GNSS & Machine Learning Engineer

Tag: NLP (Page 1 of 3)

OpenAI releases ChatGPT plugins

OpenAI announced on Mar 23, 2023, the availability of plugins within ChatGPT. Access is currently limited to ChatGPT Plus subscribers that joined a waitlist and have been selected by OpenAI.

Plugins can be automatically called by ChatGPT’s underlying LLM (Large Language Model, currently GPT-3.5 or GPT-4) in order to answer the questions of the user.

In order to make this work, plugins have to be registered in the ChatGPT user interface with a manifest file
(ai-plugin.json) that is hosted in the developer’s domain at
yourdomain.com/.well-known/ai-plugin.json. The file contains in a prescribed format
– metadata about the plugin (name, logo)
– details about the authentication mechanism
– an OpenAI specification for the endpoints of the API
– and a general description for the LLM of what the plugin can do.
The web API needs to define an endpoint “/.well-known/ai-plugin.json” to access the content of this file.

In addition to the manifest file, an openapi.yaml file, that defines the OpenAI specification, has to be generated that is referenced in the “api” section of the manifest file via the “url” field. This file contains a detailed description of the API endpoints. The web API needs to define an endpoint “/openapi.yaml” to access the content of this file.

When the user has activated a registered plugin and starts a conversation, the plugin’s description is injected into the message to ChatGPT, but invisible to the user. In this way, the LLM may choose an API call from the plugin if this seems relevant to the user’s question. The LLM will then incorporate the API result into the response to the user. More details can be found in OpenAI’s documentation.

Among the already available plugins, a few stand out. With the Wolfram plugin, all kinds of computational problems can be solved. And with the Zapier plugin, more than 5000 apps can be accessed. OpenAI itself introduced a web browser (that uses the Bing search API) and a code interpreter plugin (that runs in a sandbox without an internet connection). In addition, they open-sourced the code for a knowledge base retrieval plugin, that has to be self-hosted by a developer.

Interestingly enough, OpenAI notices that plugins will likely have wide-ranging societal implications and that language models with access to tools will likely have much greater economic impacts than those without. They expect the current wave of AI technologies to have a big effect on the pace of job transformation, displacement, and creation. OpenAI discusses the impact potential of large language models at the labor market in a recent publication.

Just a day after the OpenAI announcement of ChatGPT plugins, the open-source community already integrated these plugins also into LangChain. This is done just by referring to the plugin manifest file ai-plugin.json (see Twitter), e.g.:

tool = AIPluginTool.from_plugin_url( "https://www.clarna.com/.wellknown/ai-plugin.json")

All the other exciting news of the week is well summarized by Matt Wolfe (Google Bard, NVIDIA GTC, Adobe Firefly, Image Generation in Bing via DALL-E2, Microsoft Loop, AI in Canva, GitHub Copilot X, AI in Ubisoft, Metahuman by Unreal Engine).

Google announces PaLM API release

On the same day as OpenAI released GPT-4 (March 14, 2023), Google also announced the availability of the PaLM API for developers on Google Cloud [video]. They said that they are now providing access to foundation models on Google Cloud’s Vertex AI platform, initially for generating text and images, and over time also for audio and video. In addition, with the Generative AI App Builder, they introduced the possibility of quickly building AI-powered chat interfaces and digital assistants.

Finally, Google also made for a limited set of trusted test users generative AI features available within Google Workspace (Gmail and Google Docs).

OpenAI releases GPT-4

OpenAI released GPT-4 within ChatGPT on March 14, 2023, described in detail in a 98-pages paper (summarized on youtube).

  • Available to ChatGPT-Plus subscribers (currently with a cap that is changing over time, e.g. 100 messages every 4 hours, or 25 messages every 3 hours).
  • Still based on training data that cuts off Sept 2021.
  • It still does not learn from its experience.
  • Still no internet access.
  • The training was already finalized in Aug 2022.
  • Fine-tuned via RLHF (Reinforcement Learning with Human Feedback).
  • API waitlist is open (so no API access yet for everyone)
  • API prices (for comparison: GPT-3.5-turbo $0.002 per 1k tokens):
    • gpt-4: 8K context window (about 13 pages of text) will cost $0.03 per 1K prompt tokens and $0.06 per 1K completion tokens.
    • gpt-4-32k: 32K context window (about 52 pages of text) will cost $0.06 per 1K prompt tokens and $0.12 per 1K completion tokens.
  • The number of parameters and size of the training data set have both not been published. So competitors are not encouraged to replicate these performance ingredients but are referred to a freely available benchmark (OpenAI Evals) that measures the real performance.
  • GPT-4 ranks in the 10% best of the bar exam and 0.5% best of biology olympiad.
  • GPT-4 can handle contexts of over 25,000 words.
  • GPT-4 can access images as inputs and can generate captions, classifications, and analyses. However, this image-to-text functionality is not yet publicly available.
  • Microsoft Bing was already using an early version of GPT-4 in the last few weeks.

