Interview with OpenAI’s Greg Brockman: GPT-4 isn’t perfect, but neither are you


Greg Brockman onstage at TechCrunch Disrupt 2019
Image Credits: TechCrunch

OpenAI shipped GPT-4 yesterday, the much-anticipated text-generating AI model, and it’s a curious piece of work.

GPT-4 improves upon its predecessor, GPT-3, in key ways, for example giving more factually true statements and allowing developers to prescribe its style and behavior more easily. It’s also multimodal in the sense that it can understand images, allowing it to caption and even explain in detail the contents of a photo.

But GPT-4 has serious shortcomings. Like GPT-3, the model “hallucinates” facts and makes basic reasoning errors. In one example on OpenAI’s own blog, GPT-4 describes Elvis Presley as the “son of an actor.” (Neither of his parents were actors.)

To get a better handle on GPT-4’s development cycle and its capabilities, as well as its limitations, TechCrunch spoke with Greg Brockman, one of the co-founders of OpenAI and its president, via a video call on Tuesday.

Asked to compare GPT-4 to GPT-3, Brockman had one word: Different.

“It’s just different,” he told TechCrunch. “There’s still a lot of problems and mistakes that [the model] makes … but you can really see the jump in skill in things like calculus or law, where it went from being really bad at certain domains to actually quite good relative to humans.”

Test results support his case. On the AP Calculus BC exam, GPT-4 scores a 4 out of 5 while GPT-3 scores a 1. (GPT-3.5, the intermediate model between GPT-3 and GPT-4, also scores a 4.) And in a simulated bar exam, GPT-4 passes with a score around the top 10% of test takers; GPT-3.5’s score hovered around the bottom 10%.

Shifting gears, one of GPT-4’s more intriguing aspects is the above-mentioned multimodality. Unlike GPT-3 and GPT-3.5, which could only accept text prompts (e.g. “Write an essay about giraffes”), GPT-4 can take a prompt of both images and text to perform some action (e.g. an image of giraffes in the Serengeti with the prompt “How many giraffes are shown here?”).

That’s because GPT-4 was trained on image and text data while its predecessors were only trained on text. OpenAI says that the training data came from “a variety of licensed, created, and publicly available data sources, which may include publicly available personal information,” but Brockman demurred when I asked for specifics. (Training data has gotten OpenAI into legal trouble before.)

GPT-4’s image understanding abilities are quite impressive. For example, fed the prompt “What’s funny about this image? Describe it panel by panel” plus a three-paneled image showing a fake VGA cable being plugged into an iPhone, GPT-4 gives a breakdown of each image panel and correctly explains the joke (“The humor in this image comes from the absurdity of plugging a large, outdated VGA connector into a small, modern smartphone charging port”).

Only a single launch partner has access to GPT-4’s image analysis capabilities at the moment — an assistive app for the visually impaired called Be My Eyes. Brockman says that the wider rollout, whenever it happens, will be “slow and intentional” as OpenAI evaluates the risks and benefits.

“There’s policy issues like facial recognition and how to treat images of people that we need to address and work through,” Brockman said. “We need to figure out, like, where the sort of danger zones are — where the red lines are — and then clarify that over time.”

OpenAI dealt with similar ethical dilemmas around DALL-E 2, its text-to-image system. After initially disabling the capability, OpenAI allowed customers to upload people’s faces to edit them using the AI-powered image-generating system. At the time, OpenAI claimed that upgrades to its safety system made the face-editing feature possible by “minimizing the potential of harm” from deepfakes as well as attempts to create sexual, political and violent content.

Another perennial is preventing GPT-4 from being used in unintended ways that might inflict harm — psychological, monetary or otherwise. Hours after the model’s release, Israeli cybersecurity startup Adversa AI published a blog post demonstrating methods to bypass OpenAI’s content filters and get GPT-4 to generate phishing emails, offensive descriptions of gay people and other highly objectionable text.

It’s not a new phenomenon in the language model domain. Meta’s BlenderBot and OpenAI’s ChatGPT, too, have been prompted to say wildly offensive things, and even reveal sensitive details about their inner workings. But many had hoped, this reporter included, that GPT-4 might deliver significant improvements on the moderation front.

When asked about GPT-4’s robustness, Brockman stressed that the model has gone through six months of safety training and that, in internal tests, it was 82% less likely to respond to requests for content disallowed by OpenAI’s usage policy and 40% more likely to produce “factual” responses than GPT-3.5.

“We spent a lot of time trying to understand what GPT-4 is capable of,” Brockman said. “Getting it out in the world is how we learn. We’re constantly making updates, include a bunch of improvements, so that the model is much more scalable to whatever personality or sort of mode you want it to be in.”

