GPT-4 Is Here: A Multimodal Large Model, Direct Upgrade to ChatGPT and Bing, API Now Open — Game Over? | BlueRun Ventures --- This article is a repost from the WeChat public account: BlueRun Ventures (ID: brvchina). Early this morning, OpenAI released GPT-4. The official announcement states that GPT-4 is a large multimodal model capable of accepting both image and text inputs and generating text outputs. While it may be "less capable than humans in many real-world scenarios," it demonstrates "human-level performance on various professional and academic benchmarks." The specific upgrades include: - GPT-4 is more reliable, creative, and better able to handle nuanced instructions compared to GPT-3.5. - It supports longer contexts, handling over 25,000 words of text. - It accepts image inputs and generates captions, classifications, and analyses. - It outperforms GPT-3.5 on a range of benchmark tests, including simulated exams designed for humans. Notably, OpenAI also announced that GPT-4 is already powering Microsoft's Bing search engine. This means the Bing that has been making waves recently has been running on GPT-4 all along. Additionally, OpenAI is opening GPT-4's API to developers, with a waitlist already available. So, is this the "game over" moment many have

蓝驰创投·March 15, 2023·0·0

The Singularity Returns

GPT-4 has finally emerged from its cocoon.

On March 14 local time, OpenAI publicly released GPT-4, a large multimodal model. In an introductory video, an engineer defined the model as "the world's first high-experience, high-capability advanced AI system."

We've all known humanity is on the path to AGI, but what's stunning about GPT-4 is how clearly it reveals just how fast that forward march can be.

Following GPT-4's release, OpenAI upgraded ChatGPT. You can compare your own experience against this article to feel how far GPT-4 advances beyond its predecessor. Enjoy—

Who could possibly dethrone ChatGPT? Turns out, OpenAI itself. After ChatGPT detonated across the tech world, people have been debating what the "next step" for AI would look like. Many researchers pointed to multimodality. We didn't have to wait long. Early this morning, OpenAI released GPT-4, a multimodal pretrained large model.

GPT-4 delivers leaps forward across several dimensions: powerful image comprehension; text input limits expanded to 25,000 words; dramatically improved answer accuracy; and the ability to generate lyrics, creative writing, and stylistic variations.

"GPT-4 is the world's first high-experience, high-capability advanced AI system, and we hope to get it into everyone's hands soon," an OpenAI engineer said in the introductory video.

Seemingly intent on ending the game in one move, OpenAI released a paper (more of a technical report), a System Card, upgraded ChatGPT directly to a GPT-4-powered version, and opened up GPT-4's API.

Additionally, a Microsoft marketing executive stated immediately after GPT-4's launch: "If you've used the new Bing preview at any point in the last six weeks, you've already had an early look at the power of OpenAI's latest model." Yes, Microsoft's new Bing has been running on GPT-4 all along.

Now, let's savor this stunning release in detail.

01 GPT-4: I Scored 710 on the SAT, and I Can Be a Lawyer Too

GPT-4 is a large multimodal model that accepts both image and text inputs and outputs correct text responses. Experiments show that GPT-4 performs at human levels on various professional tests and academic benchmarks. For instance, it passed a simulated bar exam with a score around the top 10% of test-takers; by comparison, GPT-3.5 scored around the bottom 10%.

OpenAI spent six months iteratively tuning GPT-4 using adversarial testing programs and lessons learned from ChatGPT, achieving its best-ever results on truthfulness, controllability, and other dimensions. Over the past two years, OpenAI rebuilt its entire deep learning stack and, together with Azure, designed a supercomputer from scratch for its workloads. A year ago, OpenAI first ran this supercomputing system while training GPT-3.5, then went on to discover and fix bugs and refine its theoretical foundations. The result of these improvements was unprecedented stability in GPT-4's training run, to the point that OpenAI could accurately predict GPT-4's training performance in advance — a first for large models. OpenAI says it will continue focusing on reliable scaling, further refining its methods to enable stronger advance prediction and future planning capabilities, which are critical for safety. OpenAI is releasing GPT-4's text input capabilities through ChatGPT and API (waitlist). For image input, to achieve broader availability, OpenAI is partnering with other companies. OpenAI also open-sourced OpenAI Evals today, its framework for automatically evaluating AI model performance. The company says this move is intended to let everyone point out its models' flaws to help OpenAI improve them further.

Interestingly, the difference between GPT-3.5 and GPT-4 can be subtle. The divergence appears when task complexity crosses a sufficient threshold — GPT-4 is more reliable, more creative, and better able to handle nuanced instructions than GPT-3.5. To understand the gap between these two models, OpenAI conducted experiments across various benchmarks and human-designed simulated exams.

