The same LM can be a much more or less capable agent depending on the enhancements added. The researchers created and tested four different agents built on top of GPT-4 and Anthropic’s Claude:

The latest GPT-4 model from OpenAI, which is trained on human preferences using a technique called RLHFEstimated final training run compute cost: ~$50mModel version: gpt-4-0613

OpenAI is offering limited access to a text-to-voice generation platform it developed called Voice Engine, which can create a synthetic voice based on a 15-second clip of someone’s voice.

GPT-4 System CardOpenAIMarch 23, 2023

Assistant messages store previous assistant responses, but can also be written by a developer to give examples of desired behavior.

The user messages provide requests or comments for the assistant to respond to

"You are a helpful assistant."

The system message helps set the behavior of the assistant.

In June 2021, OpenAI published a paper offering a new technique for battling toxicity in GPT-3’s responses, calling it PALMS, short for ‘‘process for adapting language models to society.’’ PALMS involves an extra layer of human intervention, defining a set of general topics that might be vulnerable to GPT-3’s being led astray by the raw training data: questions about sexual abuse, for instance, or Nazism.

‘‘I think it lets us be more thoughtful and more deliberate about safety issues,’’ Altman says. ‘‘Part of our strategy is: Gradual change in the world is better than sudden change.’’

The supercomputer complex in Iowa is running a program created by OpenAI, an organization established in late 2015 by a handful of Silicon Valley luminaries, including Elon Musk; Greg Brockman, who until recently had been chief technology officer of the e-payment juggernaut Stripe; and Sam Altman, at the time the president of the start-up incubator Y Combinator.

Still, we can look for telltale signs. Another symptom of memorization is that GPT is highly sensitive to the phrasing of the question. Melanie Mitchell gives an example of an MBA test question where changing some details in a way that wouldn’t fool a person is enough to fool ChatGPT (running GPT-3.5). A more elaborate experiment along these lines would be valuable.

In fact, we can definitively show that it has memorized problems in its training set: when prompted with the title of a Codeforces problem, GPT-4 includes a link to the exact contest where the problem appears (and the round number is almost correct: it is off by one). Note that GPT-4 cannot access the Internet, so memorization is the only explanation.

To benchmark GPT-4’s coding ability, OpenAI evaluated it on problems from Codeforces, a website that hosts coding competitions. Surprisingly, Horace He pointed out that GPT-4 solved 10/10 pre-2021 problems and 0/10 recent problems in the easy category. The training data cutoff for GPT-4 is September 2021. This strongly suggests that the model is able to memorize solutions from its training set — or at least partly memorize them, enough that it can fill in what it can’t recall.

Whisper is a general-purpose speech recognition model. It is trained on a large dataset of diverse audio and is also a multi-task model that can perform multilingual speech recognition as well as speech translation and language identification.

Image generation BetaLearn how to generate or manipulate images with our DALL·E models

While OpenAI has released its algorithms to the public in the past, it has opted to keep GPT-3 locked away.

OpenAI is perhaps one of the oddest companies to emerge from Silicon Valley. It was set up as a non-profit in 2015 to promote and develop "friendly" AI in a way that "benefits humanity as a whole". Elon Musk, Peter Thiel and other leading tech figures pledged US$1 billion towards its goals.Their thinking was we couldn't trust for-profit companies to develop increasingly capable AI that aligned with humanity's prosperity. AI therefore needed to be developed by a non-profit and, as the name suggested, in an open way.In 2019 OpenAI transitioned into a capped for-profit company (with investors limited to a maximum return of 100 times their investment) and took a US$1 billion investment from Microsoft so it could scale and compete with the tech giants.

This phenomenon is characteristic of modern ML models, where an active community creates many new versions based on an original ML model that may enable greater use for different user groups. Each version may have its own license, though some model developers are now requiring all downstream models (derived models from the original model) to at least have the same use restrictions as included in the original license.

He said the new AI tutor platform collects “competency skills graphs” made by educators, then uses AI to generate learning activities, such as short-answer or multiple-choice questions, which students can access on an app. The platform also includes applications that can chat with students, provide coaching for reading comprehension and writing, and advise them on academic course plans based on their prior knowledge, career goals and interest

OpenAI and other researchers have released a few tools capable of identifying AI-generated text. These use similar AI algorithms to spot telltale signs in the text. It’s not clear if anyone is using these to protect online commenting platforms. Facebook declined to say if it is using such tools; Google and Twitter did not respond to requests for comment.

OpenAI released a more capable version of its text-generation program, called GPT-3, last June. So far, it has only been made available to a few AI researchers and companies, with some people building useful applications such as programs that generate email messages from bullet points. When GPT-3 was released, OpenAI said in a research paper that it had not seen signs of GPT-2 being used maliciously, even though it had been aware of Weiss’s research.