The early real-world results aren’t that promising, frankly. Beyond the Adversa AI tests, Bing Chat, Microsoft’s chatbot powered by GPT-4, has been shown to be highly susceptible to jailbreaking. Using carefully tailored inputs, users have been able to get the bot to profess love, threaten harm, defend the Holocaust and invent conspiracy theories.

Brockman didn’t deny that GPT-4 falls short, here. But he emphasized the model’s new mitigatory steerability tools, including an API-level capability called “system” messages. System messages are essentially instructions that set the tone — and establish boundaries — for GPT-4’s interactions. For example, a system message might read: “You are a tutor that always responds in the Socratic style. You never give the student the answer, but always try to ask just the right question to help them learn to think for themselves.”

The idea is that the system messages act as guardrails to prevent GPT-4 from veering off course.

“Really figuring out GPT-4’s tone, the style and the substance has been a great focus for us,” Brockman said. “I think we’re starting to understand a little bit more of how to do the engineering, about how to have a repeatable process that kind of gets you to predictable results that are going to be really useful to people.”

Brockman also pointed to Evals, OpenAI’s newly open sourced software framework to evaluate the performance of its AI models, as a sign of OpenAI’s commitment to “robustifying” its models. Evals lets users develop and run benchmarks for evaluating models like GPT-4 while inspecting their performance — a sort of crowdsourced approach to model testing.

“With Evals, we can see the [use cases] that users care about in a systematic form that we’re able to test against,” Brockman said. “Part of why we [open sourced] it is because we’re moving away from releasing a new model every three months — whatever it was previously — to make constant improvements. You don’t make what you don’t measure, right? As we make new versions [of the model], we can at least be aware what those changes are.”

I asked Brockman if OpenAI would ever compensate people to test its models with Evals. He wouldn’t commit to that, but he did note that — for a limited time — OpenAI’s granting select Evals users early access to the GPT-4 API.

Brockman’s conversation also touched on GPT-4’s context window, which refers to the text the model can consider before generating additional text. OpenAI is testing a version of GPT-4 that can “remember” roughly 50 pages of content, or five times as much as the vanilla GPT-4 can hold in its “memory” and eight times as much as GPT-3.

Brockman believes that the expanded context window lead to new, previously unexplored applications, particularly in the enterprise. He envisions an AI chatbot built for a company that leverages context and knowledge from different sources, including employees across departments, to answer questions in a very informed but conversational way.

That’s not a new concept. But Brockman makes the case that GPT-4’s answers will be far more useful than those from chatbots and search engines today.

“Previously, the model didn’t have any knowledge of who you are, what you’re interested in and so on,” Brockman said. “Having that kind of history [with the larger context window] is definitely going to make it more able … it’ll turbocharge what people can do.”

More TechCrunch

London-based fintech Vitesse has closed a $93 million Series C round of funding led by investment giant KKR.

Vitesse, a payments and treasury management platform for insurers, raises $93M to fuel US expansion

Zen Educate, an online marketplace that connects schools with teachers, has raised $37 million in a Series B round of funding. The raise comes amid a growing teacher shortage crisis…

Zen Educate raises $37M and acquires Aquinas Education as it tries to address the teacher shortage

“When I heard the released demo, I was shocked, angered and in disbelief that Mr. Altman would pursue a voice that sounded so eerily similar to mine.”

Scarlett Johansson says that OpenAI approached her to use her voice

A new self-driving truck — manufactured by Volvo and loaded with autonomous vehicle tech developed by Aurora Innovation — could be on public highways as early as this summer.  The…

Aurora and Volvo unveil self-driving truck designed for a driverless future

The European venture capital firm raised its fourth fund as fund as climate tech “comes of age.”

ETF Partners raises €284M for climate startups that will be effective quickly — not 20 years down the road

Copilot, Microsoft’s brand of generative AI, will soon be far more deeply integrated into the Windows 11 experience.