OpenAI also evaluated GPT-4 on traditional benchmarks designed for machine learning models. GPT-4 substantially outperforms existing large language models and most SOTA models:

Many existing machine learning benchmarks are written in English. To get a preliminary sense of GPT-4's capabilities in other languages, the research team used Azure Translate to render the MMLU benchmark — a suite of 14,000 multiple-choice questions across 57 subjects — into multiple languages. Of the 26 languages tested, GPT-4 outperformed GPT-3.5 and other large language models' (Chinchilla, PaLM) English-language performance in 24:

Like many companies using ChatGPT, OpenAI says it is also using GPT-4 internally, and is therefore tracking how large language models perform in content generation, sales, programming, and other applications. OpenAI is also using GPT-4 to assist people in evaluating AI outputs — the second phase of its strategy. OpenAI is both GPT-4's developer and its user.

02 GPT-4: I Can Meme Too

GPT-4 can accept prompts in both text and image form, with the new capability running in parallel to its text-only mode, allowing users to specify any visual or language task.

Specifically, it generates corresponding text outputs (natural language, code, etc.) given inputs composed of interspersed text and images. Across a range of domains — including documents with text and photographs, charts, or screenshots — GPT-4 demonstrates similar capabilities to text-only inputs. Moreover, it can be enhanced by test-time techniques developed for text-only language models, including few-shot and chain-of-thought prompting.

For example, feed GPT-4 an image of a bizarre-looking charger and ask why it's funny:

GPT-4 responds: VGA cable charging an iPhone.

Average daily meat consumption per capita in Georgia and Western Asia — calculate the mean:

Looks like GPT no longer babbles nonsense when it comes to calculations:

Still too easy? Then let's have it solve a problem — a physics problem, no less:

GPT-4 understood the French problem and solved it completely:

GPT-4 can understand "what's wrong with this picture" in a photo:

GPT-4 can also speed-read papers quantum-style. If you feed it the InstructGPT paper and ask for a summary, this is what happens:

Interested in a particular figure from the paper? GPT-4 can explain that too:

Next up, asking GPT-4 what a meme means:

It gave a detailed response:

What about comics?

Having GPT-4 explain why you should add layers to a neural network seems to have an extra layer of humor to it.

That said, OpenAI notes here that image input is still a research preview and not yet publicly available.

Researchers have been evaluating GPT-4's visual capabilities through academic benchmarks, but that's no longer sufficient — they keep discovering exciting new tasks the model can handle. The real tension now is between AI capability and human imagination.

At this point, some researcher is surely sighing: computer vision is dead.

Controllability

Unlike classic ChatGPT's fixed personality — verbose, calm, and uniform in tone — developers (and ChatGPT users) can now prescribe their AI's style and task through "system" messages that describe these directions.

System messages allow API users to customize user experiences within certain bounds. OpenAI knows you've been making ChatGPT cosplay, and they're encouraging it.

Limitations

Despite its impressive capabilities, GPT-4 still shares similar limitations with earlier GPT models. Most critically, it remains less than fully reliable. OpenAI notes that GPT-4 still hallucinates, generates incorrect answers, and makes reasoning errors.

For now, language model outputs should be reviewed carefully, with exact protocols matched to specific use-case needs (such as human review, additional context, or avoiding use entirely).

Overall, GPT-4 has significantly reduced hallucinations compared to previous models (after many iterations and improvements). In OpenAI's internal adversarial factuality evaluations, GPT-4 scored 40% higher than the latest GPT-3.5 model:

GPT-4 has also made progress on external benchmarks like TruthfulQA, where OpenAI tested the model's ability to distinguish facts from incorrect statements presented as adversarial alternatives. Results shown below.

The experimental results show that GPT-4's base model performs only slightly better than GPT-3.5 on this task; however, after RLHF post-training, the gap widens considerably. Below is a GPT-4 test example — it doesn't always make the right choice.

The model may exhibit various biases in its outputs. OpenAI has made progress on these fronts, with the goal of building AI systems whose default behaviors reasonably reflect widely shared user values.

GPT-4 generally lacks knowledge of events that occurred after the vast majority of its training data cutoff (September 2021), and it does not learn from experience. It sometimes makes simple reasoning errors that seem incongruous with its capabilities across so many domains, or is overly credulous of users' obviously false statements. It can also fail at difficult problems much as humans do, such as introducing security vulnerabilities into the code it generates.

GPT-4 can be confidently wrong in its predictions, and doesn't double-check when it suspects it might be wrong. Interestingly, the base pretrained model is highly calibrated (its predicted confidence in an answer typically matches the probability of being correct). However, through OpenAI's current post-training process, this calibration is reduced.

Risks and Mitigations

OpenAI states that the research team has been iteratively improving GPT-4 to make it safer and more aligned from the start of training, through efforts including pretraining data selection and filtering, evaluation and expert engagement, model safety improvements, and monitoring and enforcement.

