January 2023 to demonstrate GPT-4 and its improved "security controls" compared to other AI models.<ref name="nyt-3" />
U.S. Representatives [[Don Beyer]] and [[Ted Lieu]] confirmed to the [[New York Times]] that Altman visited [[United States Congress|Congress]] in January 2023 to demonstrate GPT-4 and its improved "security controls" compared to other AI models.<ref name="nyt-3" />

Revision as of 05:38, 17 March 2023

Generative Pre-trained Transformer 4 (GPT-4)
Original author(s)OpenAI
Initial releaseMarch 14, 2023
TypeAutoregressive multimodal transformer language model
Websiteopenai.com/gpt-4 Edit this on Wikidata

GPT-4 (Generative Pre-trained Transformer 4) is a multimodal large language model created by OpenAI,[1] the fourth in the GPT series. It was released on March 14, 2023, and will be available via API and for ChatGPT Plus users.[1] Microsoft confirmed that versions of Bing using GPT had in fact been using GPT-4 before its official release. As a transformer, GPT-4 was pretrained to predict the next token (using both public data and "data licensed from third-party providers"), and was then fine-tuned with reinforcement learning from human and AI software for human alignment and policy compliance.[2]: 2 

Training and capabilities

OpenAI wrote in their blog post announcing GPT-4 that "GPT-4 is more reliable, creative, and able to handle much more nuanced instructions than GPT-3.5."[3] The organization produced two versions of GPT-4 with context windows of 8192 and 32768 tokens, a significant improvement over GPT-3.5 and GPT-3, which were limited to 4096 and 2048 tokens respectively.[4] Unlike its predecessor, GPT-4 can take images as well as text as inputs.[5]

OpenAI adopted a closed approach with regards to the technical details of GPT-4; the technical report explicitly refrained from specifying the model size, architecture, hardware, or training method, apart from the fact it uses reinforcement learning from human and AI feedback. The report claimed that "the competitive landscape and the safety implications of large-scale models" were factors that influenced this decision.[2] The precise parameter count of GPT-4 remains unknown, but The Verge cited rumors that GPT-4 would substantially increase the parameter count from GPT-3's 175 billion to 100 trillion. OpenAI CEO Sam Altman described these rumors as "complete bullshit".[6]

U.S. Representatives Don Beyer and Ted Lieu confirmed to the New York Times that Altman visited Congress in January 2023 to demonstrate GPT-4 and its improved "security controls" compared to other AI models.[7]

Reception

The New York Times wrote that GPT-4 showed large improvements in accuracy compared to GPT-3.5, had gained the ability to summarize and comment on images, was able to summarize complicated texts, passed a bar exam and several standardized tests, but still showed a tendency to hallucinate answers.[8]

Usage

GPT-4 is available via API and for ChatGPT Plus users.[1]

Microsoft confirmed that versions of Bing using GPT had in fact been using GPT-4 before its official release.[9]

Duolingo added GPT-4 to its application, and added two new features, "Roleplay" and "Explain My Answer". It currently works only for English- speakers who learn French or Spanish.[10]

Icelandic start-up Miðeind ehf, that works on language preservation, was selected by OpenAI as one of 6 companies to participate in early beta test program of the new model.[11]

References

  1. ^ a b c Edwards, Benj (March 14, 2023). "OpenAI's GPT-4 exhibits "human-level performance" on professional benchmarks". Ars Technica. Archived from the original on March 14, 2023. Retrieved March 15, 2023.
  2. ^ a b "GPT-4 Technical Report" (PDF). OpenAI. 2023. Archived (PDF) from the original on March 14, 2023. Retrieved March 14, 2023.
  3. ^ Wiggers, Kyle (March 14, 2023). "OpenAI releases GPT-4, a multimodal AI that it claims is state-of-the-art". TechCrunch. Archived from the original on March 15, 2023. Retrieved March 15, 2023.
  4. ^ "OpenAI API". platform.openai.com. Retrieved 2023-03-16.
  5. ^ Alex Hern; Johana Bhuiyan (14 March 2023). "OpenAI says new model GPT-4 is more creative and less likely to invent facts". The Guardian. Retrieved 15 March 2023.
  6. ^ Vincent, James (January 18, 2023). "OpenAI CEO Sam Altman on GPT-4: "people are begging to be disappointed and they will be"". The Verge. Archived from the original on January 26, 2023. Retrieved January 27, 2023.
  7. ^ Kang, Cecilia (March 3, 2023). "As A.I. Booms, Lawmakers Struggle to Understand the Technology". The New York Times. Archived from the original on March 3, 2023. Retrieved March 3, 2023.
  8. ^ Metz, Cade; Collins, Keith (March 14, 2023). "10 Ways GPT-4 Is Impressive but Still Flawed". The New York Times. Archived from the original on March 14, 2023. Retrieved March 14, 2023.
  9. ^ Lardinois, Frederic (March 14, 2023). "Microsoft's new Bing was using GPT-4 all along". TechCrunch. Archived from the original on March 15, 2023. Retrieved March 14, 2023.
  10. ^ "Introducing Duolingo Max, a learning experience powered by GPT-4". Duolingo Blog. March 14, 2023.
  11. ^ "'Basically mindblowing' — What GPT-4 can do, according to one startup that's had access to it". Sifted. March 15, 2023.
Retrieved from "https://en.wikipedia.org/w/index.php?title=GPT-4&oldid=1145085655"