Microsoft wants to make Windows an AI operating system, launches Copilot+ PCs

Hello and welcome back to TechCrunch Space. For those who haven’t heard, the first crewed launch of Boeing’s Starliner capsule has been pushed back yet again to no earlier than…

TechCrunch Space: Star(side)liner

When I attended Automate in Chicago a few weeks back, multiple people thanked me for TechCrunch’s semi-regular robotics job report. It’s always edifying to get that feedback in person. While…

These 81 robotics companies are hiring

The top vehicle safety regulator in the U.S. has launched a formal probe into an April crash involving the all-electric VinFast VF8 SUV that claimed the lives of a family…

VinFast crash that killed family of four now under federal investigation

When putting a video portal in a public park in the middle of New York City, some inappropriate behavior will likely occur. The Portal, the vision of Lithuanian artist and…

NYC-Dublin real-time video portal reopens with some fixes to prevent inappropriate behavior

Longtime New York-based seed investor, Contour Venture Partners, is making progress on its latest flagship fund after lowering its target. The firm closed on $42 million, raised from 64 backers,…

Contour Venture Partners, an early investor in Datadog and Movable Ink, lowers the target for its fifth fund

Meta’s Oversight Board has now extended its scope to include the company’s newest platform, Instagram Threads, and has begun hearing cases from Threads.

Meta’s Oversight Board takes its first Threads case

The company says it’s refocusing and prioritizing fewer initiatives that will have the biggest impact on customers and add value to the business.

SeekOut, a recruiting startup last valued at $1.2 billion, lays off 30% of its workforce

The U.K.’s self-proclaimed “world-leading” regulations for self-driving cars are now official, after the Automated Vehicles (AV) Act received royal assent — the final rubber stamp any legislation must go through…

UK’s autonomous vehicle legislation becomes law, paving the way for first driverless cars by 2026

ChatGPT, OpenAI’s text-generating AI chatbot, has taken the world by storm. What started as a tool to hyper-charge productivity through writing essays and code with short text prompts has evolved…

ChatGPT: Everything you need to know about the AI-powered chatbot

SoLo Funds CEO Travis Holoway: “Regulators seem driven by press releases when they should be motivated by true consumer protection and empowering equitable solutions.”

Fintech lender SoLo Funds is being sued again by the government over its lending practices

Hard tech startups generate a lot of buzz, but there’s a growing cohort of companies building digital tools squarely focused on making hard tech development faster, more efficient and —…

Rollup wants to be the hardware engineer’s workhorse

TechCrunch Disrupt 2024 is not just about groundbreaking innovations, insightful panels, and visionary speakers — it’s also about listening to YOU, the audience, and what you feel is top of…

Disrupt Audience Choice vote closes Friday

Google says the new SDK would help Google expand on its core mission of connecting the right audience to the right content at the right time.

Google is launching a new Android feature to drive users back into their installed apps

Jolla has taken the official wraps off the first version of its personal server-based AI assistant in the making. The reborn startup is building a privacy-focused AI device — aka…

Jolla debuts privacy-focused AI hardware

The ChatGPT mobile app’s net revenue first jumped 22% on the day of the GPT-4o launch and continued to grow in the following days.

ChatGPT’s mobile app revenue saw its biggest spike yet following GPT-4o launch

Dating app maker Bumble has acquired Geneva, an online platform built around forming real-world groups and clubs. The company said that the deal is designed to help it expand its…

Bumble buys community building app Geneva to expand further into friendships

CyberArk — one of the army of larger security companies founded out of Israel — is acquiring Venafi, a specialist in machine identity, for $1.54 billion. 

CyberArk snaps up Venafi for $1.54B to ramp up in machine-to-machine security

Founder-market fit is one of the most crucial factors in a startup’s success, and operators (someone involved in the day-to-day operations of a startup) turned founders have an almost unfair advantage…

OpenseedVC, which backs operators in Africa and Europe starting their companies, reaches first close of $10M fund

A Singapore High Court has effectively approved Pine Labs’ request to shift its operations to India.

Pine Labs gets Singapore court approval to shift base to India

The AI Safety Institute, a U.K. body that aims to assess and address risks in AI platforms, has said it will open a second location in San Francisco. 

UK opens office in San Francisco to tackle AI risk

Companies are always looking for an edge, and searching for ways to encourage their employees to innovate. One way to do that is by running an internal hackathon around a…

Why companies are turning to internal hackathons

Featured Article

I’m rooting for Melinda French Gates to fix tech’s broken ‘brilliant jerk’ culture

Women in tech still face a shocking level of mistreatment at work. Melinda French Gates is one of the few working to change that.

2 days ago
I’m rooting for Melinda French Gates to fix tech’s  broken ‘brilliant jerk’ culture

Blue Origin has successfully completed its NS-25 mission, resuming crewed flights for the first time in nearly two years. The mission brought six tourist crew members to the edge of…

Blue Origin successfully launches its first crewed mission since 2022

Creative Artists Agency (CAA), one of the top entertainment and sports talent agencies, is hoping to be at the forefront of AI protection services for celebrities in Hollywood. With many…

Hollywood agency CAA aims to help stars manage their own AI likenesses