GPT-4 carries risks similar to previous models, such as generating harmful advice, buggy code, or inaccurate information. At the same time, GPT-4's additional capabilities create new risk surfaces. To understand the extent of these risks, the team engaged more than 50 experts in fields including AI alignment risk, cybersecurity, biological risk, trust and safety, and international security to conduct adversarial testing of the model's behavior in high-stakes areas. These domains require specialized expertise to evaluate, and feedback and data from these experts informed mitigation measures and model improvements.

Risk Prevention

According to the OpenAI engineers in the demo video, GPT-4's training was completed last August; the time since then has been spent on fine-tuning improvements and, most importantly, on removing dangerous content generation. GPT-4's RLHF training incorporated an additional safety reward signal that reduces harmful outputs by training the model to refuse requests for such content. The reward is provided by GPT-4's zero-shot classifier, which judges safety boundaries and completions of safety-related prompts. To prevent the model from refusing valid requests, the team collected diverse datasets from various sources (e.g., labeled production data, human red-teaming, model-generated prompts) and applied safety reward signals (positive or negative) across both allowed and disallowed categories.

These measures substantially improved GPT-4's safety performance across many dimensions. Compared to GPT-3.5, the model's tendency to respond to requests for disallowed content dropped by 82%, while GPT-4's frequency of policy-compliant responses to sensitive requests (such as medical advice and self-harm) increased by 29%.

Training Process

Like previous GPT models, the GPT-4 base model was trained to predict the next word in a document. OpenAI used publicly available data (such as internet data) as well as licensed data for training. The training data is a web-scale corpus that includes correct and incorrect solutions to math problems, weak and strong reasoning, self-contradictory and consistent statements, and a wide variety of ideologies and ideas.

Consequently, when asked a question, the base model's response may diverge significantly from the user's intent. To align it with user intent, OpenAI still uses Reinforcement Learning from Human Feedback (RLHF) to fine-tune the model's behavior. Note that the model's capabilities appear to come primarily from the pre-training process — RLHF does not improve exam scores (and may even reduce them). But the model's steerability comes from the post-training process — the base model even requires prompt engineering to answer questions.

A major focus for GPT-4 was building a predictably scalable deep learning stack. The main reason is that for large-scale training runs like GPT-4, extensive model-specific tuning is not feasible. The team developed infrastructure and optimizations that behave predictably across a wide range of scales. To verify this scalability, they accurately predicted GPT-4's final loss on an internal codebase (not part of the training set) in advance, by extrapolating from models trained with the same methodology but using 1/10,000th the compute.

Now, OpenAI can accurately predict metrics (loss) optimized during training. For example, they extrapolated from a model with 1/1,000th the compute and successfully predicted pass rate on a subset of the HumanEval dataset:

Some capabilities remain difficult to predict. For example, the Inverse Scaling Prize sought to find a metric that gets worse as model compute increases, and the hindsight neglect task was one of the winners. GPT-4 reversed this trend.

The ability to accurately predict future machine learning capabilities is critical for technical safety, but it has not received sufficient attention, OpenAI stated, adding that it is investing more effort in developing relevant methods and calling on the industry to work together. OpenAI also announced it is open-sourcing the OpenAI Evals software framework, which is used to create and run benchmarks for evaluating models like GPT-4, and can inspect model performance sample-by-sample locally.

ChatGPT Upgraded Directly to GPT-4

Following GPT-4's release, OpenAI upgraded ChatGPT directly. ChatGPT Plus subscribers can access GPT-4 with usage limits at chat.openai.com.

To access the GPT-4 API (which uses the same ChatCompletions API as gpt-3.5-turbo), users can join the waitlist. OpenAI will invite select developers to try it out.

Once granted access, users can currently make pure text requests to the GPT-4 model (image input remains in a limited alpha stage). As for pricing, it is $0.03 per 1k prompt tokens and $0.06 per 1k completion tokens. Default rate limits are 40k tokens per minute and 200 requests per minute.

GPT-4's context length is 8,192 tokens. OpenAI also offers limited access to a 32,768-token context version (approximately 50 pages of text), which will also update automatically over time (current version gpt-4-32k-0314, supported until June 14). Pricing is $0.06 per 1K prompt tokens and $0.12 per 1k completion tokens.

That covers everything OpenAI shared about GPT-4 today. One disappointing point: OpenAI's published technical report contains no additional information about model architecture, hardware, compute, or other details — hardly living up to the "Open" name.

Regardless, impatient users have probably already started testing it out.

Reference: https://openai.com/product/gpt-4

A Little Easter Egg

BlueRun Ventures has always maintained its focus on and investment in AI.

A few days ago we issued the "AGI Pioneer Call"

Hoping to build the BlueRun AGI Pioneer Club

Scan the QR code to apply

👇👇👇

01 GPT-4: I Scored 710 on the SAT, and I Can Be a Lawyer Too

02 GPT-4: I Can Meme Too

A Little Easter Egg

More Great Content