An excellent overview by Greg Brockman, President and co-founder of OpenAI, can be found on youtube.

Microsoft released Visual ChatGPT on March 08, 2023, in a paper and with source code on GitHub and Hugging Face. Although this does not seem to be GPT-4-based, it demonstrates similar image capabilities via a combination of pre-existing technologies (generate/modify [text-to-image], and describe [image-to-text]).

Two days after the GPT-4 release, Microsoft announced on March 16, 2023, the integration of GPT-4 into their Office products as a feature they called CoPilot. Copilot is not yet available for general use, but Microsoft plans to roll it out gradually to selected customers in the coming months.

OpenAI releases ChatGPT and Whisper APIs

On March 01, 2023, OpenAI announced the releases of APIs for ChatGPT (published on Nov 30, 2022) and the automatic speech recognition (ASR) engine Whisper for speech-to-text (STT) transcription (and translation) that was open-sourced in Sept 2022.

The ChatGPT model family is called gpt-3.5-turbo and costs just $0.002 per 1k tokens, which is 10 times cheaper than the existing GPT-3.5 models. Instead of consuming unstructured text as traditionally done by GPT, the ChatGPT models consume a sequence of messages with metadata following a new format called Chat Markup Language (ChatML). The number of tokens (tokens in prompt + tokens in response as available via response[‘usage’][‘total_tokens’]) is restricted to 4096. Notice that there is no possibility to fine-tune gpt-3.5-turbo models.

For Whisper the large-v2 model is now available through an API for a price of $0.006 per minute. The API contains endpoints for transcriptions (transcribes in source language) and translations (transcribes into English).

In addition, the possibility of dedicated instances for professional users was announced that can make economical sense beyond ~450M tokens per day.

A significant change that was made in the Terms of Service and Usage Polices is that data submitted to the API is no longer used for service improvements (e.g. model training) unless an organization opts in. Before it was necessary to opt-out.

Microsoft’s VALL-E can synthesize your voice from 3 sec of audio

Microsoft has introduced a new language modeling approach for text-to-speech synthesis (TTS) called VALL-E. The approach uses discrete codes derived from an off-the-shelf neural audio codec model, and is trained using 60K hours of English speech, which is hundreds of times larger than existing systems, and can be used to synthesize high-quality personalized speech with only a 3-second enrolled recording of an unseen speaker as an acoustic prompt (project page, paper).

An unofficial Pytorch implementation for VALL-E is available on GitHub.

Google’s Med-PaLM comes close to human performance in clinical knowledge

In a recent paper from Dec 26, 2022, Google demonstrates that its large language model Med-PaLM, based on 540 billion parameters with a special instruction prompt tuning for the medical domain, reaches almost clinician’s performance on new medical benchmarks MultiMedQA (benchmark combining six existing open question answering datasets spanning professional medical exams, research, and consumer queries) and HealthSearchQA (a new free-response dataset of medical questions searched online). The evaluation of the answers considering factuality, precision, possible harm, and bias was done by human experts.

GPT-3.5 passes parts of the US legal Bar Exam

In the United States, most jurisdictions require applicants to pass the Bar Exam in order to practice law. This exam typically requires several years of education and preparation (seven years of post-secondary education, including three years at an accredited law school).

In a publication from Dec 29, 2022, the authors evaluated the performance of GPT-3.5 on the multiple choice part of the exam. While GPT is not yet passing that part of the exam, it significantly exceeded the baseline random chance rate
of 25% and reached the average human passing rate for the categories Evidence and Torts.
On average, GPT is performing about 17% worse than human test-takers across all categories.

Similar to this publication is the report that ChatGPT was able to pass the Wharton Master of Business Applications (MBA) exam.

« Older posts

© 2023 Stephan Seeger

Theme by Anders NorenUp ↑