Greetings. It promises to be (another) massive week in A.I. news. And that’s leaving aside the lingering effects that the collapse of Silicon Valley Bank may have on some A.I. startups and the venture funds backing them.
Right as this newsletter was going to press, OpenAI released its long-anticipated GPT-4 model. The new model is multimodal, accepting both images and text as inputs, although it only generates text as its output. According to data released by OpenAI, GPT-4 performs much better than GPT-3.5, its latest model, and the one that powers ChatGPT, on a whole range of benchmark tests, including a battery of different tests designed for humans. For instance, GPT-4 scores well enough to be within the top 10% of test takers on a simulated bar exam. OpenAI also says that GPT-4 is safer than GPT-3.5—returning more factual answers and it's much more difficult to get GPT-4 to jump its guardrails than has been the case with GPT-3.5.
But, the company is also saying that the model is still flawed. It will still hallucinate—making up information. And OpenAI notes that in some ways hallucination might be more of an issue because GPT-4 does this less often, so people might get very complacent about the answers it produces. It is also still possible to get the model to churn out biased and toxic language. OpenAI is saying very little about how big a model GPT-4 actually is, how many specialized graphics processing units it took to train it, or exactly what data it was trained on. It says it wants to keep these details secret for both competitive and safety reasons. I'll no doubt be writing much more about GPT-4 in next week's newsletter. But my initial take is that GPT-4 looks like a big step forward, but not a revolutionary advance over what OpenAI and others have been racing to put into production over the past two months. And it will only heighten the debate about whether tech companies, including OpenAI, are being irresponsible by putting this powerful technology in the hands of consumers and customers despite its persistent flaws and drawbacks.
Meanwhile, Microsoft is expected to unveil a range of A.I.-powered enhancements to its Office software suite on Thursday. And Baidu, the Chinese search giant, has a big announcement scheduled for later this week. Google, which was caught flat-footed by the viral popularity of ChatGPT and OpenAI's alliance with Microsoft, is eager to prove that it’s not about to be sidelined in the A.I. race. And the big news today before OpenAI's GPT-4 announcement was that Google had beaten Microsoft out of the gate with a bunch of big A.I. announcements of its own.
For most people, the main news is that the search giant said it is adding generative-A.I. features to its popular Workspace productivity tools, such as Google Docs, Sheets, and Slides. Among the things people will now be able to do is use a text box to prompt Google’s A.I. to automatically draft almost any kind of document, or to create different kinds of charts for Sheets data. Users can highlight text and ask Google’s A.I. to edit it for them or rewrite it in a different tone and style. You will also be able to automatically draft emails or summarize entire email threads in Gmail. In Google Meet you will be able to generate new virtual backgrounds and automatically create notes of conversations, complete with summaries.
But equally important was the other news Google announced: The company is allowing enterprise customers to tap its most advanced family of large language models—called PaLM —through an application programming interface on Google Cloud.
Beyond PaLM, it has also launched an updated set of its Vertex AI platform for A.I. developers and data scientists. The platform allows them access to large foundation models, not just from Google, but from its growing ecosystem of allied A.I. labs, such as Anthropic and Cohere, as well as AI21 Labs and Midjourney. And it has launched a set of software, called Generative AI App Builder, that will allow slightly less technical teams to quickly build and roll out custom applications using generative A.I. models.
For both Vertex AI and the Generative AI App Builder, Google says users will have access to two new related capabilities: The first is an enterprise search tool that will allow them to perform Google searches across their own data—including data generated by CRM or ERP software, as well as internal websites and other documents—and return results only from that knowledge base. These results can then be used for natural language tasks, such as summarization, sentiment analysis, or question-answering, with less risk that the language model will simply invent information or draw information from its pretraining data rather than the customer’s own data. The other new capability is a chatbot-like “conversational A.I.” function that customers can deploy to act as the user interface for these search, natural language processing, and generative A.I. capabilities.
Google announced a group of initial “trusted testers” who will have immediate access to these new A.I. services including Toyota, Deutsche Bank, HCA Healthcare, Equifax, the television network Starz, and the Mayo Clinic, among others. The new products and features will be rolled out more broadly in the coming weeks, the company said. But it was a sign of just how intense this A.I. technology race has become that Thomas Kurian, the CEO of Google’s Cloud business, was forced to acknowledge during the press briefing that although Google was releasing these new products without having yet worked out exactly how to price them. In the past, Kurian said, Google had always made its A.I. advances available as free, open-source releases or the technology was simply “embedded in our products.” “This is the first time we are taking our new, general A.I. models and making them accessible to the developer community with an API,” he said.
Google’s press release on its new products touted the company’s commitment to “Responsible AI” and it tried to position its release under this rubric, noting that Vertex AI and Generative AI App Builder include tools to “inspect, understand, and modify model behavior” and that the information retrieval aspects of the new systems used traditional search algorithms, lessening the risk of inaccurate answers. But Kurian did not say exactly what sort of guarantees Google could offer customers that its large language models could not be prompted in ways that would elicit inaccurate responses—or worse, might morph their chatbot from a friendly assistant into a petulant, abusive, and threatening “devil-on-your-shoulder,” as testers discovered with Microsoft’s Bing. It also did not address whether Google was planning to take any steps to prevent users of its very popular Workspace tools from using the new generative A.I. features to deliberately churn out misinformation or to cheat on school essays.
Concern about this is growing. I recently debated Gary Marcus on a Canadian podcast—which hasn’t aired yet—about whether ChatGPT will wind up doing more harm than good. I am more sanguine about the technology’s potential than Gary is, but there is much we agree on. Gary is among those trying desperately to raise the alarm about the potential dangers—particularly when it comes to the industrial-scale production of misinformation that these generative A.I. systems represent. I recommend his recent piece in The Atlantic on this topic as well as his blog post questioning why those building advanced A.I. systems persist in doing so despite being fully aware of the potentially calamitous impacts the technology they are creating could have.
One reason may be that most of those researchers are now embedded inside big tech companies and if they step out of line, they get fired. Tech news site The Verge and Casey Newton’s The Platformer just revealed that Microsoft recently disbanded its A.I. ethics and society team—a central group that had been trying to raise concerns about many of the advanced A.I. systems Microsoft was building and had been urging the company to slow down the speed of its generative A.I. roll out. Some of the ethics experts were assigned to other teams. Some were fired. An audio recording of a Microsoft manager addressing the team about its restructuring that leaked to Newton made it clear that there was pressure from CEO Satya Nadella and CTO Kevin Scott to roll out OpenAI’s advanced A.I. technology throughout the company as quickly as possible and that questioning that decision or its pace was not appreciated.
Now Microsoft still has another corporate Office of Responsible AI, but its role is more to set high-level principals, frameworks, and processes—not to conduct the actual safety and ethical checks. The disbanding of the A.I. ethics group is further evidence of why the tech industry should not be trusted to self-regulate when it comes to A.I. ethics or safety and why government regulation is urgently needed.
Before we get to the rest of this week’s A.I. news, a couple of quick corrections: In last week’s newsletter, I got the surname of one of the cofounders of legal A.I. software company Casetext wrong. He is Pablo Arredondo, not Arrodondo. I also erroneously capitalized the letter 't' in Casetext. I regret both errors.
And with that here’s the rest of this week’s news in A.I.
Jeremy Kahn
@jeremyakahn
jeremy.kahn@fortune.com
